'We'll have to accept some bad things': the moral math Altman said out loud and nobody wants to do

WHAT ALTMAN SAID. On October 4, 2026, in an interview with Politico's Decoded podcast, Altman argued that the world should accept 'some bad things' happening in exchange for AI's benefits and for people having access to it. According to coverage by Fortune and TechRepublic, he expects the technology to be used for 'orders of magnitude' more good than bad. He drew a line. Bounded, understood risks would be tolerable. Catastrophic risk, such as losing control of AI, is not. He also rejected the idea that a single lab should control the technology to prevent harm, calling that a 'completely unacceptable trade-off.' Context matters. The remark came days after Donald Trump announced, on September 29, a voluntary safety accord with the major AI companies. Asked whether it was binding, Trump called it 'morally binding.' According to press reports, the accord carries no consequences for a company that ignores it.

A BARGAIN WE ALREADY SIGNED. The underlying question is uncomfortable: how many victims will a society tolerate in exchange for a technology? With cars, the answer is a lot. The WHO estimates that roads kill about 1.19 million people a year. It is the leading cause of death for people aged 5 to 29. We accept that figure in exchange for mobility, with barely a debate. Aviation chose the opposite path. It treated every crash as a system failure, with mandatory, independent investigations. According to IATA's safety report released in March 2026, there is now one fatal accident every 5.6 million flights (2021-2025 average). A decade ago it was one every 3.5 million. Medicines took a third route: regulation was born from tragedy. In 1937, a sulfanilamide elixir killed more than a hundred people in the US and led to the 1938 law requiring proof of safety before sale. Thalidomide left thousands of babies with birth defects worldwide, and in 1962 the US began requiring proof of efficacy too. Electricity and X-rays were adopted with enthusiasm and with casualties, such as the women who painted watch dials with radium, before rules existed. The lesson is clear. No technology ever arrived risk-free. The difference between cars and planes was not the technology. It was the institutional decision to count harms and learn from each one.

HOW TO DO THE MATH WITHOUT DEMAGOGUERY. Tools for this calculation exist, cold as they sound. The first is the value of a statistical life (VSL). It is not the price of a person. It is what a society is willing to pay to slightly reduce the risk of death for many people. The US Department of Transportation puts it at $13.2 million (in 2023 dollars), according to a 2024 Department of Homeland Security memo. That number decides which rules get adopted. The second is a clash of two philosophies. The precautionary principle says don't release something until it is shown to be safe. 'Permissionless innovation' says release it and fix problems as they appear. The third comes from engineering: 'acceptable risk.' You set an explicit threshold and design so you never exceed it, as with bridges or nuclear plants. Our reading: Altman sides with permissionless innovation, plus a catastrophe threshold. That is a defensible position. The problem is that he doesn't say who measures, with what data, or who pays when the harm lands on the 'acceptable' side.

THE AI LEDGER TODAY. Some harms can already be counted. According to the FBI Internet Crime Complaint Center's (IC3) 2025 report, complaints referencing AI added up to about $893 million in losses across more than 22,000 cases. People over 60 lost $352 million of that. Voice-cloning scams, where a synthetic voice imitates a relative in trouble, are part of that total. According to the ACLU, there are at least 14 known US cases of people wrongly arrested after a false facial-recognition match. The group believes these are only the cases that have come to light. OpenAI faces roughly two dozen lawsuits alleging that ChatGPT played a role in suicides and self-harm. The company denies its product caused those deaths. In January 2026, Character.AI and Google agreed to settle several similar suits, without admitting liability. Benefits can be counted too. In Sweden's MASAI trial, with about 106,000 women, AI-assisted breast screening detected 29% more cancers, with no rise in false positives and 44% less reading work for radiologists. Whether that lowers mortality is still unknown. Waymo says its driverless taxis had 82% fewer injury crashes than human drivers over 270 million miles. Those are the company's own figures, compared against a statistical model of human drivers, not an independent audit.

OUR READING. Altman is saying something true out loud. Every powerful technology causes harm, and demanding zero risk means giving it up. If AI helps catch more cancers or prevent crashes, the cost of not using it is also measured in lives. But the line hides three traps. First: the party deciding which harm is acceptable is the party selling the product. Aviation is safe because manufacturers don't investigate their own crashes. Second: harms and benefits don't fall on the same people. The retiree scammed by a cloned voice doesn't get the earlier diagnosis. The person wrongly arrested doesn't ride in a robotaxi. Third: an accord that is 'morally binding' with no consequences is not an acceptable-risk framework. It is a promise. That is why we think AI should resemble the airplane in governance, even if it spreads like the car. That means public incident registries, independent investigation of serious cases, and clear legal liability.

IMPLICATIONS. In the short term, the harms are real and growing: automated fraud, policing errors and effects on vulnerable people. The pending lawsuits will decide whether the law treats chatbots as products, with a duty of safety, or as protected platforms. That ruling will matter more than any voluntary accord. In the long term, we remain optimistic. AI that speeds up diagnosis, cuts road deaths and makes medical research cheaper could save far more lives than it puts at risk today. But that doesn't happen on its own. Cars took decades to get seatbelts and airbags, and they got them because someone counted the dead. Aviation reached its record because every crash was investigated to the bottom. If we want the horizon of abundance and health, Altman's math has to be done, but in public, with verifiable data and with someone accountable. Accepting some bad things is only reasonable if those things don't happen twice.