AI safety debate ignores everyday harms, critics say

An opinion piece in MIT Technology Review argues that the AI safety debate has become too focused on catastrophic scenarios and not focused enough on the everyday harms that AI systems can cause. The authors contend that success in technology policy is often defined too narrowly. They point to examples such as a chatbot that promises not to encourage suicide or a platform that makes it harder for children to buy fentanyl. Those protections matter, but they do not capture the routine ways AI can mislead people, discriminate against groups, exploit users or erode public trust. The critique is a reminder that safety is not only about preventing rare, dramatic failures. It is also about the ordinary failures that occur in search results, recommendations, automated decisions and customer service tools. These problems can be harder to see because they are distributed across millions of interactions and often fall hardest on people with less power to challenge them. The authors want policymakers and companies to widen their definition of harm and to measure success in terms of whether AI systems are fair, honest and accountable in daily use. That would mean looking beyond model benchmarks and disaster readiness exercises to product design, business incentives and user redress. The argument does not dismiss catastrophic risk. Rather, it says that focusing only on extreme outcomes can leave the public unprotected from the smaller but more frequent harms that shape trust in AI. For companies, the message is that safety work should include audits, transparency, complaint mechanisms and evidence that systems work well for affected communities, not just assurances that the worst-case scenario has been considered.