Fundamental Concepts in AI Ethics

Fundamental Concepts in AI Ethics by MAIEI Support

1. Automation bias

1.1. Definition

1.1.1. Automation bias principally refers to bias on behalf of human operators towards an automated system when making a decision. Here, given how our brains are wired to take the path of least cognitive resistance, a bias towards said path provided by AI machines can then take hold through using AI as a cognitive heuristic. From there, excessive dependence on the machines is then created, with an assumption that the automated outcome is always right being established. In this sense, automation bias leads to the AI being seen as an expert, rather than as an advisory tool. Being biased towards such machines in this way brings with it various problems, some of which can stem from the database itself. Here, should the AI be learning from a flawed data set (such as one which is not adequately representative), the outcomes it produces will be based on this flawed data and, consequently, so will the humans decision to enact such outcome through their automation bias. In addition, learning from large data sets also sews in the assumption that the norm established by that dataset will not be radically different in the future. That is to say, the general pattern viewed in such dataset is later applied in every use case without properly taking into account how some situations may be too unique for the norm to apply. An error of this nature can then be labelled as an automated commission error, whereby despite the AI's output (such as the dataset norm) contradicting all the present information about a certain situation, the human operators follow it blindly nonetheless. Should the output not contradict the present information, an omission error may ensue whereby the automated nature of the process fails to signal to the human operators the coming occurrence of a potential problem (for example, a mobile phone virtual assistant deleting certain photos from your phone without you knowing to save space on your device, which then goes on to delete a favourite app of yours).

1.2. Relevance in AI ethics

1.2.1. Through an AI ethics lens, a major risk can be seen in the assumption made about the veracity of the automated decision process. Through assuming through automation bias that the automated process is correct, there risks the emergence of a taboo about questioning the output of such systems, whereby doing so is frowned upon and unpracticed (as seen with automated commission errors). As a result, despite potential glaring red flags being presented in terms of the automated outcome contradicting the present reality, going against such an output is no longer seen as an option. To potentially compound the problem, this 'trusted' automated output could then be based on flawed data, bringing with it grave consequences.

1.3. Example

1.3.1. The best examples to illustrate such bias come from the aviation industry. Here, imagine yourself as a pilot flying your usual route that you have flown hundreds of times. Here, the autopilot is employed at various times during the flight, regulating the balance and height of the plane. Once the plane has reached cruising altitude, the passengers then can get out of their seat and move around/change seats to be closer to family members. For whatever reason, the autopilot then decides to take the plane upwards because it believes it can best avoid turbulence that way. The seat belt sign turns on, and the passengers sit down where they are, with the pilots trusting the decision the autopilot has made. Once the autopilot believes it sufficient to have avoided the turbulence, it brings the plane back down. The pilots then disengage the autopilot and find that the plane starts tilting to the left as a result of the passenger movement and their sitting down of where they moved to, thus unbalancing the plane. The pilots then announce to the passengers to please return to their seats, and the plane balances out before it lands safely. Here, a commission bias can be seen in the form of the pilots blindingly trusting the autopilot when it decided to take the plane further up to avoid any turbulence, where may or may not have existed. From there, an omission bias can then be seen in the form of the plane tilting to the left once the autopilot was removed, owing to the automated balancing of the autopilot hiding the problem that they plane was not equally weighted from the pilots. Such errors, although not grave in this case, have the potential to seriously complicate such situations, owing to the taboo surrounding questing the automated outcomes produced by the automated system.

1.4. Helpful links

1.4.1. (PDF) Automation Bias: Decision Making and Performance in High-Tech Cockpits

1.5. Common Contexts

1.5.1. #automationbias #bias

2. Ethical Debt

2.1. Definition

2.1.1. Ethical debt in the AI context can be described as the design, development, deployment and use of an AI system by an agent or corportation without adequately considering the ethical issues surrounding said system. In this sense, as each decision is made within this process, the ethical considerations that are not taken start collecting themselves as "debt" to be "paid back" by someone other person, group or entity once the system has been deployed. Due to the wide-reaching effects of AI, such person, group or entity which ends up paying for the ethical debt can be disproportionately related to a minority group, having to pay back the debt in a myriad of ways. These can include suffering from biased system outcomes to having to endure unfairly distributed negative consequences.

2.2. Relevance in AI Ethics

2.2.1. Under the AI ethics lens, what generally leads to the existence of ethical debt is the assumptions being made during the AI system deployment process. Aspects such as not subjecting the data set to bias checks and assuming a universal effect of the system over the general population fail to embrace the full socio-technical context of the society which the AI is deployed in. From there, such failings begin to mount and the more unaddressed assumptions made, the more difficult it becomes to solve and "pay back". One potential solution that is then proposed to solve this AI ethical dilema is a "human in the loop" solution. However, this brings its own problems through the human itself being subjected to its own bias and assumptions.

2.3. Example

2.3.1. Ethical debt can best be demonstrated through an assumption being made. For example, a government may decide that broadcasting the latest news on the new national COVID-19 lockdown restrictions is to be done solely through an AI-automated television news channel. The channel gives all the latest scientific recommendations straight from the scientists themselves, as well as the new restrictions straight from the ministers involved, with no additional rendering of the information. However, having not fully understood the socio-technical context their system finds itself in, the government has neglected and disadvantaged different sections of the population. Here, those who do not have access to a TV are not able to acess the latest restrictions, while the lack of further rendering of the information has made the information being shared unintelligible to some percentages of the population. Hence, the ethical debt starts to surface in the form of infection rates rising through a lack of an opportunity to access and understand the information being shared. In other words, the assumptions made by the government on the technology access and overall intelligence of the population has lead to the population itself having to "pay back" the ethical debt accrued by said assumptions.

2.4. Helpful Links

2.4.1. The topic of ethical debt is mentioned here in Radical AI's podcast session with Kathy Baxter, Principal Architect of Ethical AI Practice at Salesforce: Industry AI Ethics 101 with Kathy Baxter — The Radical AI Podcast

2.5. Common Contexts

2.5.1. #ethicaldebt #accountability #representation #ethics

3. AI Consciousness

3.1. Definition

3.1.1. AI consciousness is a hotly debated topic within the realm of cognitive robotics, and tends to be discussed in terms of whether AI can be held morally responsible or not if consciousness is achieved. Here, consciousness is believed to differ from 'sentience', also known as the ability to feel or perceive things, which is a function that AI can arguably already perform by virtue of processing and learning from its own data outputs. On the other hand, conscious AI or 'strong AI', if instantiated, would involve self-awareness by the machine. However, there does not even a shared understanding of how consciousness arises in humans, which is demonstrated by the 'hard problem of consciousness'. If we don't understand how consciousness arises in a carbon substrate then it is quite difficult to determine whether consciousness could arise in silicone. Our inability to understand human consciousness will likely limit our ability to determine the existence, or lack thereof, of consciousness in AI.

3.2. Relevance in AI Ethics

3.2.1. AI ethics focuses on the connection between AI consciousness and moral responsibility. As automation and 'black box' algorithms increase in prevalence, establishing moral responsibility becomes more pressing. Acknowledging 'strong AI' would help settle responsibility doubts, but then present a whole host of new challenges in accountability and justice. For example, if AI can be deemed a conscious and thus morally responsible agent, how would they be held accountable for any crimes they commit? Would we require an AI bill of rights? Hence, AI consciousness can be seen as intrinsically connected with the AI ethics world, and something that may become a hot topic in the future.

3.3. Example

3.3.1. The classic Trolley Problem can be used to demonstrate the importance of AI consciousness. Here, the self-driving car is faced with an impending crash as well as a moral dilemma of whether to save the passenger in the car or whether to save mother and her child. No matter what outcome results, the vehicle has elected to undertake a decision. Thus, can the vehicle's AI be blamed as the entity that is responsible for either the passenger's or the mother and child's death?

3.4. Helpful Links

3.4.1. https://www.frontiersin.org/articles/10.3389/fpsyg.2019.01535/full

3.4.2. https://www.washingtonpost.com/local/trafficandcommuting/when-driverless-cars-crash-who-gets-the-blame-and-pays-the-damages/2017/02/25/3909d946-f97a-11e6-9845-576c69081518_story.html

3.4.3. https://link.springer.com/article/10.1007/s10677-016-9745-2

3.4.4. The Problem of AI Consciousness

3.5. Common Contexts

3.5.1. #aiethics #trolleyproblem #autonomouscars #consciousness #morality #responsibility #accountability #justice

4. Fairness

4.1. Definition

4.1.1. AIgorithmic fairness is the principle that the outputs of an AI system should be uncorrelated with particular characteristics, such as gender, race, or sexuality. There are many ways models can be considered fair. Common approaches to AI fairness include: equal false positives across sensitive characteristics, equal false negatives across sensitive characteristics, or minimizing “worst group error” which is the number of mistakes the algorithm makes on the least represented group. While it's possible for an AI to be considered fair across sensitive characteristics independently, the AI may be unfair from an intersectional perspective (discriminating against those at the intersection of multiple sensitive characteristics). A common argument against manipulating the models to allow for AI fairness is the inaccuracy that may arise as a result.

4.2. Relevance in AI Ethics

4.2.1. AI ethics examines how social values such as fairness can be upheld in AI systems. The difficulty is that concepts of fairness, such as demographic parity and equal opportunity, which result in equal accuracy are mathematically challenging if that parity and equality does not exist in reality. If AI fairness, say in loan applications, entails achieving demographic parity between two groups of people, models might refuse loans to repaying applicants and instead give loans to defaulting applicants. One solution might be to evenly distribute the mistakes made over the number of loans given, whereby we have the same false rejection rates for the two groups. Some might consider this to be fair to the collective but not to the individual. Even in cases where the AI models do protect sets of individual sensitive attributes, we can end up with a phenomenon known as fairness gerrymandering, where specific subgroups of the population are unfairly discriminated against. AI Ethics will have to grapple with these conflicting realities of algorithmic fairness and their trade-offs when determining what constitutes an objective and just AI system. It will also need to account for the inequalities present in society, especially when these biases are compounded by AI systems

4.3. Example

4.3.1. One example of an AI system that brings up questions of fairness is the Allegheny Family Screening Tool (ASFT), which works to support social workers when deciding whether to remove a child from their home for reasons of neglect or abuse. The goal of ASFT was to optimize accuracy and reduce incidences of false negatives (reducing the number of children wrongly removed from a loving home in which the child has not been abused or neglected). The ASFT team found that their algorithm was biased against poor families. In fact, the team found that one quarter of the variables used to predict abuse and neglect were direct measures of poverty (e.g. whether the family relied on food stamps). As a result, families relying on food stamps were often rated as higher risk even though this metric is not directly correlated with child neglect or abuse. Thus, even in cases where AI developers do their best to design fair and impartial models, these models cannot be separated from the biases and injustices embedded in the data.

4.4. Helpful Links

4.4.1. Algorithmic Bias and Fairness: Crash Course AI #18

4.4.2. Michael Kearns: Algorithmic Fairness, Privacy, and Ethics in Machine Learning | AI Podcast

4.5. Common Contexts

4.5.1. #proxyindicators #injustice #human rights #blackbox

5. Explainability

5.1. Definition

5.1.1. Explainability is the principle that humans should be able to interpret and understand how an AI system derived its output. Therefore, the goal of explainability is for the human to be able to explain, in non-technical terms, how the AI's inputs led to the AI's outputs. The term explainability can refer to either global explainability or local explainability. Global explainability implies that humans can understand the relationships the model has found between inputs and outputs as a whole. For example, an AI would require global explainability in communicating to humans whether its algorithm uses racial characteristics to determine recidivism rates. Local explainability, on the other hand, refers to humans understanding why an algorithm gave a particular output following a particular input, rather than explaining a general relationships that exists in the model.

5.2. Relevance in AI Ethics

5.2.1. Explainability is critical if we ever hope to understand AI's decision making process. Since AI can be trained using data that contains latent bias, the algorithm may optimize itself to perpetuate that bias. Once AI systems have been optimized to reproduce bias, the AI's decisions will entrench systemic discrimination, inequality and a lack of access to essential services.

5.3. Example

5.3.1. In August 2019, Apple and Goldman Sachs released their joint venture credit card. In November a couple using the card realized that the husband was receiving 20x the credit limit of his wife, who had a better credit score. Apple was then investigated by the New York State Department of Financial Services for discrimination based on gender. Although the company claims that the algorithm it used does not make decisions based on age, gender or sexual orientation, the inability to explain how the AI made its decision has resulted in Apple's failure to protect against biased decision making.

5.4. Helpful Links

5.4.1. https://www.forbes.com/sites/cognitiveworld/2019/07/23/understanding-explainable-ai/#13a3b887c9ef

5.4.2. https://www.seattletimes.com/business/judge-ibm-owes-indiana-78m-for-failed-welfare-automation/

5.5. Common Contexts

5.5.1. #optimizedforracism #blackbox #injustice

6. AI Justice

6.1. Facial Recognition Technology

6.1.1. Definition

6.1.1.1. Facial recognition technology (FRT) does what it says on the tin, namely being technology which recognises faces. There are various methods as to how it does this. To begin with, the FRT analyses the measurements of a particular face (from several pictures) that it wants to be able to identify, and then makes a template of that same face (a representative array of numbers). Once this template is created, the FRT uses the template as a filter to sort through all the other faces it scans until it finds the match it's looking for. This use of the template can then be separated into two strands. Facial Verification is where somebody's face has to match to a singular template (stored locally), where the technology verifies that the face being scanned is the same as the template (such as unlocking your mobile phone with your face template already stored on said phone). Facial Identification, on the other hand, is about one-to-many matching, whereby a template is then compared to millions of images to identify which ones match (the template goes to the faces rather than the faces to the template), such as is used in scanning CCTV footage. Facial verification is more likely to be automated, with a match proving enough to warrant an action (such as unlocking your phone). Facial identification is more likely to be augmentative, being overseen by a human before a decision is made (a human verifies the quality of matches). Such verification can then be found in the similarity scores used by the business implementing the technology, sometimes with the FRT providing a rating out of 10 of how close to the template the face just scanned is. Such scanning within the identification strand of FRT is what presents the most hazards in the FRT space, especially when used live (rather than retroactively). In order to improve the FRT's scanning capablities, FRT systems are trained on sample faces at different angles in different lighting conditions to mimic those found in popular use cases, such as CCTV in public spaces. Hence, one of the key steps in training FRT is to identify the face you want to look for (whether in creating a template for verification or identification) and crop the image accordingly as to minimise the noise experienced in the background. What can then be separated from FRT is the use of biometric data. Biometrics enable the identification of individuals based on their biological and behavioural characteristics. This can be done retroactively on images, or live in terms of faces.

6.1.2. Relevance in AI Ethics

6.1.2.1. Under the AI ethics lens, FRT is particularly fruitful in terms of the problems that arise, especially in the use of live FRT (as opposed to retroactive uses such as scanning CCTV footage). As mentioned in the definition, identification FRT presents more hazards than verification FRT. One reason for this can be seen is that the template being used to identify, rather than solely verify, has a lot more variable factors to deal with than simply verifying. Identifying requires the FRT to be able to perform accurate identifications in different lighting and video quality settings on thousands of different subjects, which verification doesn't have to confront. Thus, the possibility of producing an accurate identification each time is made ever more difficult. Furthermore, identification FRT often uses similarity scores which are particular to the business employing the technology, leaving ample room for company biases to creep in, especially in setting the threshold for what constitutes a "good enough" match (such as 7/10 or 8/10). For example, if the FRT was being used to try and search for a criminal, such arbitrary establishments could then lead to unlawful convictions of members of the public being confused with whom the company is looking for. Not only the scanning itself, but the training of such models can also present ethical problems, especially in the remit of privacy. Here, the data collection required to train the models doesn't often include the explicit consent of those being facially scanned, especially if this is taking place in a popular public space. So, companies wanting to employ the technology are having to justify such collection in the name of "the public interest", which is both vague and difficult to quantify. Above all, the quality of the images or face scans acquired also matters. Collecting vast amounts of images doesn't guarantee accurate FRT systems if all the images are homogeneous, whereby the data set requires diversity to be appropriately representative of society, which is hardly ever achieved.

6.1.3. Example

6.1.3.1. In 2019, the Liberty group took the South Wales Police (SWP) to court in the UK over breaching the Data Privacy Act and Equality Act through their use of their automated (identification) FRT. The Supreme Court then ruled that they had followed the required legal frameworks, and did not use it arbitrarily, initially ruling in favour of the SWP. However, Liberty then appealed and the SWP’s use was deemed unlawful in July 2020, finding that the SWP had not conducted an adequate data impact assessment and had not sufficiently checked for racial and gender bias in their tech.

6.1.4. Helpful links

6.1.4.1. A Snapshot Series paper released by the UK government offering an introduction to FRT: https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/905267/Facial_Recognition_Technology_Snapshot_UPDATED.pdf

6.1.5. Common contexts

6.1.5.1. #facialrecognition #ukgovernment #privacy #identification #verification

6.2. Techno-solutionism

6.2.1. Definition

6.2.1.1. Techno-solutionism is the notion whereby it is believed that humankind's problems can all be solved by technological solutions alone. Here, the focus is on the solutions that the technical arena can offer rather than also exploring any human resources or political avenues. Such a fixation stems from the runaway successes that technology has had, as well as the belief that technology is not emotionally charged or as easily corruptible as a human, making it the ideal candidate to fairly decide upon a situation. As a result, technologies such as facial recognition technology and the use of technology during the pandemic have been implemented in order to demonstrate technology's capacity to solve different problems humanity faces, such as policing and the spread of the virus.

6.2.2. Relevance in AI ethics

6.2.2.1. The knock-on effect of adopting a techno-solutionist attitude in the realm of AI ethics mainly revolves around the subsequent ignoring of the biases and inequalities that technology can perpetuate. Through cutting out the more human-orientated avenues of exploration, the social problems found within society are then reflected in the technology itself. This further increases the human dependency on the digital and thus, the power that companies providing the technology possess to influence political decisions and potentially pursue their own self-interests. In this sense, the technological-solutionism attitude paves the way for technical solutions to be the norm, including the negative social costs that it brings.

6.2.3. Example

6.2.3.1. Due to the pandemic, public examinations in the UK were cancelled for the summer of 2020. The exam halls would not be big enough across the board to hold students at a social distance, while students in different personal situations had struggled to have the same quality of education as they would have done had they been in school physically. So, the UK government opted for an algorithm to predict students' grades based on 3 criteria: 1. The ranking of their school during 2017-2019, 2. The students ranking in the class of each subject based on teacher evaluation, 3. The previous test results of the student. This resulted in more than 40% of students receiving lower grades than they were predicted to get, as well as disproportionately affecting BAME students, with thousands missing out on university places as a result. This sparked mass-protest in the UK and eventually led to the government scrapping the algorithm and opting for teacher-based assessments instead. What this is to show is how a techno-solutionism's prioritization of technical means can lead to the exacerbation rather than solution of different problems human society faces. Technological solutions are not always the silver bullet to every situation and they can often make the problem worse without the exploration of other avenues to consider. Source: “F**k the algorithm”?: What the world can learn from the UK’s A-level grading fiasco

6.2.4. Helpful links

6.2.4.1. “F**k the algorithm”?: What the world can learn from the UK’s A-level grading fiasco https://knightfoundation.org/philanthropys-techno-solutionism-problem/

6.2.5. Common contexts

6.2.5.1. #technosolutionism #algorithm #humanity

6.3. Market Fundamentalism

6.3.1. Definition

6.3.1.1. Market fundamentalism is the belief that when markets are left alone, they will be able to produce the greatest possible equity in both the economic and social sector, thanks to operating without restraint. Usually a pejorative term used by critics of a laissez faire market interpretation, it has been used to argue for the natural setting of wage levels, a reduction in taxes on businesses and a de-regulation of different markets. In this way, the economy is viewed as an impersonal mechanism that produces efficient outputs, always arriving at its point of equilibrium. The theory has been widely criticised in the world of economics, being labelled as a dogmatic theory reserved for billionaires, and has come under particular scrutiny during the COVID-19 pandemic with Western economies showing their true fragile selves.

6.3.2. Relevance in AI ethics

6.3.2.1. Within the world of AI, market fundamentalism would point towards the self-regulation of the AI sector, leaving businesses to set their own rewards and punishments themselves. In this sense, the line of thought goes that eventually, the AI sector will produce the ideal market balance to be able to confront both the economic and social problems that it faces. However, the AI ethics lens revelas to us that applying market fundamentalism to the current AI sector faces a lot of red flags. Like the view of the economy as a dehmanised mechanism, AI could inherit the same view and the outputs different AI applications produce could then presumed to be objective. However, through the exposure of the bias and discrimination that AI can proliferate, such a view of AI as something objective would be truly damaging to human society. Coupling this with a lack of regulation could then prove a deadly combination in which there is no accountability for the 'objective' measures that AI produces. For example, without environmental regulation on the computing power required for different AI models, we run the risk of leaving the environment at the mercy of the business sector which, historically, has not prioritised the environment all too often. A lack of regulation then makes the task of lobbying businesses to make a change even more difficult, with no legal obligation for the business to adhere to environmental demands, nor any privacy concerns.

6.3.3. Example

6.3.3.1. In terms of a metaphor, the best way to envisage the effects market fundamentalism could have is to imagine you wee starting an AI business yourself. In a non-market-fundamentalism world, regulations on how your conduct yourself with customer data, liability and anti-discrimination laws are all in play, and they are binding. In this world, there are legal reasons with legal consequences should these laws be broken. However, in a market fundamentalism world, all is different. It will most likely be seen that having anti-discrimination laws, protecting customer privacy and liability are all good business practices, but there is nothing regulating this. There isn't an external body that enforces these considerations, or follows up should they be broken. AI, through self-regulation in market fundamentalism, will eventually land up at taking all of these considerations into account to produce the ideal balance in the economic and social sectors. So, as a business, which world would you choose? It may be safe to say that, given the choice, most business owners would choose the non-regulation option to avoid potential court dates and bureacracy in general. It may be that, as a result, your business becomes better at regulating itself, but such regulations you apply to your business now cannot be applied to others, who may prefer to regulate themselves differently. In this case, producing some generally accepted regulatory practices (which aren't super strict and don't receive harsh government sanctions should they be broken) is a way forward. However, this starts to become a more and more regulated space should this be pursued.

6.3.4. Helpful links

6.3.4.1. Market Fundamentalism — Longview Institute https://www.youtube.com/watch?v=IuPb24eixuM

6.3.5. Common contexts

6.3.5.1. #marketfundamentalism #selfregulation #aiinbusiness #regulation

7. Ethics Washing

7.1. Definition

7.1.1. Ethics Washing (also called ethics-shopping) is when organizations adopt vague ethical frameworks and/or internal policies that signal to policymakers, and/or the public, a commitment to responsible AI development. However, these organizations' frameworks and policies often do not entail any genuine obligation or accountability process. Ethics-washing is therefore used to appease regulators that the ethical obligations they seek to impose are not necessary and thereby avoid deeper public scrutiny. Ethical policies developed in this context (especially in the private sector) are less concerned with practical ethical frameworks as they are with their own political goals.

7.2. Relevance in AI Ethics

7.2.1. Policy makers believe it is important for companies to strive for ethical AI development; however, it is very difficult to identify what an effective ethical AI framework looks like. The conversation in AI ethics revolves around whether voluntary ethical approaches are genuine and effective. The use of ethics washing is particularly evident among companies that advertise their policies and investment in ethical behavior and yet fail to follow through with those policies or create internal governance mechanisms to police offending behavior. As a result, it is clear that those companies are using their ethics principles simply to distract the public and prevent people from looking into the company's practices. Furthermore, since even genuine ethical commitments are difficult to entrench across the whole organization, those ethical commitments can amount to ethics washing if there is little enforcement of the company's own standards.

7.3. Example

7.3.1. Google's DeepMind has been considered a leader in the ethical AI field and has even established its own Ethics & Society department to uphold their stated priority, which is ethics. However, DeepMind was involved in the illegal breach of 1.6 million people's health data in a project they undertook with the UK's National Health Service. DeepMind's co-founder Mustafa Suleyman wrote of the scandal in a blog post, stating, "We were almost exclusively focused on building tools that nurses and doctors wanted, and thought of our work as a technology for clinicians rather than something that needed to be accountable to and shaped by patients, the public and the NHS as a whole. We got that wrong, and we need to do better.” Clearly, internal ethical frameworks are not sufficient to mitigate the need for external oversight and regulation.

7.4. Helpful Links

7.4.1. https://www.privacylab.at/wp-content/uploads/2018/07/Ben_Wagner_Ethics-as-an-Escape-from-Regulation_2018_BW9.pdf

7.5. Common Contexts

7.5.1. #easylaw #softlaw #self-regulation #education #legallybinding

8. Bias

8.1. Definition

8.1.1. AI systems are taught to perform their function using machine learning algorithms, which train on data. Often times, that data has either intended or unintended biases encoded within it. As a result, the algorithm will learn to perpetuate that bias in its decision making to an extent that may amount to algorithmic discrimination. An AI is biased if it makes decisions that favor or penalize certain groups of people for reasons that are discriminatory or for factors that are spuriously correlated with the outcome. Bias includes, but is by no means limited to, unfair discrimination on the basis of race, gender, or sexuality. Biases can arise from: i) non-representative training data sets, which result in favorable treatment for the better represented group and worse treatment for the unrepresented group; ii) non-robust datasets, which fail to distinguish between different groups and results in the uniform treatment of people without considering important differences; iii) poor model validation which may perform well on the data collected but yield non-generalizable associations, and iv) encoded bias wherein human biases are reflected in the dataset.

8.2. Relevance in AI Ethics

8.2.1. Bias is a very popular topic in the field of AI Ethics. The reason for this is because bias in AI systems can result in a wide range of highly impactful negative consequences. For example, lots of facial recognition technology is accused of being biased because it has been predominantly trained on white male faces. As a result, when facial recognition technology is used on the general population, the technology is likely to misidentify (or not recognize at all) faces that are not white and male. The failure of this technology has resulted in devastating consequences for people like Robert Williams who was wrongly arrested following an inaccurate FRT identification. Even when the faces of people with other skin tones are included in the dataset, the disproportionate nature of the input data will continue to result in faulty FRT identifications. In order to overcome the risk of bias, some AI systems need more data from each ethnic group to be able to more accurately identify individuals. However, if the data itself reflects bias within the community (more men are represented in STEM disciplines which leads AI to perpetuate discrimination against women) bias cannot be overcome with greater amounts of data.

8.3. Example

8.3.1. A natural-language processing (NLP) algorithm is tasked with answering questions about the tens of thousands of books uploaded into its database. If the algorithm was only trained to converse in English, it would perform a lot better when answering questions about the English books than it would when answering questions about the books written in a different language. As a result, the algorithm might be biased against authors of non-English language books since the NLP fails to adequately respond to questions about their books. This bias can be highly problematic if users decide not to buy books about which the chatbot cannot converse.

8.4. Helpful Links

8.4.1. https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/905267/Facial_Recognition_Technology_Snapshot_UPDATED.pdf

8.5. Common Contexts

8.5.1. #facialrecognition #aiethics #discrimination #fairness #proxyvariables

9. Artificial Intelligence

9.1. Machine Learning

9.1.1. Deep Learning

9.1.1.1. Definition

9.1.1.1.1. Deep Learning (DL) is a Machine Learning (ML) technique that simulates how human beings learn. This learning technique uses what computer scientists call "neural networks" to make predictions. Similar to how neural networks work in the human brain, AI neural networks use multiple layers of processing units that communicate between one another and prioritize variables, which inform the computer's prediction. For example, when considering whether an image of a cat is in fact an image of a cat, the computer might prioritize the look of the eyes to the shape of the tail. This process requires extremely large datasets, is computationally intensive, and can lead to incredibly accurate outputs.

9.1.1.2. Relevance in AI Ethics

9.1.1.2.1. Like humans' neural networks, computers' neural networks cannot be seen. However, unlike humans, computers cannot explain which variables it considered when making its decision. This dilemma has been deemed the "explainability problem" and it points to the black box nature of algorithms. Without insight into how the AI made its decision in the first place, it is difficult to challenge the decision or change the computer's mind, so to speak. Furthermore, it's hard to know whether the computer made its decision based on racist, homophobic, or sexist data and values.

9.1.1.3. Example

9.1.1.3.1. Deep Learning is often used in image recognition. For example, when teaching a Deep Learning algorithm to recognize a cat, one might "feed" an algorithm many pictures of a cat and other animals that look similar. Through trial and error, the Deep Learning algorithm would learn which features are most relevant to determining whether an image contains a cat. Those relevant features are then prioritized in the computer's neural network and heavily considered in the computer's subsequent decision making.

9.1.1.4. Helpful Links

9.1.1.4.1. https://www.youtube.com/watch?v=-SgkLEuhfbg&feature=emb_logo

9.1.1.4.2. Deep learning series 1: Intro to deep learning

9.1.1.4.3. Simple Image classification using deep learning — deep learning series 2

9.1.1.5. Common Contexts

9.1.1.5.1. #datarobustness #bigdata #neuralnetworks #blackbox

9.1.2. Definition

9.1.2.1. Machine Learning (ML) is a technique within the field of Artificial Intelligence (AI). Unlike AI, whose definition is more theoretical, ML provides a more technical explanation to describe how the computer is making its predictions. When someone says that their AI uses ML, it means that their machine is teaching itself to recognize patterns. Machines don't just teach themselves to recognize any old pattern, they teach themselves using a very specific dataset to recognize patterns within that data. Based on the data it observes, and using its own statistical tools, ML adapts its own algorithms to improve the accuracy of its patterns detection. The ML process allows the computers to continue to learn on new input data and to continue to derive a meaningful and relevant output. This process can be compared to how humans learn about horses. We have an initial dataset when it comes to horses, which may include seeing a couple horses in the wild or seeing pictures online, and from this dataset, we feel that we are in a good position to determine whether future animals that have a tail and hooves are in fact horses. However, when we get data about horses that differs from our initial dataset (e.g. that a pony is also a horse) we will refine our belief about horses and, in the future, will be able to determine more accurately what is a horse without getting stumped by horses of a different size and weight.

9.1.3. Relevance in AI Ethics

9.1.3.1. The data collection that is necessary to "fuel" the Machine Learning Model present a host of ethical questions, which arise based on how the data is obtained, how it is used to train the model and how it is deployed. Ethical questions include, but are by no means limited to: whether that data is collected with the consent of individuals, whether that data, either outright or by proxy, includes information about an individual being part of a minority group, whether the data set is robust enough to make consistently accurate decisions, whether the AI makes decisions that perpetuate bias, racism, etc.

9.1.4. Example

9.1.4.1. Facebook's Home Page uses Machine Learning to post content that it predicts would be of most interest to you. Facebook's Machine Learning makes this prediction based on the data it has collected about you, including the content you like and what you're tagged in. The ML model improves its predictive capacity over time as it observes which content you spend more time reading and what content you swipe right past.

9.1.5. Helpful Links

9.1.5.1. Machine Learning Tutorial for Beginners

9.1.6. Common Contexts

9.1.6.1. #prediction #algorithm #supervisedlearning #unsupervisedlearning

9.2. Algorithmic pricing

9.2.1. Definition

9.2.1.1. Algorithmic pricing is the practice of automatically altering the listed price of a good or service as a function of available data. We can think of the price displayed as an output of some function; defined by a set of input variables and parameters. The parameters of the function control the importance of each input variable in setting the end price. The input variables and parameters may be pre-set by an engineer, or determined by an algorithm such as a neural network or decision tree. Algorithmic pricing falls into two closely-related categories: dynamic pricing and personalized pricing. Roughly speaking, dynamic pricing relies on general factors such as demand for the product, time of year/day, or location to determine the price. In contrast, personalized pricing uses information about the specific consumer, perhaps by aligning them with a group on which data about spending habits and preferences is available. Algorithmic pricing falls into two closely-related categories: dynamic pricing and personalized pricing. Roughly speaking, dynamic pricing relies on general factors such as demand for the product, time of year/day, or location to determine the price. In contrast, personalized pricing uses information about the specific consumer, perhaps by aligning them with a group on which data about spending habits and preferences is available.

9.2.2. Relevance in AI ethics

9.2.2.1. The increased prevalence of online shopping; collection of consumer data through browser cookies and social networks; and widespread use of machine learning algorithms have all made algorithmic pricing easier to implement and more profitable. Some important ethical considerations include: 1) Whether businesses must ask a consumer for explicit consent before their data is used for pricing purposes, 2) What information is fair or unfair to use for price setting, and 3) Ensuring that the practice does not facilitate price gouging.

9.2.3. Example

9.2.3.1. In 1999 Douglas Ivester, then CEO of Coca-Cola, proposed adopting vending machines which would set the price of a drink in proportion to the surrounding temperature. His argument was that the utility of a cold drink is greater on hot days, and this should be reflected in the price. This is an example of dynamic pricing, since the only determining factors are environmental (i.e. temperature). The proposal generated outrage and accusations of price gouging from consumers.

9.2.4. Helpful links

9.2.4.1. https://link.springer.com/content/pdf/10.1007/s10551-019-04371-w.pdf https://www.nytimes.com/2005/06/27/business/why-variable-pricing-fails-at-the-vending-machine.html

9.2.5. Common contexts

9.2.5.1. #businesspractices #machinelearning #consumerprivacy #dynamicpricing

9.3. Definition

9.3.1. Artificial Intelligence (AI) is a term used to describe computer systems that perform tasks and functions, which were once thought to be the exclusive domain of intellectual living beings (e.g. able to recognize faces). AI has been designed to optimize its chances of achieving a particular goal. The goals of computer systems can be quite similar to that of humans, including optimized learning, reasoning and perception. Or, computers can be designed to optimize for capabilities that exceed what's possible for humans, including finding variables that have the most influence on an outcome (e.g. AI might determine that height has the biggest influence on basketball performance) . Although "Artificial Intelligence" has retained its definition over time, examples of this technology have changed as the computer's ability to mimic human thought and behavior advances. For example, it was once believed that calculators were an example of Artificial Intelligence. However, over time, this function has been taken for granted as an inherent computer capability and not evidence of artificial intelligence. Currently, more advanced technologies such as self-driving cars and translation are cited as examples of "Artificial Intelligence".

9.4. Relevance in AI Ethics

9.4.1. Artificial Intelligence has been famously described as a "prediction machine". The term references AI's ability to use large amounts of data and, based on a pattern it finds, make inferences about the likelihood of future events. The common ethical concerns that arise as a result of this process include: data rights, privacy, security, transparency, accountability, explainability, robustness, and fairness. On top of these concerns, which point to the functioning of the technology, there are also ethical questions surrounding the trade-offs that have been made when this technology is implemented. For example, the accuracy and productivity of AI has already given it a competitive advantage in the workplace. Therefore, if a company implements AI technology it may be to the detriment of human employment which has socio-economic consequences.

9.5. Example

9.5.1. Voice Assistants, such as Siri and Alexa, are prime examples of Artificial Intelligence. This technology is capable of mimicking human behavior in terms of understanding language, "thinking" about a relevant response and translating that response into speech. Although human communication was once thought to be the exclusive domain of humans, computer programs have also become capable of performing this function. Thus, we call it "Artificial Intelligence".

9.6. Helpful Links

9.6.1. https://www.youtube.com/watch?v=nASDYRkbQIY&feature=emb_logo

9.7. Common Contexts

9.7.1. #AIethics #robotics #futureofwork #productivity #labormarket #databias #racistmachines

9.8. Open source

9.8.1. Definition

9.8.1.1. Open source denotes both how the programming/software involved in an AI is set up, as well as a set of values. In terms of the setup, open source software usually involves the sharing of a base code at the algorithmic level with the knowledge/intention of it being seen, tweaked and used by people who have no relation to you or are not part of your team. In terms of values, open source advocates for how information and source code material should be accessible for all those who want to use it (just like the internet being free). However, this doesn't mean that the software is free in terms of cost, with some companies charging users for access to the code. What does make a line of code open source is that when it is in fact accessed, the code can be tweaked and modified for different uses which the original author can keep track of. Given this, the author can see how the code is being tweaked and learn from such adjustments and thus improve their own coding skills. What this also means is that the author can view how the code is evolving and, given that each programme comes with legal rights for the author, legally put a stop to any undesired evolution of the code should they wish.

9.8.2. Relevance in AI Ethics

9.8.2.1. Within the world of AI ethics, open source brings some invaluable attributes as opposed to privately owned programmes. For example, due to the greater level of accessibility, startups looking to enter into the realm of AI are able to get up to speed and start making a difference faster through having a base source code to work from. As a result, open source serves to reduce the barrier to entry in the tech business and allows for very much needed collaboration by people from all different walks of life. The ability for a wider range of people to tweak the code can help to account for the variety of different experiences that different people can offer, as well as improve security through increased scrutiny of the base code. A valuable feedback loop can then arise, allowing practitioners to be able to help each other out within the domain and strive towards eradicating the perilous issues that flawed code can bring. One example of this is practitioners being able to collaborate over the development of autonomous vehicles (AVs). Due to AVs requiring tonnes of data points from all different aspects of the driving experience, acquiring such data sets and maintaining them is greatly aided by having different collaborators refining, cleaning up and adding to the dataset at hand. This also allows everyone to keep up-to-date in a constantly evolving field.

9.8.3. Example

9.8.3.1. While open source is great in terms of allowing different point of collaboration between different people, as well as the business benefits it brings, there still remain some perils if you use it incorrectly. For example, some believe that open source code is whatever code that is displayed on a public domain (such as GitHub) and thus free to use. However, each public domain has its own licensing agreements and rules of governance, while each author of code has varying degrees of rights over their code. For example, if Paula wrote a programme on GitHub and left it there without caring whether it was repurposed by someone else or not, she would still have the right to sue Pablo if he repurposed her code in a way she was not happy with. In this sense, while open source code is there for wider use, even its wider use still has rules.

9.8.4. Helpful links

9.8.4.1. Opensource.com | Opensource.com

9.8.5. Common Contexts

10. Neural Networks

10.1. Definition

10.1.1. Artificial neural networks (or simply 'neural networks') get their name from being structured somewhat similarly to the neurons in the human brain. Neural networks are used to perform Deep Learning. The name 'neural network' designates a Machine Learning model that has hundreds of thousands (sometimes even billions) of artificial "neurons" (which are actually perceptrons). Each perceptron receives an input in the form of a number, and has an assigned weight within the overall model. Perceptrons are usually arranged in numerous layers, and each is directly connected to the ones in the layer before it and after it. As the model is trained, the weight assigned to each perceptron can be adjusted to improve how accurately the model performs its task. There are different types of neural networks (e.g. convolutional or recurrent). The features of each neural network type render them more or less useful depending on the specific task or application.

10.2. Relevance in AI Ethics

10.2.1. There are two prominent ethical concerns surrounding neural networks. First, neural networks create very complex Machine Learning models. As a result, the model's outputs are often not explainable. This means that us humans don't fully understand, and cannot clearly explain why a Machine Learning model, which use neural networks, gave a particular answer or outcome. This becomes particularly problematic when a person wants to contest a decision made by an AI system that uses neural networks. One may want to contest an AI's decision if they are arrested through a facial recognition match, or if they are denied a loan by their bank. In addition, although neural networks have allowed for outstanding progress in computer science, this progress often leads to greater risk. For example, neural networks have allowed for articles to be written by computers that are perceived to be written by humans. Although this signifies progress in the realm of computer science, this capability can spread misinformation online at unprecedented rates.

10.3. Example

10.3.1. Neural networks are often compared to neurons in the human brain. Our brains are constantly reacting to stimulus. What this means is that information is being transported to our neurons, which results in our neurons firing and triggering other neurons to fire. The precise way that our neurons fire dictate our responses to external stimuli. Similarly, the perceptrons that are activated in artificial neural networks fire in a pattern, which is dictated by the "weights" the computer system has determined. It is through this process that the model derives its outputs.

10.4. Helpful Links

10.4.1. How Convolutional Neural Networks work

10.4.2. How Convolutional Neural Networks work

10.4.3. Neural networks

10.5. Common Contexts

10.5.1. #deeplearning #machinelearning #imagerecognition #explainability #transparency #blackbox

11. Proxy Variables

11.1. Definition

11.1.1. Proxy indicators are seemingly neutral data points about an individual which, in practice, reveal sensitive information about that individual. Proxy data does this by virtue of serving as a “proxy” for another variable (i.e. although explicit race or gender data, for example, is not collected, the use of ZIP codes, grade point averages, credit card purchase information, etc. can serve as a proxy for data about race, gender, sex, sexual orientation or religion).

11.2. Relevance in AI Ethics

11.2.1. On the surface, AI systems that do not collect data about an individuals' protected class status may be perceived as fair - if those data points aren't collected, the AI can't make racist or sexist decisions, right? Unfortunately this is not the case. There are other variables that can disclose sensitive information by proxy and give rise to biased algorithms. Therefore, it is not sufficient for a system to simply not have data about an individual's sex, gender, racial identity or sexual orientation but rather, the system must also demonstrate a lack of proxy for those data points.

11.3. Example

11.3.1. Amazon deployed an AI-based hiring tool that sifted through candidate's applications and recommended applicants to Amazon's talent team. The model worked by prioritizing candidates with similar applications to those that Amazon hired in the past. After Amazon stopped feeding the algorithm information about the candidate's gender, to prevent the AI from perpetuating bias against women, developers found that the AI began to favor candidates who described themselves using verbs commonly found on male engineers’ resumes, such as “executed” and “captured". This case is evidence of discrimination by proxy variable.

11.4. Helpful Links

11.4.1. https://towardsdatascience.com/how-discrimination-occurs-in-data-analytics-and-machine-learning-proxy-variables-7c22ff20792

11.5. Common Contexts

11.5.1. #bias #discrimination #injustice #fairness

12. Supervised Learning

12.1. Classification

12.1.1. Definition

12.1.1.1. Classification is one approach to machine learning. Classification teaches machine learning models to classify input data into designated categories. As a result, the machine's output is a determination of which category the input data belongs. In AI, classification is achieved using different algorithms (e.g. Decision Tree, Random Forest, Naive Bayes, and Support Vector Machine). There are four main types of classification tasks: i) Binary Classification (2 class types); ii) Multi-class Classification (more than 2 class types); iii) Multi-label classification (2+ class types and the model predicts multiple class types); iv) imbalanced classification (uneven distribution of items being classified into different class types).

12.1.2. Relevance in AI Ethics

12.1.2.1. The classification technique is vulnerable to adversarial attacks. These adversarial attacks can have ethical implications in as far as they trick the model into performing poorly. This occurs when data put through the model is on the margins of classification types. This can lead to misclassification with various implications ranging from mild (spam mail not being detected) to severe (inappropriate content on Youtube being recommended to children).

12.1.3. Example

12.1.3.1. Classification can be compared to the process we undertake when sorting our recycling. During the recycling process, we group our plastic, cardboard and glass recycling together and match that recycling to the appropriate bin. In this scenario, we are classifying our recycling items by putting them into the appropriate bin.

12.1.4. Helpful Links

12.1.4.1. 4 Types of Classification Tasks in Machine Learning

12.1.5. Common Contexts

12.1.5.1. #machinelearning #adversarialattacks

12.2. Regression

12.2.1. Definition

12.2.1.1. Regression analysis is an approach to machine learning that teaches a machine learning models to predict a value based on the relationships between the data points it is trained on. Once the machine is able to understand how, for example, the size of my house affects my house's retail price, the machine can make quantitative predictions about the retail price of other homes based on their size. In order for this to work, variables in the data must have some relationship with the outcome. For example, regression analyses cannot tell you whether a picture contains a cat or a dog because classifying the photo as a cat or a dog is not a quantitative prediction problem. Rather, regression analyses can predict an individual's height given variables such as age, weight, and geography. In order to forecast, statistical models are used to predict a value based on the relationships between the variables in the training data. The most common type of regression analysis performed is a 'linear regression'. Performing a linear regression analysis involves characterizing the relationship between variables as a line of best fit. You can imagine this line of best fit connecting various points on a graph to determine the line (slope and y-intercept) that best characterizes the relationship between the incoming data points and the AI's prediction.

12.2.2. Relevance in AI Ethics

12.2.2.1. There are three major AI Ethics problems associated with regression analyses: 1) biased input data, which overvalues independent variables with low levels of predictive value (e.g. when predicting the risk of individual drivers, the independent variable of zip codes is overvalued relative to an individual's driving record. As a result, interest rates on car insurance are higher for individuals living in minority neighborhoods regardless of their driving record); 2) poor inferences about variables that, while correlated, are not causally connected (e.g. using facial recognition technology to determine an individual's IQ); and, 3) algorithms that perpetuate discrimination (e.g. extending credit based on zip code, which correlates with race and minority status and can amount to redlining).

12.2.3. Example

12.2.3.1. Regression analyses are designed to solve quantitative problems. For example, given an individual's age and employment status, how much time will they spend on YouTube in one sitting.

12.2.4. Helpful Links

12.2.4.1. https://towardsdatascience.com/supervised-learning-basics-of-linear-regression-1cbab48d0eba#:~:text=Regression%20analysis%20is%20a%20subfield%20of%20supervised%20machine,that%20someone%20will%20spend%20watching%20a%20video.%202.

12.2.4.2. Introduction to Statistical Learning

12.2.5. Common Contexts

12.2.5.1. #lineofbestfit #quantitativepredictions #linearregression #bias #racism #discrimination #correlationisnotcausation

12.3. Definition

12.3.1. Supervised learning is a technique used to train all types of artificial intelligence (i.e. machine learning and neural networks). This approach relies on the software's pattern recognition skills. It works by teaching the algorithm which data is associated with which label. Labels are simply tags associated with a particular data point that provides information about that data. This technique teaches algorithms to recognize data inputs and to "know" the corresponding data output, or label. For example, supervised learning is used to teach machine learning systems to recognize pictures of cats. It works by providing the algorithm photos of cats with corresponding labels that say "cat". With enough training data, the computer is able to recognize future photos of cats and provide the "cat" label on its own. The algorithm can even assign a probability of having successfully labelled the data input it was given. In order to perform this functioning, supervised learning leverages algorithms such as classification and regression.

12.4. Relevance in AI Ethics

12.4.1. In order to assess the ethics of a particular algorithm, we must be able to understand how an algorithm derived its output. However, black box supervised learning models such as complex trees and neural networks lack interpretability and transparency. These black box supervised learning models are increasingly common and are used to observe the connection between two variables that do not have a linear relationship. Transparency is important not only to validate that the model works but also to ensure that the model is not biased. For example, black box algorithms may determine consumer credit risk based on age, race etc. and without the ability to determine how an algorithm made its decision, it cannot be assured that an algorithm's output is not racist, homophobic, sexist etc.

12.5. Example

12.5.1. If you want your machine learning model to predict the time it will take you to get to work you might teach your model using supervised learning. With the supervised learning model, you would feed your algorithm information that is relevant to the length of your commute such as weather, time of day and chosen route as well as the time it takes you to get to work. Once the algorithm has been trained, the algorithm can recognize the relationship between data points and be able to predict how long it will take you to get to work based on new input data about the weather, time of day and chosen route. The machine may also see connections in the labeled data that you may not have realized. For example, the machine may able to detect that one route will take you longer than another only at a particular time and with particular weather conditions.

12.6. Helpful Links

12.6.1. Supervised Learning: Crash Course AI #2

12.6.2. https://arxiv.org/pdf/1612.08468.pdf#:~:text=black%20box%20supervised%20learning%20models%20%28e.g.%2C%20complex%20trees%2C,regard%20is%20their%20lack%20of%20interpretability%20or%20transparency.

12.6.3. 6 Types of Supervised Learning You Must Know About in 2020

12.7. Common Contexts

12.7.1. #teachingalgorithms #labeling #blackboxalgorithms #bias

13. Reinforcement Learning

13.1. Definition

13.1.1. Reinforcement learning is a way of teaching algorithms using positive and/or negative reinforcement. The two components of reinforcement learning are: i) the agent (aka the algorithm) and, ii) the environment in which the agent finds itself (aka the features of the space that define the agent's options). The agent follows a 3-step process to learn in this way. First, it observes the state of the environment before any action is taken. Second, it decides on a course of action within the environment. Third, the agent is met with a reward signal (a number) to reflect how good or bad the agent's action was. Here, the agent wants to maximize its cumulative reward over time, so it learns, based on feedback, to behave in ways that achieve the highest score. As a result, the algorithm is said to be making decisions according to its "policy" (the methodology the agent has learned in order to make its decision). The agent may "exploit" its existing knowledge and continue the same approach as before or it may "explore" new options to achieve a higher reward. If the agent finds a more efficient path using exploration, a new policy will replace the old one. The actions that an algorithm can take are limited by the author of the environment (data scientist), denoting what's called the "action space" (what actions the agent can take). For example, some "action spaces" are more discrete than others (have a more limited number of possible actions) based on the desired outcomes of the author. For example, in the traditional game of GO the agent, just like a human player, is only permitted to make certain types of moves.

13.2. Relevance in AI Ethics

13.2.1. In reinforcement learning, only the final action triggers the reward or punishment, rather than any intermediate action that the agent took. As a result, there is a risk that the agent will pursue "reward hacking". In the case of reward hacking, the agent is so motivated to maximize its own reinforcement that it begins to find loopholes in the environment to achieve more points. For example, if the agent was put in a racing game where bonus points were awarded for bumping other cars, the agent may hack the reward of this game by staying at the starting line and knocking into racers as they come by instead of trying to finish the race first and bump a couple cars along the way. In this sense, the ethical concern is centered on the consequences of such loopholes being found and exploited. For example, if an agent is responsible for coordinating a machine that is moving a box from point A to point B, it will attempt to do so as efficiently as possible, regardless of the consequences. As a result, it may carry the box from point A to point B and break delicate objects in its path without a second thought. Therefore, ill-defined objective and constraints when teaching AI using reinforcement learning can have serious consequences. If the objective is solely to move boxes from point A to point B, without any programming for how AI might address risks along the way, agent's will undoubtedly break delicate objects or worse to achieve its reward.

13.3. Example

13.3.1. Many self-driving cars are taught to drive using reinforcement learning. In this case, self-driving cars rely on IoT sensors for their blind spots, IoT connectivity for in-car weather and maps as well as software algorithms to determine the best course of action. Autopilot systems use cameras to collect data on the environment, which the agent uses to make decisions during navigation.

13.4. Helpful Links

13.4.1. Reinforcement Learning: Crash Course AI#9

13.4.2. AlphaGo: The story so far

13.5. Common Contexts

13.5.1. #selfdrivingcars #rewardhacking #agent #environment #AlphaGo

14. Unsupervised Learning

14.1. Definition

14.1.1. Unsupervised learning is a way of training machine learning algorithms or neural networks to recognize patterns in an unstructured and non-labelled data set. Unstructured or non-labelled data sets are those in which data is not pre-defined or pre-organized. As a result, it is up to the unsupervised learning algorithm to organize the data. The unsupervised learning algorithm structures the data by finding similarities in the data and grouping those similar data points together. To achieve this, the unsupervised learning algorithm uses techniques, such as K-means clustering. These algorithm can detect anomalies in the data because those anomalies cannot be grouped with similar data.

14.2. Relevance in AI Ethics

14.2.1. The "black box" nature of unsupervised learning algorithms are what makes them concerning from an AI ethics perspective. Since the unstructured data is not pre-labeled, the algorithm has to label the data itself, producing the labels it believes to be most accurate. Hence, due to the "black box" nature of the algorithm, there is a lack of oversight into the algorithm's decision making, with potentially harmful implications.

14.3. Example

14.3.1. When thinking about the difference between supervised and unsupervised learning, consider students learning at school. In math class, the teacher needs to explicitly train the student how to perform various math equations. This process is quite hands on with the teacher instructing the students in highly explicit ways, which can be compared to the way in which a supervised learning algorithm learns. Conversely, during lunch time, students learn by observation how to successfully socialize with their peers. This can be compared to the way in which an unsupervised learning algorithm learns, without explicit instruction and through its own observation.

14.4. Helpful Links

14.4.1. https://intellipaat.com/blog/supervised-learning-vs-unsupervised-learning-vs-reinforcement-learning/

14.4.2. Unsupervised learning: the curious pupil

14.4.3. Unsupervised Learning: Crash Course AI #6

14.5. Common Contexts

14.5.1. #machinelearning #blackbox #unsupervisedclustering #K-meansclusteringalgorithm

15. Differential Privacy

15.1. Definition

15.1.1. Differential privacy is a technique that allows an AI system to learn from an aggregated dataset without compromising on individual privacy. The process can be compared to learning about a community without learning about individuals within that community. This process involves adding "noise" to the dataset such that, once aggregated, the data can demonstrate meaningful information about the group but cannot be cross referenced with a different dataset to infer information about a single individual. This process allows AI systems to overcome privacy concerns by virtue of the aggregated data and the additional noise. According to the pioneer of differential privacy, Cynthia Dwork, "differential privacy is a definition of privacy that is tailored to privacy preserving data analysis".

15.2. Relevance in AI Ethics

15.2.1. Differential privacy is a creative way of preserving privacy without compromising on the robustness of a dataset. Furthermore, differential privacy can maintain the dataset's confidentiality if the system is ever hacked. Unlike anonymizing data, differential privacy adds noise that cannot be removed by cross-checking the dataset with another database. This technique is considered a win in the field of AI ethics since data privacy is a common concern among AI projects. However, the process of training a model using differential privacy also risks reducing the accuracy of the model's performance, which, depending on the application, may present a host of new ethical questions.

15.3. Example

15.3.1. Apple has improved user experience using differential privacy to gain insight on what users are doing without compromising their privacy. This technique has allowed Apple to gain insight into the words that are trending among its clients in order to improve the relevance of its Autofill suggestion.

15.4. Helpful Links

15.4.1. Privacy and accuracy: How Cynthia Dwork is making data analysis better - The AI Blog

15.5. Common Contexts

15.5.1. #federatedlearning #homomorphicencryption #privacy

16. Chatbots

16.1. Definition

16.1.1. Chatbots can come in various forms. However, the common thread between them is that they are an automated system that engages in online dialogue with at least one other human, where the majority of its actions take place without the explicit involvement of a human agent. For example, website and banking assistants can be classed as chatbots due to their engagement in online dialogue with at least one other human, with the majority of their output not requiring human intervention. This allows for a separation between a chatbot and a social bot, where a social bot is more about the automated process of disseminating information, as well as forming that basis of sending spam emails. In this sense, Twitter and Facebook automated trolling accounts are to be classed as social bots, for they do not explicitly engage with a human in dialogue, but rather their purpose to spread information.

16.2. Relevance in AI Ethics

16.2.1. In terms of AI ethics, the main query surrounding chatbots is their potential for deceit, as well as their scalability. Chatbots are starting to be able to reach the level that they can convince some human agents that they are in fact human, leading to phishing scams alongside human emotional manipulation. For example, there is now the possibility of social media accounts being decorated with all the hallmarks of a regular account, and thus engaging other humans in dialogue to develop a relationship to then manipulate in order to receive banking details or passwords. In this way, chatbots ability to deceive should not be underestimated, especially through its use of developing a false sense of trust. What is probably more alarming is then how this can be scaled to millions of conversations happening simultaneously at any one minute. Chatbots are able to engage in a multitude of conversations with the same programmed responses, increasing the odds in its favour of finding one or two humans willing to give up personal details through its increased audience. As a result, chatbot regulation has emerged in order to combat this, which can be found in the example below!

16.3. Example

16.3.1. Paradoxically, the previous cornerstone of measuring AI intelligence was based in the Turing Test, aimed at testing whether machines could imitate being human, and consequently convince an actual human that they're not in fact a machine. If they could do so, the machine was "intelligent". However, in 2019, the California Law in California emerged in order to better regulate bot practice and render this form of intelligence illegal. Here, both chatbots and social bots are viewed as having the potential to deceive human agents into believing something which isn't true (such as a conspiracy theory being disseminated by a social bot, or believing that a chatbot is in fact a real human). In this way, the California law makes it such that both sets of bots must have an explicit statement (whether in its account bio or at the start of a conversation) that it is in fact a bot, not a real person. This is aimed towards helping to prevent the likelihood of humans being lead into a false sense of trust with chatbots, or adding credibility to the statements made by social bots.

16.4. Helpful links

16.4.1. A California law now means chatbots have to disclose they’re not human

16.5. Common Contexts

16.5.1. #chatbot #disinformation #California #automation

17. Monte-Carlo methods

17.1. Definition

17.1.1. Monte Carlo methods are a class of techniques for randomly sampling a probability distribution. There are many problems where describing or estimating the probability distribution is relatively straightforward, but calculating a desired quantity exactly is intractable. This happens all the time and for many reasons, including the stochastic nature of the domain, or there being an exponential number of random variables. Instead, of calculating the desired quantity exactly, it can be approximated by using random sampling, referred to as Monte Carlo methods. Monte Carlo methods vary, but tend to follow a particular pattern: 1. Define a domain of possible inputs, 2. Generate inputs randomly from a probability distribution over the domain, 3. Perform a deterministic computation on the inputs (a decision rule), 4. Aggregate the results (by looking at statistics, like mean, average, and the distribution of the results)

17.2. Relevance in AI Ethics

17.2.1. Monte Carlo methods give the illusion of being unbiased because they rely on sampling random variables. If the variables are random, how can bias arise? In reality there are two sources of bias: the probability distribution, and the decision rule. The choice of probability distribution governs the kinds of solution a monte carlo method will generate. For example, a uniform distribution will generate different random numbers than a normal distribution. The researcher chooses the probability distribution based on intuitions about the problem they're solving– and as with all human intuitions, bias can creep in. The decision rule works similarly– the researcher chooses the kind of decision rule to use. This is another way for bias to creep into the algorithm.

17.3. Helpful links

17.3.1. A Gentle Introduction to Monte Carlo Sampling for Probability

17.4. Common Contexts

17.4.1. #inverseproblems #numericalintegration #functionoptimisation #ensemblelearning

18. Blockchain

18.1. Definition

18.1.1. Blockchain (also known as distributed ledger technology) is a distributed data base network across computer nodes to securely and safely store data. The goal is for digital information to be accessible and distributive, but not alterable. Here, data is stored in blocks (groups of data) with a limited capacity. Once the block is filled, no more data can placed within the block. From there, the block 'locks' and forms a link with the previous block through cryptography to join the chain of already filled blocks. Once this process has occurred, the data within the locked block can no longer be edited. Furthermore, each block has its own unique hash added to the hash of the previous block, as well as its own time stamp for when it joined the chain. Hence, if a hacker was to enter the system and attempt to alter any of the blocks, it would be detectable as the other nodes would not be altered. Any alteration to filled blocks, from a hacker or not, requires consensus from a majority of the users linked to the blockchain account.

18.2. Relevance in AI Ethics

18.2.1. With a focus on AI Ethics, blockchain systems have a strong implication for secure data storage. Instead of having one central server location, data is distributed across multiple locations. In this way, should there be a power outage at one of the locations, not all of the data are in jeopardy. Furthermore, the linear and chronological nature of the chain makes any attempt to steal the data costly, time-consuming and very difficult to do. Blockchain's transparency is also of great use to AI. The provenance of the data the AI is using is clear for all to see. This makes it an ideal candidate for private medical data and financial transactions to be stored on the network and utilised by AI.

18.3. Example

18.3.1. Blockchain is widely used in food distribution tracking. Here, companies such as IBM have implemented blockchain to track food distribution, possessing the ability to identify where different ingredients have come from in case of any food contamination taking place. The technology also forms the basis of the current cryptocurrencies. Bitcoin, non-fungible tokens (NFTs) and the like are stored on blockchain technologies to take advantage of its community ownership framework, as well as its transparency benefits.

18.4. Helpful links

18.4.1. How blockchain and AI link. Blockchain and artificial intelligence (AI) | IBM A good resource on blockchain itself https://www.investopedia.com/terms/b/blockchain.asp .

18.5. Common contexts

18.5.1. #blockchain #data #datagovernance #privacy

19. The Internet of Things (IOT)

19.1. Definition

19.1.1. The Internet of Things (IoT) as a term was first coined in 1999 by Kevin Ashton. At its core, the concept refers to devices connected to the same IoT platform which allows them to share data without human intervention. This network of physical devices and sensors possesses the ability to store and protect information, whereby its processing can take place locally. The IoT has a wide variety of applications from the domestic to the corporate. For example, Amazon's Alexa home range can communicate and share data between each other, or a smart gas meter communicating gas usage to the provider. On the corporate side, there is the presence of the Industrial Internet of Things (IoT), where all purposes are focused on improving the industrial process. For example, sensors can be fixed to various machinery that communicates the performance of the machine in question to the IoT platform, clearly signalling when its operating below optimum efficiency. In this way, the IoT boasts more data and access to a business' own systems, making it easier to affect any changes necessary. Consequently, production systems become more responsive. The IoT, as a result, can also be seen as providing actionable information being compiled by smart devices.

19.2. Relevance in AI Ethics

19.2.1. In terms of AI Ethics, the ultra interconnectedness of the network means there are security issues to consider. For example, a singular devastating hack could open a treasure trove of personal data to steal. Combining all the IoT data available on one person can create a surprisingly detailed image of their habits and personal life. Furthermore, a more subtle intrusion could allow hackers to eavesdrop on conversations in people's homes. The IoT can also offer the a more positive side in the potential to mitigate the risks involved with the use of AI technology. Offering performance analytics in real time can allow a faster response to sorting out any potential issues. In this way, should a problem occur, the consequences can be minimised.

19.3. Example

19.3.1. An example of the IoT can be self-driving cars. Here, all sorts of data is collected whilst the vehicle is on the road, including information about engine performance. Hence, should there be a problem, the vehicle could notify the insurance provider and have them schedule an appointment to fix the issue, all without human involvement.

19.4. Helpful Links

19.4.1. A useful list of examples https://www.ibm.com/blogs/internet-of-things/what-is-the-iot/ How IoT applies to business https://www.zdnet.com/article/what-is-the-internet-of-things-everything-you-need-to-know-about-the-iot-right-now/

19.5. Common Contexts

19.5.1. #IoT #autonomousvehicles #connectivity

20. Anthropomorphism

20.1. Definition

20.1.1. Anthropomorphism relates to the attribution of human attributes to non-human objects. For example, we could be seen to be anthropomorphising an object when we attribute goal-orientated agency to said object, such as saying that 'the leaf decided to move closer to the rock'. In the context of AI, we tend to employ anthropomorphic language to describe the actions of an AI, like how 'the AI decided to produce result X' or 'the AI thought it was best that we take Y plan of action'. From there, anthropomorphism can come both in form (the AI looks like a human) and in action (the AI behaves like a human). In terms of application, anthropomorphism is a technique deployed in order to increase human interaction with a certain technology. By making the technology seem more familiar, it becomes more likely that a human will engage with the technology. Manifestations of this technique can include giving a chatbot a name and gender, constructing a human-like physical embodiment for an AI or describing an AI as a 'thinking, feeling' entity.

20.2. Relevance in AI Ethics

20.2.1. Within the field of AI Ethics, the main issue with anthropomorphism arises in conversations about deception. Championed by the Turing test, human deception was seen as a key component to AI being seen as intelligent. Within the test, if the technology could deceive the human into thinking that it was a fellow human counterpart, the technology was deemed as intelligent. While this has now gone out of fashion as a mark of intelligence, the danger of deception still exists. The potential damage such deception can cause was recognised in the California Chatbot Law of 2019. The law stipulates that any chatbot within the state of California must declare that it is a chatbot within the first instances of interaction with a human agent. One of the core inspirations behind such a law was the potential that chatbots possess to deceive a human both through disseminating false information on an automated basis, as well as through direct one-to-one interaction. The chatbot could in fact be operated by a human pretending to be the chatbot of an official website, such as Woolworth's, in order to acquire the personal details of a human agent or get money off them. Or, it could be that a chatbot is set up to impersonate a human agent. Hence, in terms of AI Ethics, the crucial issue lies in chatbot deception.

20.3. Example

20.3.1. An interesting example of the effect anthropomorphism can have can be seen in the military field. Here, Boomer was a machine designed to detect explosives in the battlefield, saving countless lives. However, the technology eventually met its fate and when it did, it was given a full-on military funeral, as if it was a fellow human soldier. This shows the extent to which anthropomorphism can develop a deep bond between humans and technology, and how such a relationship has the potential to be exploited.

20.4. Helpful links

20.4.1. An interesting article by Heather Knight on human responses to anthropomorphic technology: How humans respond to robots: Building public Policy through good design

20.5. Common contexts

20.5.1. #anthropomorphism #chatbots #deception

21. Large Language Models

21.1. Definition

21.1.1. Large language models (LLMs) such as Open AI's GPT models are language models based on transformer neural networks. These networks are made up of nodes which link together like neurons in the human brain. The transformer element (introduced in 2017 by Google Brain) allows for large datasets to be processed all at once through unsupervised learning (where the model is trained on an unlabelled dataset). Consequently, the model is able to understand the patterns of natural language through learning the dataset. From there, the model is provided with a prompt (such as by a human user) which it uses to continue the conversation in a coherent manner. Here, different speech possibilities are assigned different weights by the transformer to convey meaning (the more weight assigned to a word, the more confident the model is on that word being the most appropriate for the sequence). In GPT models, this weight assignment is carried out by the encoder (for inputs) and decoder (for outputs).

21.2. Relevance in AI Ethics

21.2.1. The ability of LLMs to convincingly convey information (thanks to its grasp of language) boasts a strong possibility of such persuasive ability being used for manipulation. Whether it be producing convincing arguments, or creating a false sense of trust between the model and the user, there is scope for LLMs to be able to trick humans (such as Google engineer Blake Lemoine believing that one of Google's LLms, LaMDA, was sentient. Beyond its capabilities, the content of the datasets that LLMs train on is also noteworthy. ChatGPT has already displayed signs of political bias (such as around president Biden and former president Trump) in some of the answers it has provided. Furthermore, these datasets are being maintained by under-payed workers (such as OpenAI paying Kenyan workers $2/hour to give feedback on the answers it was producing).

21.3. Example

21.3.1. A prime example of the power of these LLMs to proliferate misinformation is when a New York lawyer used ChatGPT to supplement their legal research through finding specific case studies. It turned out that the cases it produced (and that were later submitted to court) were completely made up.

21.4. Helpful links

21.4.1. https://www.forbes.com/sites/mattnovak/2023/05/27/lawyer-uses-chatgpt-in-federal-court-and-it-goes-horribly-wrong/#:~:text=Sadly%2C%20the%20lawyer%20used%20the,hit%20his%20knee%20in%202019 . https://towardsdatascience.com/how-chatgpt-works-the-models-behind-the-bot-1ce5fca96286 https://www.brookings.edu/blog/techtank/2023/05/08/the-politics-of-ai-chatgpt-and-political-bias/

21.5. Common contexts

21.5.1. #llms #aiethics #unsupervised learning

22. Hallucination

22.1. Definition

22.1.1. Large lanuage models (LLMs) optimise for probability in their response to a prompt. Consequently, they do not prioritise factual truth when giving an answer. Instead, they use their training dataset (which is not a dataset of facts, but of text or code in general) to provide the most probable next word in their sentence. As a result, a hallucination is when a LLM provides false information as if it were true. This can come in the form of factually incorrect information, providing non-existent sources, nonsensical answers (such as repeating lots of words in a poem), relating two different concepts that do not correlate, and others.

22.2. Relevance in AI Ethics

22.2.1. LLMs say with as much confidence what is factually true and what is factually false. Hence, the danger of misinformation (especially given the uptake of models such as ChatGPT) is very pronounced given the scale at which such information can be spread, as well as the lack of universal scrutiny over the answers given by LLMs. With this in mind, LLMs can facilitate the dissemination of the biases within the model. Should the model inadequately represent certain cultures, the false information it produces as a result will be widely consumed and believed. In this case, with hallucinations in mind, treating LLMs as infallible runs the risk of such false information being taken as gospel. This can lead to societal and political harms spread on a global scale.

22.3. Example

22.3.1. ChatGPT's hallucinations have already cost OpenAI a lawsuit. Here, the LLM provided false information about Georgia radio host Mark Walters where it accused him of having embezzled, which was untrue. Walters then sued OpenAI for defamation.

22.4. Helpful links

22.4.1. https://towardsdatascience.com/llm-hallucinations-ec831dcd7786 https://www.zdnet.com/article/chatgpts-hallucination-just-got-openai-sued-heres-what-happened/

22.5. Common contexts

22.5.1. #llms #aiethics #hallucination

23. Diffusion models

23.1. Definition

23.1.1. Diffusion models (such as Stable Diffusion, DALL.E 2, Imagen and Midjourney) generate images or audio via input prompts from the user. These images/audios are similar to the ones located in the dataset it has been trained on. They add noise to the available training data through a Markov chain (a set of mathematical functions) in a process called forward diffusion. Then, it takes the noise out (reverse diffusion) to recover the data. Effectively, they destroy the data and then learn how to recover it. In that process, the model learns what is image/audio, and what is noise. Hence, the model can then produce coherent non-noised images/audios (with slight variation) from a noised images. When a user puts in a prompt, the model selects random noised images from the dataset that best match the prompt, and then produces a related non-noised image. The adding of noise underlies 3 different diffusion model categories: denoising diffusion probabilistic models (DDPMs), noise-conditioned score-based generative models (SGMs) and stochastic differential equations (SDEs). DDPMs adds and removes noise from images or sound. SGMs are more-so based around estimating a score function to indicate the log density of an image to then generate calculated assumptions about the data set (used in deepfakes). SDEs are used to generate predictions while taking into account random factors (such as in market calculations).

23.2. Relevance in AI Ethics

23.2.1. With the wide accessibility that diffusion models offer, many people can produce images/sounds on demand. This raises many questions in the art world, mainly focused on dangers of job replacement, as well as copyright issues.

23.3. Example

23.3.1. For example, at time of writing, Getty Images is suing Stable Diffusion in the US for copyright infringement. The company claims that Stable Diffusion used 12 million of its images in its training data without permission. In a similar vein, artists are also complaining that their images have been used in the training of these models without due credit nor permission.

23.4. Helpful Links

23.4.1. https://www.theverge.com/2023/1/17/23558516/ai-art-copyright-stable-diffusion-getty-images-lawsuit https://www.euronews.com/culture/2023/03/27/from-lawsuits-to-tech-hacks-heres-how-artists-are-fighting-back-against-ai-image-generatio

23.5. Common Contexts

23.5.1. #art #diffusion

24. General adversarial network (GAN)

24.1. Definition

24.1.1. A GAN consists of two competing models: a generator and a discriminator. Once trained on a precured dataset, the generator aims to generate images that are not located in the dataset, but look as if they were. The discriminator then tries to determine whether the image generated is part of the dataset (and is therefore accepted) or whether it is not (and is therefore rejected). Once the generator can fool the discriminator with one of its generated images, the process is complete. This comes after much human effort involved in curting the dataset, tweaking the system to generate desired content and post-curation of the results. In the art world, this process is used by artists to create machine learning art.

24.2. Relevance in AI Ethics

24.2.1. Within this process, what is included in the dataset comes under ethical scrutiny. The understandable concerns surrounding how the data was curated, was it obtained with consent and what happens to the images after the desired result is obtained are all questions that can be directed at using a GAN. Furthermore, despite the heavy human touch involved in generating machine learning art, there is also the risk of glossing over this human involvement in favour of AI authorship. The results could lead to viewers deeming the GAN as creative or autonomous, which then potentially leads to legal consequences (such as in copyright).

24.3. Example

24.3.1. Due to the speed at which art pieces can be produced using GANS, some artists have turned to NFTs and blockchain to sell their work.

24.4. Helpful links

24.4.1. https://www.artnews.com/list/art-in-america/features/gans-and-nfts-1234594335/robbie-barrat-ai-generated-nude-portrait/ https://www.stirworld.com/see-features-exploring-gan-art-with-amir-zhussupov

25. Smart City

25.1. Definition

25.1.1. A smart city refers to the growing involvement of data analytics in city living. This involves the implementation of hardware to monitor different aspects of city life in real time. For example, using data to note the optimum time for when street lights should come on and off, or which water pipes need replacing based on need. Such implementation usually involves three layers: the hardware, the visual representation of the data collected, and the layer where the decisions are made. These initiatives aim to attract capital (either supplied by the local authorities or by private companies) in order to improve the running of a city, usually involving some environmental considerations (such as reducing the city's carbon footprint). Consequently, the more data involved, the better the efficiency of the city and the less impact it has on the environment.

25.2. Relevance in AI Ethics

25.2.1. Due to the intense focus on data analysis, questions surrounding the use of citizen surveillance arise. For example, monitoring the electricity usage in certain areas could reveal the night/day routines of the citizens living there without their consent. The organisation collecting this data also comes into question, with private companies potentially selling on this data to other private entities. Even if this data was collected by government entities, the risk of surveillance and data auction are still present.

25.3. Example

25.3.1. The smart city initiative in Toronto by Google-affiiliated Sidewalk Labs started in 2017 was ultimately abandoned in 2020. The official reasons given were the economic effects of the pandemic, but there were also grave concerns held by local residents of private sector encroachment on governance. Furthermore, some believed that the initiative did not pay enough attention to local needs. Instead, it focused on what the technology can do, rather than what it should do.

25.4. Helpful Links

25.4.1. https://www.technologyreview.com/2022/06/29/1054005/toronto-kill-the-smart-city/ https://montrealethics.ai/reports-on-communication-surveillance-in-botswana-malawi-and-the-drc-and-the-chinese-digital-infrastructure-surveillance-in-zambia/ https://montrealethics.ai/ai-ethics-inclusivity-in-smart-cities/

26. AI-text detection tools

26.1. Definition

26.1.1. As AI language capabilities continue to reach higher levels of sophistication, worries about how to decipher between text written by AI technologies (such as large language models) and humans have risen (especially in the education and political sectors). Consequently, AI-text detection tools have been touted as one potential avenue to counter this issue. This typically involves a language model trained on a structured dataset of texts written by both humans and other language models. The detection model is then tested against an unfamiliar dataset of AI-written and human-written texts, attempting to assess whether the text is written by a human or an AI.

26.2. Relevance in AI Ethics

26.2.1. In terms of AI ethics, not being able to tell whether a piece of text is authored by a human or a machine can have negative consequences. Students can submit work written by an AI that is not their own (potentially containing false information), as well as grave potential for the spread of misinformation. For example, using an AI technology to generate text that presents harmful information and attributing said text to a prominent politician. At the time of writing, OpenAI has shut down their AI detection tool due to its heavy innacuracies when attempting to detect AI-written text. In this case, it is not just mistaking AI-written text for human authorship that is an issue, but also determining human-written text as authored by an AI. This has been recorded in some universities in the US, where students are having their papers returned to them classified as AI-generated, especially international students. Hence, the quicker AI-detection tools become more sophisticated, the better for both the education and political sectors.

26.3. Example

26.3.1. Students who are non-native English speakers have had their papers disproportionately flagged for being AI-generated compared to native speakers. These false positives mainly stem from AI-text detectors targetting sentences which are simple and when the next word is predictable. Unfortunately, non-native English speakers' essays tend to fit this bill. As a result, thousands of non-native English speaker students have had their papers wrongly flagged as AI-generated.

26.4. Helpful links

26.4.1. https://openai.com/blog/new-ai-classifier-for-indicating-ai-written-text https://themarkup.org/machine-learning/2023/08/14/ai-detection-tools-falsely-accuse-international-students-of-cheating https://www.theguardian.com/technology/2023/jul/10/programs-to-detect-ai-discriminate-against-non-native-english-speakers-shows-study#:~:text=Tests%20on%20seven%20popular%20AI,students%2C%20academics%20and%20job%20applicants .

27. Computer vision

27.1. Definition

27.1.1. Computer vision aims to allow computers to 'see' through replicating the human vision system in order to analyse and extract information from images and videos. Cameras, data and algorithms are used to analyse visual data as opposed to retinas, cortexes and eye balls. The algorithm trains on vast amount of structured visual data, which allows it to be able to recognise similar images in a new and unseen dataset with unstructured (non-labelled) images. To do this, a convolutional neural network (CNN) involving deep learning is deployed. Deep learning allows the algorithm to recognise images on its own. This comes through how the CNN breaks down the images into pixels, which are then labelled by the CNN. The algorithm then performs 'convolutions' (a mathematical function that acts on two functions to generate a third) to 'see' the image. It then test its prediction using images already stored in its database. Consequently, information can be extracted from pixels. Hence, it plays crucial roles in the self-driving car, virtual reality, health and facial recognition industries.

27.2. Relevance in AI Ethics

27.2.1. Given that computer vision relies so heavily on data and visual input from images, it involves great risk to adversarial attacks and data privacy. Adversarial attacks can be where malicious actors 'poision' the dataset by supplying incorrectly labelled images, which could then lead a self-driving car to misidentify a 30 mph sign for an 80 mph sign. Furthermore, what images computer vision models have access to also matters. These images that form part of their training must have been accessed through the consent of their proprietary owners, where a failure to do so can lead to severe harm to the owner of the image, as well as serious data protection violations.

27.3. Example

27.3.1. As a metaphorical example, you find yourself as a passenger in a self-driving car driving late at night in winter. Suddenly, snow starts to fall and covers the lines separating the lanes. As a result, your car cannot recognise where it is supposed to drive, and comes to a stop, refusing to start again until it can reliably recognise the lane separation lines.

27.4. Helpful links

27.4.1. https://www.ibm.com/topics/computer-vision#:~:text=Computer%20vision%20is%20a%20field,recommendations%20based%20on%20that%20information . https://www.mygreatlearning.com/blog/what-is-computer-vision-the-basics/

28. Compute

28.1. Definition

28.1.1. Compute refers to both the calculations needed to complete a certain task and the hardware involved in doing so (note: compute can sometimes be used to refer solely to calculations or solely to the hardware). These software and hardware components are usually refered to in a stack that can include chips (like GPUs), software, domain-specific languages, data-management software and infrastructure (like data centres). Compute is then measured in terms of floating point operation (FLOP), which are mathematical formulas that can represent large numbers with great precision. Compute performance is then measured in FLOP/second (how many calculations a resource can carry out per second). Definition adapted from the AI Now Institute Computational Power and AI report: https://ainowinstitute.org/publication/policy/compute-and-ai

28.2. Relevance in AI Ethics

28.2.1. The high entry cost associated with the chip manufacturing industry necessary for compute means few actors are able to have any skin in the game. Consequently, this also incentivises these actors (such as cloud manufacturers) to maintain their position in the market and take advantage of the craze over compute power which, given the insatiable desire to release bigger and bigger models, may be ecologically unsustainable. From there, this results in these actors having a strong influence over who gets access to the manufactured chips, meaning they have a say in geopolitical discussions surrounding AI chip manufacturing.

28.3. Example

28.3.1. DeepMind's cofunder Mustafa Sulleyman has argued that US chips should only be sold to businesses who can show ethical intent and responsible use of the technology. Determining what that looks like is the next stage of the process.

28.4. Helpful links

28.4.1. https://www.verdict.co.uk/us-should-only-sell-ai-chips-to-those-with-ethical-intent-mustafa-suleyman-says/ https://ainowinstitute.org/publication/policy/compute-and-ai

29. Overfitting and underfitting

29.1. Definition

29.1.1. Overfitting is where an algorithmic model learns its training data too precisely to the extent that it models itself on the noise present in the dataset, leading to poor performance when confronted with generalising to new datasets (often found in complex machine learning models). Underfitting, contrastingly, is where the model cannot model neither the training data nor generalise to new data (likely caused by an insufficient amount of data points). Both of these issues lead to a negative impact on the performance of the model, which can be measured and monitored when compared against the model's performance. A decrease in performance on the new dataset means that the model has overfitted the training data, which is also reflected in a decrease in performance on the training data set as the model has taken into account the irrelevant noise.

29.2. Relevance in AI Ethics

29.2.1. Given that overfitting means the model trains on the training dataset almost verbatim, should the training dataset contain any biases these will also be baked into the model. From there, should similar training datasets be used for multiple models, this can lead to a composition bias whereby the model generalises untrue inferences from a sample (its training dataset) to its generalised dataset (a population). Furthermore, concentrating on fewer datasets (and overfitting as a result) can create a fake sense of progress, with the model's positive performance on the dataset feigning progress made by the model, when it is in fact performing well as it has learnt the training dataset well.

29.3. Example

29.3.1. Overfitting and underfitting can be explained in reference to taking an exam. Overfitting would be including information in your answer that is not relevant to the question, perhaps meaning you lose marks. Furthermore, fake progress can be seen in the model learning the same training data 'exm' really well, and thus scoring well on that particular paper. Underfitting is then where you do not include the relevant information you should have in your answer, also leading to a loss of marks.

29.4. Helpful links

29.4.1. https://machinelearningmastery.com/overfitting-and-underfitting-with-machine-learning-algorithms/ https://montrealethics.ai/game-changer-or-not-an-evaluation-of-interpretable-machine-learning-models/

30. Griefbots

30.1. Definition

30.1.1. As the result of combining "grief" and "chatbot", a griefbot is a chatbot trained on the digital footprint left behind by a someone who is now deceased. Said training can include data from the person's social media, emails, articles etc. The primary intention of the technology is to allow its user to interact with the person's digital representation after their death to help assuage their grief. Consequently, griefbots are usually digital representations of a now deceased loved one, which the users interact with to help them deal with their grief. Said process can be compared to being a more conversational experience than simply talking to the grave.

30.2. Relevance in AI Ethics

30.2.1. Griefbots prove very wide-reaching in terms of the AI ethics considerations it requires. There are questions of consent (whether the griefbot subjected consented to the process), privacy (programming the griefbot will require an external person viewing some potentially sensitive information), as well as some perverse business incentives. For the latter, this could include companies who supply the griefbot trying to retain active users, meaning users ultimately become addicted to the service.

30.3. Example

30.3.1. The founder of ReplikaAI, Eugenia Kuyda, programmed a griefbot based on 8,000 lines of text messages between her best friend, Roman Mazurenko, and herself after his sudden and tragic death.

30.4. Helpful links

30.4.1. https://www.theverge.com/a/luka-artificial-intelligence-memorial-roman-mazurenko-bot https://www.researchgate.net/publication/359255833_Griefbots_A_New_Way_of_Communicating_With_The_Dead

31. Frontier AI Models

31.1. Definition

31.1.1. Frontier AI Models are an example of foundation models (general AI models capable of performing a wide variety of different tasks to a high level of precision) which match or exceed the capabilities of the present most-advanced AI models. These models are trained on a broad spectrum of data to allow it to be competent at their different tasks. At this point, examples include cutting edge large language models which, consequently, require huge amounts of compute (such as data) and natural resources (such as electricity, water and raw materials) to run.

31.2. Relevance in AI Ethics

31.2.1. These types of models are the subject of current regulatory efforts such as the EU's AI act and the Executive Order on AI issued by President Biden. Related AI Ethics issues, according to the Center for Governance on AI, can include the unexpected capabilities problem (where the abilities of such systems are hard to predict and, thus, hard to regulate), the deployment safety problem (where unintended harm results despite the developer and users' best interests) and the proliferation problem (where nefarious actors can use the models despite not possessing the resources to create them).

31.3. Example

31.3.1. Even the most well-intentioned initiatives can have unintended consequences. For example, as a result of the emergency alert system in the US, contraband technological devices possessed by some of the Amish community members in Ohio went off, leading to those members being shunned. Hence, when you try and design a highly capable and sophisticated frontier AI model with capabilities applicable in many different areas, the likelihood of the designers' intentions translating seamlessly are slim. Thus, the crux of the issue comes in how these almost inevitable unintended consequences can ben mitigated.

31.4. Helpful links

31.4.1. https://www.gov.uk/government/publications/frontier-ai-capabilities-and-risks-discussion-paper/frontier-ai-capabilities-and-risks-discussion-paper#:~:text=For%20the%20purposes%20of%20the,%5Bfootnote%202%5D https://www.governance.ai/post/frontier-ai-regulation https://www.independent.co.uk/news/world/americas/amish-phone-us-government-alert-b2426792.html

32. G.I.G.O

32.1. Definition

32.1.1. G.I.G.O stands for "garbage in, garbage out", which is an old Computer Science and Mathematics adage (usually accredited to George Fuechsel) referring to how the quality of the output of a system depends on the quaity of the input. Sometimes referred to as "R.I.R.O" (rubbish in, rubbish out), if an AI system is based on poor quality data (such as being non-representative, filled with noise etc.), then its output will also be of poor quality.

32.2. Relevance in AI Ethics

32.2.1. When it comes to AI Ethics, one of the main struggles is when the data of a system is biased and, as a result, produces biased and harmful outcomes. Recently, generative AI diffusion models such as Midjourney and Stable Diffusion have given a visual representation of some of these biases: the prompt "Indian person" is nearly always an old man with a beard, a "Mexican person" is usually a man in a sombrero, Indonesian food is almost exclusively shown as being served on banana leaves etc. When this type of stereotyped information is fed into a system, you will get a stereotyped output, providing false and harmful information about different cultures.

32.3. Example

32.3.1. In 2015, Google's ad services showed more prestigious and high-paying jobs when the user identified as male and not female. In 2018, Amazon's hiring algorithm disproportionately disciminated against women when considering them for positions. Also in 2018, Joy Buolamwini's study showed that facial recognition algorithms were woefully inaccurate when identifying black female faces compared to white males faces.

32.4. Helpful links

32.4.1. https://www.independent.co.uk/tech/google-s-algorithm-shows-prestigious-job-ads-to-men-but-not-to-women-10372166.html https://www.reuters.com/article/idUSKCN1MK0AG/ https://news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212 https://montrealethics.ai/target-specification-bias-counterfactual-prediction-and-algorithmic-fairness-in-healthcare/ https://restofworld.org/2023/ai-image-stereotypes/?mc_cid=7632ce2a62&mc_eid=fa6fa6c768

33. AI companions

33.1. Definition

33.1.1. An AI companion typically refers to social AI (or conversational AI) software (such as chatbots, avatars etc.) aimed to establish some form of relationship with a human user. This could range from a platonic friendship to a trusted romantic partner, while varying from taking an anthropomorphic form to a pet. Conversations could include videos, images and text, whereby transformer architecture is typically used in order to create more natural-sounding dialogue (as opposed to traditional decision tree architecture). This form of conversational AI is to be defined away from therapy, mental health and wellbeing chatbots as the core idea of these is not to establish a human relationship (although that could form part of their final goal). Also, this type of companionship differs to business use cases, which may define AI assistants (such as Zoom's AI Companion) as 'companions', given its human relationship development focus.

33.2. Relevance in AI Ethics

33.2.1. In terms of AI Ethics, there are several strands of different fears that come into play when it comes to relationships with AI companions. The major strand is privacy, whereby intimate conversations on AI companion apps (such as Replika and Romantic AI) are not classed as protected data, meaning that companies can sell such data to third parties. A further strand is attachment concerns, where it is feared that users will become so fixated on using the app that it harms their in-person relationships. Stemming from this are similar concerns surrounding people replacement, and social deskilling, where the constant use of such applications leads to a reduction in the social skills possessed by their users.

33.3. Example

33.3.1. The Mozilla Foundation found that none of the companion apps that it reviewed passed the mimimum requirements for its "privacy included in this product" label. Shockingly, all apps reviewed sell data to third parties, or do not provide enough information to say that they do not.

33.4. Helpful links

33.4.1. https://foundation.mozilla.org/en/privacynotincluded/articles/happy-valentines-day-romantic-ai-chatbots-dont-have-your-privacy-at-heart/

34. Parameters

34.1. Definition

34.1.1. Parameters (e.g. coefficients (weights) of linear and logistic regression models) are the 'dials' (weights and biases) of a language model that you can turn to optimise its performance (such as predicting and producing the correct tokens of a sentence). These are deemed 'internal' and part of the model, as opposed to hyperparameters (which determine the learning rate of the model), which are deemed 'external' (they don't form part of the model). With this, you initialise training by setting the hyperparameters and parameters to your desired values, before optimising and adjusting the model to produce the correct outcome (labelled 'parameter or hyperparameter tuning'). If you wanted to fine tune only some of the parameters, this is called 'parameter-efficient tuning'.

34.2. Relevance in AI Ethics

34.2.1. The more parameters housed by the model, the more environmentally costly (producing more carbon, using more electricity, requiring more water for cooling) and expensive the model is to run. Consequently, the biggest (and highest-perfomring) models require a huge amount of resources to run, meaning only certain organisations can afford to do so. OpenAI have released GPT-4o mini to provide a cost efficient alternative, but it holds true for its and its competitors larger models that the environmental cost only grows as models get bigger.

34.3. Example

34.3.1. In the cannonical paper by Strubell, Ganesh and McCallum in 2019, training large language models was estimated to produce around 626,000 lbs of CO2 emissions, more than an average person's use of a car in their lifetime (126,000 lbs).

34.4. Helpful links

34.4.1. https://montrealethics.ai/efficiency-is-not-enough-a-critical-perspective-of-environmentally-sustainable-ai/ https://arxiv.org/abs/1906.02243 https://kelvin.legal/understanding-large-language-models-what-are-paramters/#:~:text=What%20are%20parameters%3F,tokens%20and%20generating%20new%20ones . https://medium.com/m/global-identity-2?redirectUrl=https%3A%2F%2Ftowardsdatascience.com%2Fparameters-and-hyperparameters-aa609601a9ac