Tech critics are now beginning to draw specific attention to the ‘harms’ arising from AI. As Rene Shelby and colleagues (2022, p.2) have outlined, these can be defined as “adverse lived experiences resulting from a system’s deployment and operation in the world”. This builds on Hoffmann’s (2016) earlier distinction between different ways that technological systems can result in specific ‘distributional harms’ (i.e. injustices in the way that “rights, responsibilities, and resources” are distributed), as well as ‘dignitary harms’ (i.e. affronts to an individual’s self-respect). Following these leads, various researchers have begun to document how AI and algorithmic technologies result in ‘harmful’ forms of algorithmic discrimination, exclusion, exploitation and disadvantage.

Such talk of ‘harm’ marks a conscious effort to advance discussions around AI past vague notions of AI ethics, bias, fairness, accountability, transparency, and safety. All these latter terms have gained prominence in tech circles as a way of characterising and codifying the social outcomes arising from new AI technologies. However, such framings of ‘AI ethics’ and ‘AI safety’ are notably limited in their outcomes – often being used simply to justify the development of tech-driven ‘solutions’ and adjustments that purport to remedy the issues at hand. Indeed, tech developers are now well-adept at engineering improvements and interventions designed to make their technologies ethical, fair, transparent and/or accountable. In short, these reductive ways of talking about AI-related problems that tend to result in ‘fixes’.

In comparison, the notion of ‘harm’ points to the messy ways in which AI is caught up in social structures, disadvantage, oppression and societal stratification – all entanglements that cannot be easily categorised and ‘fixed’. Such elaboration is certainly welcome in terms of advancing discussions around the societal impact of AI. However, as tech critics have begun to address the problem of AI ‘harms’ in-depth, the limitations of the term have become increasingly apparent. In particular, there is considerable uncertainty (if not ambiguity) in terms of what (and for whom) different authors assume to constitute AI harm. Indeed, when read as a whole, the critical tech literature now contains significant inconsistences around what counts as AI harm. Often critics will lapse into intuitive ‘know it when you see it’ identifications of harm, meaning that their accounts are open to push-back by those wishing to underplay (or outright deny) the adverse impact of the technology’s deployment.

With this in mind, a new paper from Nathalie DiBerardino, Luke Stark and Clair Baleshta takes a deep dive into philosophical definitions of ‘harm’ – in particular, drawing on the work of legal philosopher Joel Feinberg and his distinction between “harms” and “wrongs”. Feinberg reasons that harm is a non-normative concept. In basic terms, then, a harm is any event that constitutes “the thwarting, setting back, or defeating” of an individual’s interests. Here Feinberg is using ‘interest’ to cover anything which an individual has a stake in, such as “physical integrity, intellectual acuity, a tolerable social and physical environment, and a certain amount of freedom from interest and coercion”. Key to any ‘harm’ occurring is the understanding that the individual is worse off than if the event had not happened. In contrast, Feinberg reasons that a wrong is a normative concept. In short, a wrong can be classed as any harm that is morally objectionable. In this sense, someone can be said to have been wronged when they have treated ‘unjustly’ by someone else. This notion of ‘unjust’ treatment relates to any action or omission that is “morally indefensible”.

These distinctions raise a number of important points. First, is the notion that while not all harms have to involve an agent, anything that constitutes a wrong will involve an agent in some way. Feinberg gives the example that someone can be harmed by an earthquake, but it is not possible for someone to be wronged by an earthquake. Any wrong requires additional agentic involvement – for example, being hurt in an earthquake due to shoddy construction or inadequate building codes.

This distinction raises the key point that all wrongs are contextual. In other words, when distinguishing between harms and wrongs we need to be mindful of the broader social contexts within which these actions and events are taking place. One important distinction is whether the harm/wrong takes place in a context where there are moral responsibilities or obligations between different people and authorities. For example, someone cutting their finger with a knife in their own kitchen would certainly constitute a harm. However, the same incident taking place in an unsafe workplace environment is likely to constitute a wrong – in short, we are ‘owed’ safety at work from our employers.

Any distinction between harms and wrongs also relates to the actions and events that take place around the main action. For example, being grievously injured by another person certainly constitutes a harm, but only constitutes a wrong if the attack is unprovoked. Obviously, If someone injures you in response to you having just attacked them, then you have not been wronged per se. Another complication are incidents of ‘pre-emptive’ harm – where an initial small harm might actually prevent a subsequent greater harm from occurring. As such examples illustrate, distinguishing between harms and wrongs is tricky. On one hand, not all harms are wrongs. On the other hand, people can sometimes be grievously wronged but only see themselves as being harmed – not being aware of the full moral dimension of the action/incident.

Such distinctions force us into much more detailed understandings of the adverse outcomes that might result from the deployment of AI and other data-driven systems in social settings. Clearly any system that is designed to make decisions or recommendations is going to set back some people’s interests over the interests of others. However, when these outcomes are socially embedded in already unjust or discriminatory contexts such as healthcare, policing or education systems then the chances increase of some of these being unjust setbacks to interests. DiBerardino makes the point that the most commonly reported AI ‘biases’ tend to be closely related to unjust or wrongful discrimination in contexts such as education, healthcare and criminal justice. Similarly, what are often described as privacy and security harms are actually instances of rights violations, and therefore morally indefensible wrongs.

In this sense, DiBerardino reminds us that judging what constitutes an algorithmic ‘wrong’ relies on engaging with diverse populations that might come into contact with the technology – especially historically disadvantaged groups that are not usually factored into tech design and development. For example, a technology such as facial recognition has long been compromised by being technically more accurate when identifying people who are white and/or male. The inconvenience of a white person being mis-recognised occasionally by FRT is not likely to be experienced as a ‘wrong’. In contrast, the outcomes that might occur from the misrecognition of a Black person by FRT – especially in social contexts such as a criminal justice system that are already more racially discriminatory or oppressive – is much more likely to constitute an unjust wrong. As such, we cannot discount the propensity of FRT to misrecognise faces as simply a minor technical aberration that is not fundamentally a social problem. The harms and wrongs arising from this technology depend on whose face is being misrecognised and in which contexts. It is no coincidence that news reports of people getting wrongfully arrested after being falsely identified from FRT are almost always Black men.

DiBerardino concludes that tech developers and regulators need to be clear on the distinction between algorithmic harms and wrongs if they are to develop robust technology and policy responses. Similarly, for tech critics, it is suggested that this more nuanced approach might help “ground a more powerful and rigorous normative critique or the baleful effects of many algorithmic systems today”. While it is good to have moved on from weak notions of AI bias, fairness and transparency, it is important to not downplay or underemphasise any blatantly unjust outcomes of AI use simply as ‘harms’. We need to be confident of calling out the unjust and disriminantory outcomes of AI and algorithmic technologies for what they are … wrong!

REFERENCES

DiBerardino, N., Baleshta, C., & Stark, L. (2024, June). Algorithmic Harms and Algorithmic Wrongs. In The 2024 ACM Conference on Fairness, Accountability, and Transparency (pp. 1725-1732).

Feinberg, J. (1986). Wrongful life and the counterfactual element in harming. Social Philosophy and Policy, 4(1):145-178.

Hoffmann, A. (2016). Beyond distributions and primary goods: assessing applications of Rawls in information science and technology literature since 1990. Journal of the Association for Information Science and Technology 68(7):1601–1618.

Shelby, R., Rismani, S., Henne, K., Moon, A., Rostamzadeh, N., Nicholas, P., Yilla, N., Gallegos, J., Smart, A., Garcia, E. and Virk, G. (2022). Sociotechnical harms: scoping a taxonomy for harm reduction. https://arxiv.org/abs/2210.05791v1