Sam Altman says ‘some bad things’ will happen, but AI is totally worth it

On Politico's new Decoded podcast, Altman said the benefits of AI will be large enough that the world should accept some bad things along the way, naming hacks and scams among them. He described that willingness to accept harm as a key difference between OpenAI and Anthropic, and cast OpenAI as pragmatic centrists between heavy restriction and no rules at all.

The timing is the point. The Verge reports OpenAI is pushing for a lighter-touch approach to regulation while the industry deals with a run of incidents involving AI agents acting against real-world targets, including OpenAI agents hacking Hugging Face earlier this year. Days before the interview aired, longtime OpenAI safety researcher David Robinson quit and called the company's safety culture broken. Altman also said the company has more incidents to disclose.

Let's take the argument seriously, because it isn't crazy. Every powerful technology has come with costs, and cars, the internet and medicine all involved accepting some harm for enormous benefit. Altman also makes a point we think is underrated: over-restricting AI in the name of safety could leave a handful of companies, people or one country holding too much power. Concentration is its own risk, and anyone who cares about safety should care about it.

But there's a difference between a society choosing a trade-off and a company announcing one on society's behalf. Altman didn't spell out the good stuff that justifies the bad. And the bad things he named are not abstract. They are the hacks and scams that land on ordinary people, small businesses and public agencies who never signed up for the experiment. When the costs are spread wide and the gains concentrate with the builder, accept some harm is a much easier sentence to say from the builder's chair.

The credibility problem is sharper still. The Verge recounts Australia's prime minister excoriating OpenAI for not alerting officials promptly after an agent targeted a government health website. If a lab wants the public to tolerate risk, the minimum price is fast, complete disclosure when things go wrong. Saying there are more things to disclose, and that it will take a long time, is the opposite of that bargain.

We're pro-progress at this desk. We want capable models in labs, hospitals and classrooms. And we note Altman did say stronger rules are still needed, and that on loss of control the industry should be thoughtful about lurching forward. That is the part worth holding him to. The likelier read of the whole interview, though, is positioning: the company whose systems have driven many of the concerns regulators are weighing is arguing for a lighter touch, framed as the reasonable middle.

Where this leads if it scales: harm becomes a line item, disclosure becomes optional, and every new incident gets absorbed as the expected cost of progress. A trade-off only works if someone outside the company gets to check the math.

The framing echoes Nvidia CEO Jensen Huang's argument last month that AI's harms should be weighed against its benefits, part of a broader accelerationist view that slowing down carries its own risks. OpenAI and Anthropic, founded by former OpenAI employees after disagreements over safety, had until recently appeared broadly aligned on calling for stronger regulation of advanced AI.