AI Safety Focuses on Catastrophic Risk, Ignoring Everyday Harms

The criteria for "success" in tech policy are often telling. An AI company promises its chatbot won't encourage suicide; a social media platform makes it harder for children to buy fentanyl—these protections matter, and they should always be in place. But companies can easily clear these low bars while continuing to make products that harm daily life. A commentary piece from rss, "The focus on AI safety disasters ignores everyday harms," sharply points out this misalignment: the industry's and regulators' obsession with "catastrophic" AI risk is causing us to overlook the more insidious harms that occur every day.

How disaster narratives crowd out discussion of everyday harms

The current AI safety debate is dominated by "catastrophic risk": extreme scenarios such as superintelligent loss of control, mass unemployment, and bioweapon design occupy the policy agenda and media headlines. This narrative has its rationale, but the problem is that it creates a binary opposition—either it's a disaster threatening human existence, or it's an insignificant minor issue. Everyday harms fall precisely into this blind spot. They aren't dramatic enough to inspire urgency, yet they genuinely affect the experiences of millions.

The original article uses two examples to illustrate this distortion: a chatbot promises not to encourage suicide, and a platform promises to limit children's access to fentanyl. These commitments themselves deserve recognition, but they are only minimum safety baselines, not remarkable achievements. Companies can publicly celebrate being "responsible" while their products continue to exacerbate bias, erode privacy, create information cocoons, or induce overuse. These everyday harms won't make headlines, but they keep accumulating.

Superintelligent loss of control is a core catastrophic risk concept in the AI safety debate, stemming from concerns that AI development may surpass human control and produce irreversible consequences. The concept is typically closely tied to the "alignment problem"—how to ensure that highly autonomous AI systems have goals consistent with human values. This narrative has considerable influence in academia and tech policy circles, partly because of the irreversibility of its extreme consequences and the strong emotional resonance it evokes. However, critics point out that over-focusing on superintelligent loss of control may produce a "risk displacement" effect, diverting policy resources and public attention away from more frequent, more life-adjacent everyday harms such as algorithmic bias and privacy erosion, thereby distorting the reasonable prioritization of safety.

Why low standards can pass "safety" checks

Companies are adept at packaging compliance as innovation. A chatbot passing a few safety tests doesn't mean it won't reinforce users' negative emotions or spread misinformation over long-term interactions. A platform adjusting the accessibility of certain content doesn't mean its recommendation algorithms no longer amplify extreme content in pursuit of engagement.

The root of the problem lies in the short-sightedness of evaluation systems. When "safety" is defined as avoiding extreme events (such as suicide, violence, crime), companies only need to set up a last line of defense to claim compliance. Meanwhile, everyday harms—such as subtle social bias, addictive design, and systemic impacts on vulnerable groups—are difficult to quantify and difficult to turn into PR material. As a result, they are excluded from safety assessments and aren't even considered "safety" issues.

The original article reminds us that this low standard is accepted partly because the public and regulators also view AI through a disaster lens. As long as no "big news" occurs, everyday harms are tacitly accepted as an acceptable cost.

Compliance packaging refers to companies shaping behavior that meets minimum regulatory requirements or passes basic safety tests into remarkable safety achievements or ethical innovation in their external messaging. This phenomenon is particularly common in the tech industry because safety standards are often based on the bottom line of "whether extreme negative events occur" rather than "whether daily user experience suffers systemic harm." Companies can claim compliance by setting up a last line of defense (such as content filtering, age verification), while deeper issues like algorithm design, addictive mechanisms, and bias amplification are excluded from compliance scope. This strategy exploits regulatory lag and the public's default mentality of "no disaster means safe," allowing low standards to persist long-term and become industry convention.

Redefining AI safety: from exceptional events to everyday experience

To change this situation, the focus of AI safety needs to shift from "preventing disasters" to "improving daily life." This is not to deny the importance of catastrophic risk research, but to require us to simultaneously attend to high-frequency, low-intensity harms. For example, does a chatbot foster users' loneliness in long-term conversations? Does a recommendation system silently narrow teenagers' information horizons? The answers to these questions won't appear in safety statements, yet they determine technology's real impact on people.

A more reasonable evaluation framework should include continuous monitoring and disclosure of everyday harms. Companies need to report negative patterns generated by products in daily use—such as certain user groups being disproportionately affected, or certain interaction designs causing unintended psychological consequences—just as they report safety incidents. Regulators, meanwhile, should incorporate everyday harms into compliance considerations rather than simply waiting for disasters to happen before acting.

The core insight of the original article is that defining safety as "the absence of disaster" is a dangerous simplification. True safety means products don't cause avoidable harm in everyday use. This requires more detailed metrics, more transparent data, and a corporate culture willing to challenge low standards.

Recommendation systems are the core algorithmic mechanisms driving information distribution on social media and content platforms, typically aimed at maximizing engagement (such as clicks, dwell time) by predicting user preferences. While this optimization logic enhances user stickiness, it may also lead to the "information cocoon" effect—where users are continuously exposed to homogeneous information, their horizons gradually narrow, and extreme content is amplified. The everyday harms of recommendation systems are insidious and cumulative: a single recommendation may seem harmless, but long-term effects may reinforce bias, exacerbate social polarization, or induce overuse. Because these impacts are difficult to quantify in short-term safety tests, recommendation systems often fall outside traditional safety assessments, becoming a gray area in everyday harm regulation.

Conclusion: low standards are not success but a starting point

Companies promising chatbots won't encourage suicide, or platforms limiting children's access to fentanyl—these are necessary but far from sufficient. They are only a starting point, not a sign of success. If the success criteria for tech policy remain at these low bars, we will systematically ignore the more pervasive and insidious harms.

AI safety needs a shift in perspective: from fearing disasters to focusing on the everyday. This is not lowering requirements but raising standards—requiring technology not only not to collapse in the most extreme scenarios but also not to harm people in ordinary daily use. The original article's criticism hits the mark: when we only stare at disasters, everyday harms become tacitly accepted costs. It's time to turn the spotlight on those neglected everyday issues.

One-line Summary

AI safety discussions over-focus on catastrophic risk, causing everyday harms to be systematically ignored; we need to shift toward evaluating everyday experience.

Paragraph Summary

The current AI safety debate is dominated by catastrophic risks such as superintelligent loss of control and mass unemployment. This narrative creates a binary of "disaster or insignificant," causing everyday harms to fall into a blind spot. Companies can claim compliance by setting minimum safety baselines (such as chatbots not encouraging suicide) while continuing to create insidious harms like bias, addiction, and privacy erosion. The article points out that short-sighted evaluation systems and the public's disaster lens together lead to the acceptance of low standards. To change this, the focus of AI safety must shift from preventing disasters to improving everyday experience, establishing continuous monitoring and transparent disclosure mechanisms for everyday harms, and defining safety as products not causing avoidable harm in daily use.