MIT Technology Review is raising an uncomfortable question about where the AI safety conversation has been focused — and who gets left out when the benchmark for responsible technology is set barely above catastrophe.
The publication's argument, distilled from its report, is that the technology policy debate has become captured by a kind of threshold logic: if a product does not actively help someone die, it is treated as a success. An AI chatbot that refrains from encouraging suicide, a social media platform that makes it marginally harder for children to obtain dangerous drugs — these are held up as evidence that the industry is self-correcting. The implicit message is that meeting a floor nobody should have fallen through in the first place constitutes progress.
To understand why this framing matters, it helps to trace how the AI safety conversation arrived here. For much of the last decade, the loudest voices in the room were concentrated on what researchers call existential or catastrophic risk — the possibility that increasingly powerful systems might, at some future point, behave in ways that threaten human welfare on a civilizational scale. That concern attracted serious philosophers, well-funded institutes, and eventually the attention of policymakers on both sides of the Atlantic. It is not a frivolous concern. But it has a particular effect on regulatory imagination: it trains attention on the dramatic and the hypothetical while the mundane and the present accumulates quietly in the background.
What fills that background is a long catalogue of documented, ordinary harms. Hiring algorithms that encode historical discrimination. Facial recognition systems with sharply higher error rates for darker-skinned faces, deployed in consequential contexts like law enforcement. Content recommendation systems that have been shown to push vulnerable users toward material that worsens their mental health, not by some science-fiction mechanism but by simple optimization for engagement. Credit-scoring tools that produce outcomes correlated with race or geography without anyone designing them to do so. These are not hypothetical injuries awaiting future technology. They are happening at scale, to real people, and they attract comparatively little of the policy energy that pools around the catastrophic scenarios.
The industry dynamic here is worth noting. For technology companies, the catastrophic-risk framing is, paradoxically, not entirely unwelcome. Engaging with long-range existential debates is expensive in public relations terms but cheap in operational ones. A company can fund a safety institute, publish a responsible-use policy, and announce that its model will not assist in the synthesis of chemical weapons — all without changing a single decision that affects the millions of ordinary users interacting with its products every day. The bar is high in drama and low in practice. Everyday harms, by contrast, require unglamorous and continuous work: auditing datasets, revisiting product decisions that generate revenue, and accepting that the definition of harm extends beyond the lurid cases.
The regulatory environment has broadly followed the industry's lead. Landmark AI legislation in various jurisdictions has devoted significant attention to defining and guarding against high-risk applications in the abstract, sometimes without building the monitoring infrastructure that would catch lower-level harms as they compound. The United States has lagged on comprehensive AI legislation, and the policy debate there has often bifurcated between those focused on frontier model safety and those focused on civil rights applications — with limited overlap between the two communities.
The likely consequences of this imbalance fall unevenly. People who are already subject to automated decision-making in high-stakes settings — benefits determinations, parole assessments, loan applications — are the most exposed to everyday harms and typically the least represented in the rooms where safety definitions are negotiated. If the policy conversation continues to treat catastrophe-avoidance as synonymous with safety, those communities will continue absorbing costs that never register as policy failures, because the systems producing those costs will never trip the alarm.
For companies, the medium-term risk is arguably regulatory backlash driven not by an AI catastrophe but by accumulated grievance. If the gap between what companies claim their safety commitments mean and what affected users actually experience grows wide enough, the political response could be less targeted and less thoughtful than anything the industry would prefer.
What to watch in the coming months is whether the bodies currently drafting AI governance frameworks — including those with active legislative processes — begin incorporating metrics for diffuse, everyday harm rather than relying almost exclusively on prohibited-use lists. Equally worth watching is whether the academic and advocacy communities working on algorithmic accountability manage to build more durable coalitions with the mainstream AI safety movement, which has historically been skeptical of what it sees as a distraction from larger risks. MIT Technology Review's intervention is a sign that at least some influential voices are pushing back on that hierarchy. Whether that pressure shifts institutional priorities is the open question.
