Four Recommendations for Ethical AI in Healthcare

Lindsey Jarrett

Center for Practical Bioethics

For several decades now, we have been having conversations about the impact that technology, from the voyage into space to the devices in our pockets, will have on society. The force with which technology alters our lives at times feels abrupt. It has us feeling excited one day and fearful the next.

If your experiences in life are not dependent on the use of technology — especially if your work still allows for you to disconnect from the virtual world – it may feel like technology is working at a decent pace. However, many of us require some sort of technology to work, to communicate with others, to develop relationships, and to disseminate ideas into the world. Further, we also increasingly need technology to help us make decisions. These decisions vary in complexity from auto-correcting our messages to connecting to someone on a dating app, and without access to a piece of technology, it is increasingly challenging to rely on anything but technology.

Is the use of technology for decision making a problem in and of itself due to its entrenched use across our lives, or are there particular components and contexts that need attention? Your answer may depend on what you want to use it for, how you want others to use it to know you, and why the technology is needed over other tools. These considerations are widely discussed in the areas of criminal justice, finance, security, hiring practices, and conversations are developing in other sectors as issues of inequity, injustice and power differentials begin to emerge.

Issues emerging in the healthcare sector is of particular interest to many, especially since the coronavirus pandemic. As these conversations unfold, people start to unpack the various dilemmas that exist within the intersection of technology and healthcare. Scholars have engaged in theoretical rhetoric to examine ethical implications, researchers have worked to evaluate the decision-making processes of data scientists who build clinical algorithms, and healthcare executives have tried to stay ahead of regulation that is looming over their hospital systems.

However, recommendations tend to focus exclusively on those involved with algorithm creation and offer little support to other stakeholders across the healthcare industry. While this guidance turns into practice across data science teams building algorithms, especially those building machine learning based tools, the Ethical AI Initiative sees opportunities to examine decisions that are made regarding these tools before they get to a data scientist’s queue and after they are ready for production. These opportunities are where systemic change can occur, and without that level of change, we will continue to build products to put on the shelf and more products to fill the shelf when those fail.

Healthcare is not unique in facing these types of challenges, and I will outline a few recommendations on how an adapted, augmented system of healthcare technology can operate, as the industry prepares for more forceful regulation of the use of machine learning-based tools in healthcare practice.

The info is here.

The Petrov dilemma: Moral responsibility in the age of ChatGPT

Julian Savulescu

NUS News

Originally published 13 FEB 23

Here is an excerpt:

3. Act ethically

What is to act ethically has been debated for thousands of years. But at its heart is a simple answer: add value to the lives of other people. For Lt-Col Petrov, doing nothing was the hardest choice: “All I had to do was to reach for the phone; to raise the direct line to our top commanders – but I couldn’t move. I felt like I was sitting on a hot frying pan.”

ChatGPT is an extremely powerful technology with enormous potential for human benefit. It can distill human knowledge and wisdom, quickly create factual and normative content, enable communication between speakers of different languages, and ultimately enhance or even surpass human creativity.

As with any powerful technology, there are significant risks. At present, there are numerous mistakes and false claims, but these presumably will be ironed out. Longer term, risks include to threats to privacy; large-scale perpetuation of biases; plagiarism; displacement of human workers and creativity; deception and malicious use.

Acting ethically requires realising the value of technology while ensuring the risks are reasonable. Institutional or structural changes may be necessary, for example, labelling and creation of watermarks to prevent plagiarism.

To be human is to make your own ethical decisions

One ethical question that confronts regulators and industry is: should a driverless car swerve and sacrifice an innocent bystander (pedestrian or cyclist) to save a greater number of lives in the vehicle? I asked ChatGPT whether it is right to kill one to save five. Here is the answer:

“Whether or not it is right to kill one person to save five people is a complex ethical question. Some people may argue that taking one life to save five is the more logical option, and is therefore justified. However, it is also important to consider the ethical principle of respect for autonomy.

“This principle states that all individuals have an inherent right to make autonomous and informed decisions about their lives, including their own death. Taking this into consideration, it may not be morally right to force someone to accept the sacrifice of their life in order to save the lives of five other people. Ultimately, the answer to this question is subjective and depends on one’s own ethical and moral beliefs.”

But as driverless cars are rolled out, we won’t have the luxury of remaining on the fence, or relying on drivers’ reflexes to get us out of a decision. We will have to make moral choices and give reasons, and apply them consistently to programming the vehicles. This requires deep – and human – ethical thinking and expertise.

The info is here.

Definition drives design: Disability models and mechanisms of bias in AI technologies

Newman-Griffis, D., et al. (2023).

First Monday, 28(1).

https://doi.org/10.5210/fm.v28i1.12903

Abstract

The increasing deployment of artificial intelligence (AI) tools to inform decision-making across diverse areas including healthcare, employment, social benefits, and government policy, presents a serious risk for disabled people, who have been shown to face bias in AI implementations. While there has been significant work on analysing and mitigating algorithmic bias, the broader mechanisms of how bias emerges in AI applications are not well understood, hampering efforts to address bias where it begins. In this article, we illustrate how bias in AI-assisted decision-making can arise from a range of specific design decisions, each of which may seem self-contained and non-biasing when considered separately. These design decisions include basic problem formulation, the data chosen for analysis, the use the AI technology is put to, and operational design elements in addition to the core algorithmic design. We draw on three historical models of disability common to different decision-making settings to demonstrate how differences in the definition of disability can lead to highly distinct decisions on each of these aspects of design, leading in turn to AI technologies with a variety of biases and downstream effects. We further show that the potential harms arising from inappropriate definitions of disability in fundamental design stages are further amplified by a lack of transparency and disabled participation throughout the AI design process. Our analysis provides a framework for critically examining AI technologies in decision-making contexts and guiding the development of a design praxis for disability-related AI analytics. We put forth this article to provide key questions to facilitate disability-led design and participatory development to produce more fair and equitable AI technologies in disability-related contexts.

Conclusion

The proliferation of artificial intelligence (AI) technologies as behind the scenes tools to support decision-making processes presents significant risks of harm for disabled people. The unspoken assumptions and unquestioned preconceptions that inform AI technology development can serve as mechanisms of bias, building the base problem formulation that guides a technology on reductive and harmful conceptualisations of disability. As we have shown, even when developing AI technologies to address the same overall goal, different definitions of disability can yield highly distinct analytic technologies that reflect contrasting, frequently incompatible decisions in the information to analyse, what analytic process to use, and what the end product of analysis will be. Here we have presented an initial framework to support critical examination of specific design elements in the formulation of AI technologies for data analytics, as a tool to examine the definitions of disability used in their design and the resulting impacts on the technology. We drew on three important historical models of disability that form common foundations for policy, practice, and personal experience today—the medical, social, and relational models—and two use cases in healthcare and government benefits to illustrate how different ways of conceiving of disability can yield technologies that contrast and conflict with one another, creating distinct risks for harm.

The information is here.

AI in the hands of imperfect users

Kostick-Quenet, K.M., Gerke, S. 

npj Digit. Med. 5, 197 (2022). 

https://doi.org/10.1038/s41746-022-00737-z

Abstract

As the use of artificial intelligence and machine learning (AI/ML) continues to expand in healthcare, much attention has been given to mitigating bias in algorithms to ensure they are employed fairly and transparently. Less attention has fallen to addressing potential bias among AI/ML’s human users or factors that influence user reliance. We argue for a systematic approach to identifying the existence and impacts of user biases while using AI/ML tools and call for the development of embedded interface design features, drawing on insights from decision science and behavioral economics, to nudge users towards more critical and reflective decision making using AI/ML.

(cut)

Impacts of uncertainty and urgency on decision quality

Trust plays a particularly critical role when decisions are made in contexts of uncertainty. Uncertainty, of course, is a central feature of most clinical decision making, particularly for conditions (e.g., COVID-1930) or treatments (e.g., deep brain stimulation or gene therapies) that lack a long history of observed outcomes. As Wang and Busemeyer (2021) describe, “uncertain” choice situations can be distinguished from “risky” ones in that risky decisions have a range of outcomes with known odds or probabilities. If you flip a coin, we know we have a 50% chance to land on heads. However, to bet on heads comes with a high level of risk, specifically, a 50% chance of losing. Uncertain decision-making scenarios, on the other hand, have no well-known or agreed-upon outcome probabilities. This also makes uncertain decision making contexts risky, but those risks are not sufficiently known to the extent that permits rational decision making. In information-scarce contexts, critical decisions are by necessity made using imperfect reasoning or the use of “gap-filling heuristics” that can lead to several predictable cognitive biases. Individuals might defer to an authority figure (messenger bias, authority bias); they may look to see what others are doing (“bandwagon” and social norm effects); or may make affective forecasting errors, projecting current emotional states onto one’s future self. The perceived or actual urgency of clinical decisions can add further biases, like ambiguity aversion (preference for known versus unknown risks38) or deferral to the status quo or default, and loss aversion (weighing losses more heavily than gains of the same magnitude). These biases are intended to mitigate risks of the unknown when fast decisions must be made, but they do not always get us closer to arriving at the “best” course of action if all possible information were available.

(cut)

Conclusion

We echo others’ calls that before AI tools are “released into the wild,” we must better understand their outcomes and impacts in the hands of imperfect human actors by testing at least some of them according to a risk-based approach in clinical trials that reflect their intended use settings. We advance this proposal by drawing attention to the need to empirically identify and test how specific user biases and decision contexts shape how AI tools are used in practice and influence patient outcomes. We propose that VSD can be used to strategize human-machine interfaces in ways that encourage critical reflection, mitigate bias, and reduce overreliance on AI systems in clinical decision making. We believe this approach can help to reduce some of the burdens on physicians to figure out on their own (with only basic training or knowledge about AI) the optimal role of AI tools in decision making by embedding a degree of bias mitigation directly into AI systems and interfaces.

The info is here.

The AI Ethicist's Dirty Hands Problem

H. S. Sætra, M. Coeckelbergh, & J. Danaher

Communications of the ACM, January 2023, 

Vol. 66 No. 1, Pages 39-41

Assume an AI ethicist uncovers objectionable effects related to the increased usage of AI. What should they do about it? One option is to seek alliances with Big Tech in order to "borrow" their power to change things for the better. Another option is to seek opportunities for change that actively avoids reliance on Big Tech.

The choice between these two strategies gives rise to an ethical dilemma. For example, if the ethicist's research emphasized the grave and unfortunate consequences of Twitter and Facebook, should they promote this research by building communities on said networks? Should they take funding from Big Tech to promote the reform of Big Tech? Should they seek opportunities at Google or OpenAI if they are deeply concerned about the negative implications of large-scale language models?

The AI ethicist’s dilemma emerges when an ethicist must consider how their success in communicating an

identified challenge is associated with a high risk of decreasing the chances of successfully addressing the challenge.  This dilemma occurs in situations in which the means to achieve one’s goals are seemingly best achieved by supporting that which one wishes to correct and/or practicing the opposite of that which one preaches.

(cut)

The Need for More than AI Ethics

Our analysis of the ethicist’s dilemma shows why close ties with Big Tech can be detrimental for the ethicist seeking remedies for AI related problems.   It is important for ethicists, and computer scientists in general, to be aware of their links to the sources of ethical challenges related to AI.  One useful exercise would be to carefully examine what could happen if they attempted to challenge the actors with whom they are aligned. Such actions could include attempts to report unfortunate implications of the company’s activities internally, but also publicly, as Gebru did. Would such actions be met with active resistance, with inaction, or even straightforward sanctions? Such an exercise will reveal whether or not the ethicist feels free to openly and honestly express concerns about the technology with which they work. Such an exercise could be important, but as we have argued, these individuals are not necessarily positioned to achieve fundamental change in this system.

In response, we suggest the role of government is key to balancing the power the tech companies have

through employment, funding, and their control of modern digital infrastructure. Some will rightly argue that political power is also dangerous.   But so are the dangers of technology and unbridled innovation, and private corporations are central sources of these dangers. We therefore argue that private power must be effectively bridled by the power of government.  This is not a new argument, and is in fact widely accepted.

The info is here.

San Francisco will allow police to deploy robots that kill

Janie Har

Associated Press

Originally posted 29 Nov 22

Supervisors in San Francisco voted Tuesday to give city police the ability to use potentially lethal, remote-controlled robots in emergency situations -- following an emotionally charged debate that reflected divisions on the politically liberal board over support for law enforcement.

The vote was 8-3, with the majority agreeing to grant police the option despite strong objections from civil liberties and other police oversight groups. Opponents said the authority would lead to the further militarization of a police force already too aggressive with poor and minority communities.

Supervisor Connie Chan, a member of the committee that forwarded the proposal to the full board, said she understood concerns over use of force but that “according to state law, we are required to approve the use of these equipments. So here we are, and it’s definitely not a easy discussion.”

The San Francisco Police Department said it does not have pre-armed robots and has no plans to arm robots with guns. But the department could deploy robots equipped with explosive charges “to contact, incapacitate, or disorient violent, armed, or dangerous suspect” when lives are at stake, SFPD spokesperson Allison Maxie said in a statement.

“Robots equipped in this manner would only be used in extreme circumstances to save or prevent further loss of innocent lives,” she said.

Supervisors amended the proposal Tuesday to specify that officers could use robots only after using alternative force or de-escalation tactics, or concluding they would not be able to subdue the suspect through those alternative means. Only a limited number of high-ranking officers could authorize use of robots as a deadly force option.

The info is here.

Artificial intelligence and consent: a feminist anti-colonial critique

Varon, J., & Peña, P. (2021). 

Internet Policy Review, 10(4).

https://doi.org/10.14763/2021.4.1602

Abstract

Feminist theories have extensively debated consent in sexual and political contexts. But what does it mean to consent when we are talking about our data bodies feeding artificial intelligence (AI) systems? This article builds a feminist and anti-colonial critique about how an individualistic notion of consent is being used to legitimate practices of the so-called emerging Digital Welfare States, focused on digitalisation of anti-poverty programmes. The goal is to expose how the functional role of digital consent has been enabling data extractivist practices for control and exclusion, another manifestation of colonialism embedded in cutting-edge digital technology.

Here is an excerpt:

Another important criticism of this traditional idea of consent in sexual relationships is the forced binarism of yes/no. According to Gira Grant (2016), consent is not only given but also is built from multiple factors such as the location, the moment, the emotional state, trust, and desire. In fact, for this author, the example of sex workers could demonstrate how desire and consent are different, although sometimes confused as the same. For her there are many things that sex workers do without necessarily wanting to. However, they give consent for legitimate reasons.

It is also important how we express consent. For feminists such as Fraisse (2012), there is no consent without the body. In other words, consent has a relational and communication-based (verbal and nonverbal) dimension where power relationships matter (Tinat, 2012; Fraisse, 2012). This is very relevant when we discuss “tacit consent” in sexual relationships. In another dimension of how we express consent, Fraisse (2012) distinguishes between choice (the consent that is accepted and adhered to) and coercion (the "consent" that is allowed and endured).

According to Fraisse (2012), the critical view of consent that is currently claimed by feminist theories is not consent as a symptom of contemporary individualism; it has a collective approach through the idea of “the ethics of consent”, which provides attention to the "conditions" of the practice; the practice adapted to a contextual situation, therefore rejecting universal norms that ignore the diversified conditions of domination (Fraisse, 2012).

In the same sense, Lucia Melgar (2012) asserts that, in the case of sexual consent, it is not just an individual right, but a collective right of women to say "my body is mine" and from there it claims freedom to all bodies. As Sarah Ahmed (2017, n.p.) states “for feminism: no is a political labor”. In other words, “if your position is precarious you might not be able to afford no. [...] This is why the less precarious might have a political obligation to say no on behalf of or alongside those who are more precarious”. Referring to Éric Fassin, Fraisse (2012) understands that in this feminist view, consent will not be “liberal” anymore (as a refrain of the free individual), but “radical”, because, as Fassin would call, seeing in a collective act, it could function as some sort of consensual exchange of power.

The info is here.

A new control problem? Humanoid robots, artificial intelligence, and the value of control

Nyholm, S. 

AI Ethics (2022).

https://doi.org/10.1007/s43681-022-00231-y

Abstract

The control problem related to robots and AI usually discussed is that we might lose control over advanced technologies. When authors like Nick Bostrom and Stuart Russell discuss this control problem, they write in a way that suggests that having as much control as possible is good while losing control is bad. In life in general, however, not all forms of control are unambiguously positive and unproblematic. Some forms—e.g. control over other persons—are ethically problematic. Other forms of control are positive, and perhaps even intrinsically good. For example, one form of control that many philosophers have argued is intrinsically good and a virtue is self-control. In this paper, I relate these questions about control and its value to different forms of robots and AI more generally. I argue that the more robots are made to resemble human beings, the more problematic it becomes—at least symbolically speaking—to want to exercise full control over these robots. After all, it is unethical for one human being to want to fully control another human being. Accordingly, it might be seen as problematic—viz. as representing something intrinsically bad—to want to create humanoid robots that we exercise complete control over. In contrast, if there are forms of AI such that control over them can be seen as a form of self-control, then this might be seen as a virtuous form of control. The “new control problem”, as I call it, is the question of under what circumstances retaining and exercising complete control over robots and AI is unambiguously ethically good.

From the Concluding Discussion section

Self-control is often valued as good in itself or as an aspect of things that are good in themselves, such as virtue, personal autonomy, and human dignity. In contrast, control over other persons is often seen as wrong and bad in itself. This means, I have argued, that if control over AI can sometimes be seen or conceptualized as a form of self-control, then control over AI can sometimes be not only instrumentally good, but in certain respects also good as an end in itself. It can be a form of extended self-control, and therefore a form of virtue, personal autonomy, or even human dignity.

In contrast, if there will ever be any AI systems that could properly be regarded as moral persons, then it would be ethically problematic to wish to be in full control over them, since it is ethically problematic to want to be in complete control over a moral person. But even before that, it might still be morally problematic to want to be in complete control over certain AI systems; it might be problematic if they are designed to look and behave like human beings. There can be, I have suggested, something symbolically problematic about wanting to be in complete control over an entity that symbolizes or represents something—viz. a human being—that it would be morally wrong and in itself bad to try to completely control.

For these reasons, I suggest that it will usually be a better idea to try to develop AI systems that can sensibly be interpreted as extensions of our own agency while avoiding developing robots that can be, imitate, or represent moral persons. One might ask, though, whether the two possibilities can ever come together, so to speak.

Think, for example, of the robotic copy that the Japanese robotics researcher Hiroshi Ishiguro has created of himself. It is an interesting question whether the agency of this robot could be seen as an extension of Ishiguro’s agency. The robot certainly represents or symbolizes Ishiguro. So, if he has control over this robot, then perhaps this can be seen as a form of extended agency and extended self-control. While it might seem symbolically problematic if Ishiguro wants to have complete control over the robot Erica that he has created, which looks like a human woman, it might not be problematic in the same way if he wants to have complete control over the robotic replica that he has created of himself. At least it would be different in terms of what it can be taken to symbolize or represent.

The info is here.

In the corner of an Australian lab, a brain in a dish is playing a video game - and it’s getting better

Liam Mannix

Sydney Morning Herald

Originally posted 13 NOV 22

Here is an excerpt:

Artificial intelligence controls an ever-increasing slice of our lives. Smart voice assistants hang on our every word. Our phones leverage machine learning to recognise our face. Our social media lives are controlled by algorithms that surface content to keep us hooked.

These advances are powered by a new generation of AIs built to resemble human brains. But none of these AIs are really intelligent, not in the human sense of the word. They can see the superficial pattern without understanding the underlying concept. Siri can read you the weather but she does not really understand that it’s raining. AIs are good at learning by rote, but struggle to extrapolate: even teenage humans need only a few sessions behind the wheel before the can drive, while Google’s self-driving car still isn’t ready after 32 billion kilometres of practice.

A true ‘general artificial intelligence’ remains out of reach - and, some scientists think, impossible.

Is this evidence human brains can do something special computers never will be able to? If so, the DishBrain opens a new path forward. “The only proof we have of a general intelligence system is done with biological neurons,” says Kagan. “Why would we try to mimic what we could harness?”

He imagines a future part-silicon-part-neuron supercomputer, able to combine the raw processing power of silicon with the built-in learning ability of the human brain.

Others are more sceptical. Human intelligence isn’t special, they argue. Thoughts are just electro-chemical reactions spreading across the brain. Ultimately, everything is physics - we just need to work out the maths.

“If I’m building a jet plane, I don’t need to mimic a bird. It’s really about getting to the mathematical foundations of what’s going on,” says Professor Simon Lucey, director of the Australian Institute for Machine Learning.

Why start the DishBrains on Pong? I ask. Because it’s a game with simple rules that make it ideal for training AI. And, grins Kagan, it was one of the first video game ever coded. A nod to the team’s geek passions - which run through the entire project.

“There’s a whole bunch of sci-fi history behind it. The Matrix is an inspiration,” says Chong. “Not that we’re trying to create a Matrix,” he adds quickly. “What are we but just a goooey soup of neurons in our heads, right?”

Maybe. But the Matrix wasn’t meant as inspiration: it’s a cautionary tale. The humans wired into it existed in a simulated reality while machines stole their bioelectricity. They were slaves.

Is it ethical to build a thinking computer and then restrict its reality to a task to be completed? Even if it is a fun task like Pong?

“The real life correlate of that is people have already created slaves that adore them: they are called dogs,” says Oxford University’s Julian Savulescu.

Thousands of years of selective breeding has turned a wild wolf into an animal that enjoys rounding up sheep, that loves its human master unconditionally.

The info is here.

Public attitudes value interpretability but prioritize accuracy in Artificial Intelligence

Nussberger, A. M., Luo, L., Celis, L. E., 

& Crockett, M. J. (2022). 

Nature communications, 13(1), 5821.

Abstract

As Artificial Intelligence (AI) proliferates across important social institutions, many of the most powerful AI systems available are difficult to interpret for end-users and engineers alike. Here, we sought to characterize public attitudes towards AI interpretability. Across seven studies (N = 2475), we demonstrate robust and positive attitudes towards interpretable AI among non-experts that generalize across a variety of real-world applications and follow predictable patterns. Participants value interpretability positively across different levels of AI autonomy and accuracy, and rate interpretability as more important for AI decisions involving high stakes and scarce resources. Crucially, when AI interpretability trades off against AI accuracy, participants prioritize accuracy over interpretability under the same conditions driving positive attitudes towards interpretability in the first place: amidst high stakes and scarce resources. These attitudes could drive a proliferation of AI systems making high-impact ethical decisions that are difficult to explain and understand.

Discussion

In recent years, academics, policymakers, and developers have debated whether interpretability is a fundamental prerequisite for trust in AI systems. However, it remains unknown whether non-experts–who may ultimately comprise a significant portion of end-users for AI applications–actually care about AI interpretability, and if so, under what conditions. Here, we characterise public attitudes towards interpretability in AI across seven studies. Our data demonstrates that people consider interpretability in AI to be important. Even though these positive attitudes generalise across a host of AI applications and show systematic patterns of variation, they also seem to be capricious. While people valued interpretability as similarly important for AI systems that directly implemented decisions and AI systems recommending a course of action to a human (Study 1A), they valued interpretability more for applications involving higher (relative to lower) stakes and for applications determining access to scarce (relative to abundant) resources (Studies 1A-C, Study 2). And while participants valued AI interpretability across all levels of AI accuracy when considering the two attributes independently (Study 3A), they sacrificed interpretability for accuracy when these two attributes traded off against one another (Studies 3B–C). Furthermore, participants favoured accuracy over interpretability under the same conditions that drove importance ratings of interpretability in the first place: when stakes are high and resources are scarce.

Our findings highlight that high-stakes applications, such as medical diagnosis, will generally be met with enhanced requirements towards AI interpretability. Notably, this sensitivity to stakes parallels magnitude-sensitivity as a foundational process in the cognitive appraisal of outcomes. The impact of stakes on attitudes towards interpretability were apparent not only in our experiments that manipulated stakes within a given AI-application, but also in absolute and relative levels of participants’ valuation of interpretability across applications–take, for instance, ‘hurricane first aid’ and ‘vaccine allocation’ outperforming ‘hiring decisions’, ‘insurance pricing’, and ‘standby seat prioritizing’. Conceivably, this ordering would also emerge if we ranked the applications according to the scope of auditing- and control-measures imposed on human executives, reflecting interpretability’s essential capacity of verifying appropriate and fair decision processes.

The research is here.

AI Isn’t Ready to Make Unsupervised Decisions

Joe McKendrick and Andy Thurai

Harvard Business Review

Originally published September 15, 2022

Artificial intelligence is designed to assist with decision-making when the data, parameters, and variables involved are beyond human comprehension. For the most part, AI systems make the right decisions given the constraints. However, AI notoriously fails in capturing or responding to intangible human factors that go into real-life decision-making — the ethical, moral, and other human considerations that guide the course of business, life, and society at large.

Consider the “trolley problem” — a hypothetical social scenario, formulated long before AI came into being, in which a decision has to be made whether to alter the route of an out-of-control streetcar heading towards a disaster zone. The decision that needs to be made — in a split second — is whether to switch from the original track where the streetcar may kill several people tied to the track, to an alternative track where, presumably, a single person would die.

While there are many other analogies that can be made about difficult decisions, the trolley problem is regarded to be the pinnacle exhibition of ethical and moral decision making. Can this be applied to AI systems to measure whether AI is ready for the real world, in which machines can think independently, and make the same ethical and moral decisions, that are justifiable, that humans would make?

Trolley problems in AI come in all shapes and sizes, and decisions don’t necessarily need to be so deadly — though the decisions AI renders could mean trouble for a business, individual, or even society at large. One of the co-authors of this article recently encountered his own AI “trolley moment,” during a stay in an Airbnb-rented house in upstate New Hampshire. Despite amazing preview pictures and positive reviews, the place was poorly maintained and a dump with condemned adjacent houses. The author was going to give the place a low one-star rating and a negative review, to warn others considering a stay.

However, on the second morning of the stay, the host of the house, a sweet and caring elderly woman, knocked on the door, inquiring if the author and his family were comfortable and if they had everything they needed. During the conversation, the host offered to pick up some fresh fruits from a nearby farmers market. She also said she doesn’t have a car, she would walk a mile to a friend’s place, who would then drive her to the market. She also described her hardships over the past two years, as rentals slumped due to Covid and that she is caring for someone sick full time.

Upon learning this, the author elected not to post the negative review. While the initial decision — to write a negative review — was based on facts, the decision not to post the review was purely a subjective human decision. In this case, the trolley problem was concern for the welfare of the elderly homeowner superseding consideration for the comfort of other potential guests.

How would an AI program have handled this situation? Likely not as sympathetically for the homeowner. It would have delivered a fact-based decision without empathy for the human lives involved.

The article is here.

The uselessness of AI ethics

Munn, L. The uselessness of AI ethics.

AI Ethics (2022).

https://doi.org/10.1007/s43681-022-00209-w

Abstract

As the awareness of AI’s power and danger has risen, the dominant response has been a turn to ethical principles. A flood of AI guidelines and codes of ethics have been released in both the public and private sector in the last several years. However, these are meaningless principles which are contested or incoherent, making them difficult to apply; they are isolated principles situated in an industry and education system which largely ignores ethics; and they are toothless principles which lack consequences and adhere to corporate agendas. For these reasons, I argue that AI ethical principles are useless, failing to mitigate the racial, social, and environmental damages of AI technologies in any meaningful sense. The result is a gap between high-minded principles and technological practice. Even when this gap is acknowledged and principles seek to be “operationalized,” the translation from complex social concepts to technical rulesets is non-trivial. In a zero-sum world, the dominant turn to AI principles is not just fruitless but a dangerous distraction, diverting immense financial and human resources away from potentially more effective activity. I conclude by highlighting alternative approaches to AI justice that go beyond ethical principles: thinking more broadly about systems of oppression and more narrowly about accuracy and auditing.

(cut)

Meaningless principles

The deluge of AI codes of ethics, frameworks, and guidelines in recent years has produced a corresponding raft of principles. Indeed, there are now regular meta-surveys which attempt to collate and summarize these principles. However, these principles are highly abstract and ambiguous, becoming incoherent. Mittelstadt suggests that work on AI ethics has largely produced “vague, high-level principles, and value statements which promise to be action-guiding, but in practice provide few specific recommendations and fail to address fundamental normative and political tensions embedded in key concepts.” The point here is not to debate the merits of any one value over another, but to highlight the fundamental lack of consensus around key terms. Commendable values like “fairness” and “privacy” break down when subjected to scrutiny, leading to disparate visions and deeply incompatible goals.

What are some common AI principles? Despite the mushrooming of ethical statements, Floridi and Cowls suggest many values recur frequently and can be condensed into five core principles: beneficence, non-maleficence, autonomy, justice, and explicability. These ideals sound wonderful. After all, who could be against beneficence? However, problems immediately arise when we start to define what beneficence means. In the Montreal principles for instance, “well-being” is the term used, suggesting that AI development should promote the “well-being of all sentient creatures.” While laudable, clearly there are tensions to consider here. We might think, for instance, of how information technologies support certain conceptions of human flourishing by enabling communication and business transactions—while simultaneously contributing to carbon emissions, environmental degradation, and the climate crisis. In other words, AI promotes the well-being of some creatures (humans) while actively undermining the well-being of others.

The same issue occurs with the Statement on Artificial Intelligence, Robotics, and Autonomous Systems. In this Statement, beneficence is gestured to through the concept of “sustainability,” asserting that AI must promote the basic preconditions for life on the planet. Few would argue directly against such a commendable aim. However, there are clearly wildly divergent views on how this goal should be achieved. Proponents of neoliberal interventions (free trade, globalization, deregulation) would argue that these interventions contribute to economic prosperity and in that sense sustain life on the planet. In fact, even the oil and gas industry champions the use of AI under the auspices of promoting sustainability. Sustainability, then, is a highly ambiguous or even intellectually empty term that is wrapped around disparate activities and ideologies. In a sense, sustainability can mean whatever you need it to mean. Indeed, even one of the members of the European group denounced the guidelines as “lukewarm” and “deliberately vague,” stating they “glossed over difficult problems” like explainability with rhetoric.

Ethical considerations for precision psychiatry: A roadmap for research and clinical practice

Fusar-Poli, P., Manchia, M., et al. (2022, October). 

European Neuropsychopharmacology, 63, 17–34.

https://doi.org/10.1016/j.euroneuro.2022.08.001

Abstract

Precision psychiatry is an emerging field with transformative opportunities for mental health. However, the use of clinical prediction models carries unprecedented ethical challenges, which must be addressed before accessing the potential benefits of precision psychiatry. This critical review covers multidisciplinary areas, including psychiatry, ethics, statistics and machine-learning, healthcare and academia, as well as input from people with lived experience of mental disorders, their family, and carers. We aimed to identify core ethical considerations for precision psychiatry and mitigate concerns by designing a roadmap for research and clinical practice. We identified priorities: learning from somatic medicine; identifying precision psychiatry use cases; enhancing transparency and generalizability; fostering implementation; promoting mental health literacy; communicating risk estimates; data protection and privacy; and fostering the equitable distribution of mental health care. We hope this blueprint will advance research and practice and enable people with mental health problems to benefit from precision psychiatry.

From the Results section

3.1. Ethics of precision psychiatry: Key concepts

Broadly speaking, ethical issues concern the development of ‘practical ought claims’ (Sheehan and Dunn, 2013) (i.e. normative claims that are practical in nature), which arise when we face ethical uncertainty in precision psychiatry. These practical claims come schematically like this: how should somebody or a group of people act in relation to a particular issue when they face certain circumstances? For example, how should researchers inform patients about their individualised risk estimates after running a novel clinical prediction model? To address these questions, four overarching ethical principles have been suggested (by Beauchamp and Childress) (Beauchamp and Childress, 2019), which include autonomy, beneficence, non-maleficence and justice. These can be applied to precision psychiatry, complemented by an extra principle of “explainability/interpretability” (Panel 1) which has been specifically introduced for artificial intelligence (Floridi et al., 2018) (for a more detailed discussion of ethical platforms for big data analytics see eSupplementary 1).

Although these four principles have become the cornerstones of biomedical ethics in healthcare practice, they have been criticised as they are often conflicting with no clear hierarchy and are not very specific (i.e. these principles are somewhat implicit, representing general moral values), leading to “imprecise ethics” that may not fit the needs of precision psychiatry (Table 1). Rather we should ask ourselves “why” a certain act may be harmful or beneficial. For example, let's imagine having a risk assessment; what would that mean for the individual, their family planning, workplace, choosing their studies, or their period of life? Alternatively, let's imagine that the risk assessment is not performed; what would be the results in a few years’ time? To address these sorts of questions, this study will consider ethical values in a broader sense, for example, by taking into account some of the different principles present in the charter of fundamental rights of the European Union – starting from dignity, freedom, equality, solidarity, citizens’ rights and justice (Table 1) (European Union, 2012; Hallinan, 2021). In particular, human dignity and human flourishing are the most crucial elements from an ethical point of view that are tightly linked to autonomy and self-determination (which is modulated by several factors such as physical health, psychological state, sociocultural environment, as well as values and beliefs). The loss of insight associated with some psychiatric disorders may incapacitate the individual to make autonomous decisions. For example, autonomy emerged as the driving decision component for undergoing risk prediction testing among young populations (Mantell et al., 2021a), regardless of whether a person would decide for or against risk profiling. Finally, it is important to highlight that unique ethical considerations may be associated with the historically complex socio-political perceptions and attitudes towards severe mental disorders and psychiatry (Ball et al., 2020a; Manchia et al., 2020a).

The info is here.

Technology and moral change: the transformation of truth and trust

Danaher, J., Sætra, H.S. 

Ethics Inf Technol 24, 35 (2022).

https://doi.org/10.1007/s10676-022-09661-y

Abstract

Technologies can have profound effects on social moral systems. Is there any way to systematically investigate and anticipate these potential effects? This paper aims to contribute to this emerging field on inquiry through a case study method. It focuses on two core human values—truth and trust—describes their structural properties and conceptualisations, and then considers various mechanisms through which technology is changing and can change our perspective on those values. In brief, the paper argues that technology is transforming these values by changing the costs/benefits of accessing them; allowing us to substitute those values for other, closely-related ones; increasing their perceived scarcity/abundance; and disrupting traditional value-gatekeepers. This has implications for how we study other, technologically-mediated, value changes.

(cut)

Conclusion: lessons learned

Having examined our two case studies, it remains to consider whether or not there are similarities in how technology affects trust and truth, and if there are general lessons to be learned here about how technology may impact values in the future.

The two values we have considered are structurally similar and interrelated. They are both intrinsically and instrumentally valuable. They are both epistemic and practical in nature: we value truth and trust (at least in part) because they give us access to knowledge and help us to resolve the decision problems we face on a daily basis. We also see, in both case studies, similar mechanisms of value change at work. The most interesting, to our minds, are the following:

• Technology changes the costs associated with accessing certain values, making them less or more important as a result Digital disinformation technology increases the cost of finding out the truth, but reduces the cost of finding and reinforcing a shared identity community; reliable AI and robotics gives us an (often cheaper and more efficient) substitute for trust in humans, while still giving us access to useful cognitive, emotional and physical assistance.

• Technology makes it easier, or more attractive to trade off or substitute some values against others Digital disinformation technology allows us to obviate the need for finding out the truth and focus on other values instead; reliable machines allow us to substitute the value of reliability for the value of trust. This is a function of the plural nature of values, their scarcity, and the changing cost structure of values caused by technology.

• Technology can make some values seem more scarce (rare, difficult to obtain), thereby increasing their perceived intrinsic value Digital disinformation makes truth more elusive, thereby increasing its perceived value which, in turn, encourages some moral communities to increase their fixation on it; robots and AI make trust in humans less instrumentally necessary, thereby increasing the expressive value of trust in others.

• Technology can disrupt power networks, thereby altering the social gatekeepers to value to the extent that we still care about truth, digital disinformation increases the power of the epistemic elites that can help us to access the truth; trust-free or trust-alternative technologies can disrupt the power of traditional trusted third parties (professionals, experts etc.) and redistribute power onto technology or a technological elite.

The info is here.

A Normative Approach to Artificial Moral Agency

Behdadi, D., Munthe, C.

Minds & Machines 30, 195–218 (2020).

https://doi.org/10.1007/s11023-020-09525-8

Abstract

This paper proposes a methodological redirection of the philosophical debate on artificial moral agency (AMA) in view of increasingly pressing practical needs due to technological development. This “normative approach” suggests abandoning theoretical discussions about what conditions may hold for moral agency and to what extent these may be met by artificial entities such as AI systems and robots. Instead, the debate should focus on how and to what extent such entities should be included in human practices normally assuming moral agency and responsibility of participants. The proposal is backed up by an analysis of the AMA debate, which is found to be overly caught in the opposition between so-called standard and functionalist conceptions of moral agency, conceptually confused and practically inert. Additionally, we outline some main themes of research in need of attention in light of the suggested normative approach to AMA.

Free will and Autonomy

Several AMA debaters have claimed that free will is necessary for being a moral agent (Himma 2009; Hellström 2012; Friedman and Kahn 1992). Others make a similar (and perhaps related) claim that autonomy is necessary (Lin et al. 2008; Schulzke 2013). In the AMA debate, some argue that artificial entities can never have free will (Bringsjord 1992; Shen 2011; Bringsjord 2007) while others, like James Moor (2006, 2009), are open to the possibility that future machines might acquire free will.Footnote15 Others (Powers 2006; Tonkens 2009) have proposed that the plausibility of a free will condition on moral agency may vary depending on what type of normative ethical theory is assumed, but they have not developed this idea further.

Despite appealing to the concept of free will, this portion of the AMA debate does not engage with key problems in the free will literature, such as the debate about compatibilism and incompatibilism (O’Connor 2016). Those in the AMA debate assume the existence of free will among humans, and ask whether artificial entities can satisfy a source control condition (McKenna et al. 2015). That is, the question is whether or not such entities can be the origins of their actions in a way that allows them to control what they do in the sense assumed of human moral agents.

An exception to this framing of the free will topic in the AMA debate occurs when Johnson writes that ‘… the non-deterministic character of human behavior makes it somewhat mysterious, but it is only because of this mysterious, non-deterministic aspect of moral agency that morality and accountability are coherent’ (Johnson 2006 p. 200). This is a line of reasoning that seems to assume an incompatibilist and libertarian sense of free will, assuming both that it is needed for moral agency and that humans do possess it. This, of course, makes the notion of human moral agents vulnerable to standard objections in the general free will debate (Shaw et al. 2019). Additionally, we note that Johnson’s idea about the presence of a ‘mysterious aspect’ of human moral agents might allow for AMA in the same way as Dreyfus and Hubert’s reference to the subconscious: artificial entities may be built to incorporate this aspect.

The question of sourcehood in the AMA debate connects to the independence argument: For instance, when it is claimed that machines are created for a purpose and therefore are nothing more than advanced tools (Powers 2006; Bryson 2010; Gladden 2016) or prosthetics (Johnson and Miller 2008), this is thought to imply that machines can never be the true or genuine source of their own actions. This argument questions whether the independence required for moral agency (by both functionalists and standardists) can be found in a machine. If a machine’s repertoire of behaviors and responses is the result of elaborate design then it is not independent, the argument goes. Floridi and Sanders question this proposal by referring to the complexity of ‘human programming’, such as genes and arranged environmental factors (e.g. education). 

The info is here.

Preventing an AI-related catastrophe

Benjamin Hilton

8000 Hours

Originally Published August 25th, 2022

Summary

We expect that there will be substantial progress in AI in the next few decades, potentially even to the point where machines come to outperform humans in many, if not all, tasks. This could have enormous benefits, helping to solve currently intractable global problems, but could also pose severe risks. These risks could arise accidentally (for example, if we don’t find technical solutions to concerns about the safety of AI systems), or deliberately (for example, if AI systems worsen geopolitical conflict). We think more work needs to be done to reduce these risks.

Some of these risks from advanced AI could be existential — meaning they could cause human extinction, or an equally permanent and severe disempowerment of humanity. There have not yet been any satisfying answers to concerns — discussed below — about how this rapidly approaching, transformative technology can be safely developed and integrated into our society. Finding answers to these concerns is very neglected, and may well be tractable. We estimate that there are around 300 people worldwide working directly on this.3 As a result, the possibility of AI-related catastrophe may be the world’s most pressing problem — and the best thing to work on for those who are well-placed to contribute.

Promising options for working on this problem include technical research on how to create safe AI systems, strategy research into the particular risks AI might pose, and policy research into ways in which companies and governments could mitigate these risks. If worthwhile policies are developed, we’ll need people to put them in place and implement them. There are also many opportunities to have a big impact in a variety of complementary roles, such as operations management, journalism, earning to give, and more — some of which we list below.

(cut)

When can we expect transformative AI?

It’s difficult to predict exactly when we will develop AI that we expect to be hugely transformative for society (for better or for worse) — for example, by automating all human work or drastically changing the structure of society. But here we’ll go through a few approaches.

One option is to survey experts. Data from the 2019 survey of 300 AI experts implies that there is 20% probability of human-level machine intelligence (which would plausibly be transformative in this sense) by 2036, 50% probability by 2060, and 85% by 2100. There are a lot of reasons to be suspicious of these estimates,8 but we take it as one data point.

Ajeya Cotra (a researcher at Open Philanthropy) attempted to forecast transformative AI by comparing modern deep learning to the human brain. Deep learning involves using a huge amount of compute to train a model, before that model is able to perform some task. There’s also a relationship between the amount of compute used to train a model and the amount used by the model when it’s run. And — if the scaling hypothesis is true — we should expect the performance of a model to predictably improve as the computational power used increases. So Cotra used a variety of approaches (including, for example, estimating how much compute the human brain uses on a variety of tasks) to estimate how much compute might be needed to train a model that, when run, could carry out the hardest tasks humans can do. She then estimated when using that much compute would be affordable.

Cotra’s 2022 update on her report’s conclusions estimates that there is a 35% probability of transformative AI by 2036, 50% by 2040, and 60% by 2050 — noting that these guesses are not stable.22

Tom Davidson (also a researcher at Open Philanthropy) wrote a report to complement Cotra’s work. He attempted to figure out when we might expect to see transformative AI based only on looking at various types of research that transformative AI might be like (e.g. developing technology that’s the ultimate goal of a STEM field, or proving difficult mathematical conjectures), and how long it’s taken for each of these kinds of research to be completed in the past, given some quantity of research funding and effort.

Davidson’s report estimates that, solely on this information, you’d think that there was an 8% chance of transformative AI by 2036, 13% by 2060, and 20% by 2100. However, Davidson doesn’t consider the actual ways in which AI has progressed since research started in the 1950s, and notes that it seems likely that the amount of effort we put into AI research will increase as AI becomes increasingly relevant to our economy. As a result, Davidson expects these numbers to be underestimates.

The info is here.

‘The entire protein universe’: AI predicts shape of nearly every known protein

Ewen Callaway

Nature (608)

Posted with correction 29 July 22

From today, determining the 3D shape of almost any protein known to science will be as simple as typing in a Google search.

Researchers have used AlphaFold — the revolutionary artificial-intelligence (AI) network — to predict the structures of more than 200 million proteins from some 1 million species, covering almost every known protein on the planet.

The data dump is freely available on a database set up by DeepMind, the London-based AI company, owned by Google, that developed AlphaFold, and the European Molecular Biology Laboratory’s European Bioinformatics Institute (EMBL–EBI), an intergovernmental organization near Cambridge, UK.

“Essentially you can think of it covering the entire protein universe,” DeepMind chief executive Demis Hassabis said at a press briefing. “We’re at the beginning of a new era of digital biology.”

The 3D shape, or structure, of a protein is what determines its function in cells. Most drugs are designed using structural information, and the creation of accurate maps of proteins’ amino-acid arrangement is often the first step to making discoveries about how proteins work.

DeepMind developed the AlphaFold network using an AI technique called deep learning, and the AlphaFold database was launched a year ago with more than 350,000 structure predictions covering nearly every protein made by humans, mice and 19 other widely studied organisms. The catalogue has since swelled to around 1 million entries.

“We’re bracing ourselves for the release of this huge trove,” says Christine Orengo, a computational biologist at University College London, who has used the AlphaFold database to identify new families of proteins. “Having all the data predicted for us is just fantastic.”

(cut)

But such entries tend to be skewed toward human, mouse and other mammalian proteins, Porta says. It’s likely that the AlphaFold dump will add significant knowledge, because it includes such a diverse range of organisms. “It’s going to be an awesome resource. And I’m probably going to download it as soon as it comes out,” says Porta.

The info is here.

Dual use of artifcial-intelligence-powered drug discovery

Urbina, F., Lentzos, F., Invernizzi, C. et al. 

Nat Mach Intell 4, 189–191 (2022). 

https://doi.org/10.1038/s42256-022-00465-9

The Swiss Federal Institute for NBC (nuclear, biological and chemical) Protection —Spiez Laboratory— convenes the ‘convergence’ conference series set up by the Swiss government to identify developments in chemistry, biology and enabling technologies that may have implications for the Chemical and Biological Weapons Conventions. Meeting every two years, the conferences bring together an international group of scientific and disarmament experts to explore the current state of the art in the chemical and biological fields and their trajectories, to think through potential security implications and to consider how these implications can most effectively be managed internationally.  The meeting convenes for three days of discussion on the possibilities of harm, should the intent be there, from cutting-edge chemical and biological technologies.  Our drug discovery company received an invitation to contribute a presentation on how AI technologies for drug discovery could potentially be misused.

Risk of misuse

The thought had never previously struck us. We were vaguely aware of security concerns around work with pathogens or toxic chemicals, but that did not relate to us; we primarily operate in a virtual setting.  Our work is rooted in building machine learning models for therapeutic and toxic targets to better assist in the design of new molecules for drug discovery. We have spent decades using computers and AI to improve human health—not to degrade it. We were naive in thinking about the potential misuse of our trade, as our aim had always been to avoid molecular features that could interfere with the many different classes of proteins essential to human life. Even our projects on Ebola and neurotoxins, which could have sparked thoughts about the potential negative implications of our machine learning models, had not set our alarm bells ringing.

(cut)

Broader effects on society

There is a need for discussions across traditional boundaries and multiple disciplines to allow for a fresh look at AI for de novo design and related technologies from different perspectives and with a wide variety of mindsets. Here, we give some recommendations that we believe will reduce potential dual-use concerns for AI in drug discovery. Scientific conferences, such as the Society of Toxicology and American Chemical Society, should actively foster a dialogue among experts from industry, academia and policy making on the implications of our computational tools.

A pdf is here.

Robots became racist after AI training, always chose Black faces as ‘criminals’

Pranshu Verma

The Washington Post

Originally posted 16 JUL 22

As part of a recent experiment, scientists asked specially programmed robots to scan blocks with people’s faces on them, then put the “criminal” in a box. The robots repeatedly chose a block with a Black man’s face.

Those virtual robots, which were programmed with a popular artificial intelligence algorithm, were sorting through billions of images and associated captions to respond to that question and others, and may represent the first empirical evidence that robots can be sexist and racist, according to researchers. Over and over, the robots responded to words like “homemaker” and “janitor” by choosing blocks with women and people of color.

The study, released last month and conducted by institutions including Johns Hopkins University and the Georgia Institute of Technology, shows the racist and sexist biases baked into artificial intelligence systems can translate into robots that use them to guide their operations.

Companies have been pouring billions of dollars into developing more robots to help replace humans for tasks such as stocking shelves, delivering goods or even caring for hospital patients. Heightened by the pandemic and a resulting labor shortage, experts describe the current atmosphere for robotics as something of a gold rush. But tech ethicists and researchers are warning that the quick adoption of the new technology could result in unforeseen consequences down the road as the technology becomes more advanced and ubiquitous.

“With coding, a lot of times you just build the new software on top of the old software,” said Zac Stewart Rogers, a supply chain management professor from Colorado State University. “So, when you get to the point where robots are doing more … and they’re built on top of flawed roots, you could certainly see us running into problems.”

Researchers in recent years have documented multiple cases of biased artificial intelligence algorithms. That includes crime prediction algorithms unfairly targeting Black and Latino people for crimes they did not commit, as well as facial recognition systems having a hard time accurately identifying people of color.

The info is here.

AI bias can arise from annotation instructions

K. Wiggers & D. Coldeway

TechCrunch

Originally posted 8 MAY 22

Here is an excerpt:

As it turns out, annotators’ predispositions might not be solely to blame for the presence of bias in training labels. In a preprint study out of Arizona State University and the Allen Institute for AI, researchers investigated whether a source of bias might lie in the instructions written by dataset creators to serve as guides for annotators. Such instructions typically include a short description of the task (e.g., “Label all birds in these photos”) along with several examples.

The researchers looked at 14 different “benchmark” datasets used to measure the performance of natural language processing systems, or AI systems that can classify, summarize, translate and otherwise analyze or manipulate text. In studying the task instructions provided to annotators that worked on the datasets, they found evidence that the instructions influenced the annotators to follow specific patterns, which then propagated to the datasets. For example, over half of the annotations in Quoref, a dataset designed to test the ability of AI systems to understand when two or more expressions refer to the same person (or thing), start with the phrase “What is the name,” a phrase present in a third of the instructions for the dataset.

The phenomenon, which the researchers call “instruction bias,” is particularly troubling because it suggests that systems trained on biased instruction/annotation data might not perform as well as initially thought. Indeed, the co-authors found that instruction bias overestimates the performance of systems and that these systems often fail to generalize beyond instruction patterns.

The silver lining is that large systems, like OpenAI’s GPT-3, were found to be generally less sensitive to instruction bias. But the research serves as a reminder that AI systems, like people, are susceptible to developing biases from sources that aren’t always obvious. The intractable challenge is discovering these sources and mitigating the downstream impact.

The info is here.