The AI Ethics Winter
AI Ethics is one-sided and increasingly irrelevant at the time it is needed most
I recently gave a talk at a flagship AI Ethics conference and thought: If the proverbial Martian visitor attended this conference, they’d have a fine-grained understanding of all the issues highlighted by this self-styled “AI hater”:
“the environmental harms, the reinforcement of bias and generation of racist output, the cognitive harms and AI supported suicides, the problems with consent and copyright, the way AI tech companies further the patterns of empire, how it’s a con that enables fraud and disinformation and harassment and surveillance, the exploitation of workers, as an excuse to fire workers and de-skill work, how they don’t actually reason and probability and association are inadequate to the goal of intelligence…”
And the list goes on.
But our Martian would be excused for not knowing that AI researchers just won two Nobel Prizes, or that Indigenous researchers are using AI to preserve their languages, or that scientists are using AI to search for life on other planets, improve prediction of natural disasters, find political consensus, and improve animal welfare. Nor would they know that some people become less anxious or depressed after talking to chatbots, or that AI is detecting breast cancer earlier and more accurately than radiologists.
The “AI Winter” described a period (roughly, 1990s) when the field’s credibility collapsed because its practitioners overpromised and underdelivered. The one-sidedness of AI Ethics is putting the field on a path toward a similar collapse, an AI Ethics winter.
How did this happen?
AI ethicists often point out, correctly, that people associated with AI companies have incentives to hype AI capabilities. What is less often appreciated is that AI ethicists have their own incentives.
Peter Königs argues that AI ethicists face three structural pressures: the field is organized around commenting on new technologies, there are strong norms against purely positive commentary, and academics must publish and get grants to survive.
The result is that raising ethical concerns about AI is the most viable career path. As Königs puts it, ethicists cannot simply say to an AI researcher: “Look, what you’re doing is terrific. I don’t see any problems.” No matter how good the project, the ethicist’s job requires them to problematize it. If they don’t, they’re out of a job.
Dan Williams has argued that these incentive structures shape both what researchers study and how they reason. When researchers are strongly incentivized to reach these kinds of critical conclusions, they engage in motivated reasoning, lowering standards for evidence and arguments that support such conclusions and ignoring information or possibilities in tension with them.
Consider Karen Hao’s widely acclaimed Empire of AI. In a chapter framing data center water use as a continuation of colonial extraction, Hao claimed that a Google data center in Chile “could use more than one thousand times the amount of water consumed by the entire population of Cerrillos, roughly eighty-eight thousand residents, over the course of a year.”
But Andy Masley showed these figures were off by three orders of magnitude. It resulted from confusing cubic meters with liters. This mistake passed through teams of fact-checkers and reviewers, and major press coverage before a correction was issued.
If you don’t believe Williams’s claim about motivated reasoning, just imagine the immediate and withering scrutiny that would follow from a high-profile journalist making a claim of comparable magnitude in the other direction: “AI is saving a thousand times more water than it uses!”
These incentives and motivated reasoning tendencies conspire to produce the one-sidedness that the opening Martian parable illustrates. Williams’s analysis of news media provides a helpful analogy here. Most individual stories published by mainstream outlets are accurate, but because outlets cover a non-random sample of negative events, audiences end up radically misinformed about the state of the world:
“People develop mental pictures of reality far more negative than the objective facts warrant. They overestimate poverty, crime rates, and many other social pathologies and dangers, and believe most trends are going in the wrong direction.”
Something similar is true of AI Ethics. Individual papers are often valuable. But the one-sided market for critical content means these papers add up to a body of work that is overwhelmingly and misleadingly negative.
So What?
One obvious rejoinder here is: So what if AI ethics is one-sided? Given the AI Hype Machine, we need a counterweight!
This seems plausible at first but falls apart under scrutiny. The counterweight argument implies that, without AI Ethics, society would be pro-AI. But public opinion on AI is not very positive, and so the idea that society needs AI ethicists to provide a corrective is dubious. And even if the counterweight theory held water, Königs points out that it is very strange indeed to deliberately design a scholarly community to be biased.
Painting AI in an overwhelmingly negative light also makes it very difficult to understand why nearly a billion people use ChatGPT each week, and why 50 million pay for it. One-sidedness also distorts the space for productive debate. Pointing out the negativity bias, or suggesting that AI can be beneficial reliably invites charges of being a “tech bro” or Silicon Valley apologist.
Incentives
One might also object (without resort to ad hominems) that the incentive argument cuts both ways. Both optimists and pessimists about AI are subject to incentives that can distort reasoning. The appropriate response is to prioritize adversarial collaboration (otherwise known as “philosophy”) where both sides subject their work to scrutiny from the other. What I am arguing against here is a disciplinary monoculture where only one of these orientations is professionally rewarded.
Irrelevance
As a card-carrying AI Ethicist, I think the single biggest problem with one-sidedness is that the discipline is critiquing its way into irrelevance. AI Ethics should be an applied discipline. This means engaging with real trade-offs in real institutions under real constraints. A medical ethicist who only says, “here’s the negative side effects” and never says, “here’s how to weigh risks, benefits, alternatives” would be useless to, and eventually ignored by, healthcare providers. That’s roughly where AI Ethics stands today.
Can it be fixed?
The oldest question
Critics will be quick to point to counterexamples1 so let me state unequivocally: Work in AI Ethics has, and continues to, identify important ways in which algorithmic systems perpetuate harms and contribute to various forms of oppression. I’ve written about this. I teach classes about it. I’ve advised and supported more than 20 graduate students working on projects in this vein. Moreover, we have every reason to adopt a sharply critical posture towards the technology companies currently racing to build AGI. It’d be foolish not to.
What I’m arguing is that this is not the ONLY thing AI Ethics should do.
AI Ethics is a sub-discipline of Ethics. Ethics is about living well. And living well requires much more than identifying and avoiding harms.
Consider another parallel, this time with an (oversimplified) example from epistemology. W.K. Clifford famously argued that “it is wrong always, everywhere, and for everyone, to believe anything upon insufficient evidence.” William James countered that we have two epistemic duties: avoiding error and gaining truth. Someone who only cared about never being wrong would believe almost nothing and miss most of what matters.
An ethics that only asks, “what could go wrong?” is as incomplete as an epistemology that only asks, “how might I be deceived?” And yet, AI Ethics has become so fixated on harms that it has lost sight of the equally important task of understanding how AI might contribute to human flourishing.
On its current trajectory, AI Ethics is making itself irrelevant at precisely the moment it should be indispensable. Luckily, the path back to relevance is simple. A renewed and honest commitment to the oldest question in philosophy: “How should we live?”
Obvious counterexamples include Gebru et al. (2018) on data sheets, Mitchell et al. (2019) on model cards, and Raji et al. (2020) on algorithmic audits. These are highly cited AI Ethics papers comprised of constructive tools, not purely negative critique. This is perfectly compatible with the observation that the vast majority of AI Ethics is still problematically one-sided.
I hate to tell someone else about their field, and I still think you're hitting on one component of the problem.
1. AI ethics has, certainly from my outside view, coalesced around a particular empirical/metaphysical claim (on the empirical side, AI is useless / a con; on the metaphysical side, that AI does something different from human brains, is just a stochastic parrot, and can't reason or produce useful text). This view was wrong when stochastic parrots was first published, and the obviousness of that wrongness is continually increasing. And yet so far I have seen relatively little motion to recognize that. It's not just being obviously wrong: it's the failure to build a track record of accurate predictions.
2. Because of the uniformly critical slant of the field, false claims (about AI and water, for example, which long predate Hao) that are negative about AI go unchallenged, resulting in bad thinking filling the field, bad thinking that outsiders can see. Intellectual diversity is useful, but the field understands itself, it seems, as having an agenda, and that's almost always harmful for making true statements.
3. Because the field is, at least as perceived from the outside, so overwhelmingly left-wing, it's been unable to usefully engage with anti-AI populist conservatives, and the AI hostility means that the only ability to work with anti-AI regulation people is hating EAs. So to the extent that the field talks to policymakers, it's limited to a fraction of people on the left.
4. As you point out, the constructive work is good and important. It's also incredibly rare, and I think that tends to make a field have less influence. This also seems to fit, say, sociologists vs economists: the latter so much more neutral and constructive work that policymakers can use, rather than just complaining.
5. My personal favorite AI Ethics Moment is the complaints about facial recognition being worse at identifying Black faces. It's obvious that if models had come out and were reliably *better* at identifying Black faces than white faces, that also would have been taken as evidence of anti-Black racism and something that needed to be changed because of how it empowered the carceral surveillance state. If you make the same critique regardless of the data, you're not providing information. And yet that's what I expect from the field.
I really want AI Ethics to be a vibrant field that raises good points and helps our culture adapt to AI better. I agree that you have raised some important points. But I have little optimism for the field.
You put your finger on something that's almost philosophically incoherent at its core. If your ethical framework is entirely organized around harm-avoidance, you don't actually have an ethics — you have a risk assessment protocol. A real ethics has to be able to say "this is worth doing, and here's why the benefits justify the costs." Without that capacity, you can't reason about trade-offs at all. You can only ever say no.
And "only ever say no" is a position that real decision-makers — in hospitals, in governments, in companies — will simply route around. They're not going to stop doing things because an ethicist told them it was bad. They're going to stop consulting ethicists. Which is exactly the irrelevance your warning is about.
The water data error in the Hao book is instructive here — not because one factual mistake damns a whole argument, but because of what it reveals about the direction of motivated reasoning. A factor-of-a-thousand mistake in the alarming direction slips through teams of fact-checkers. The same mistake in the reassuring direction would have been caught by the third pair of eyes.
What I think sits just underneath the piece, and isn't quite said explicitly, is the distinction between ethics as a practice and ethics as a posture. A lot of what currently passes for AI ethics is posture — it signals values, performs concern, positions the speaker on the right side of history. That's not nothing, but it's not the same as helping anyone navigate a hard decision. And the incentive structures Königs describes reward posture over practice almost perfectly.
The negative bell only has force if people take it seriously. And people stop taking it seriously when it rings constantly. Remember “The boy who cried: Wolf?