Altman says accept some AI harms. Scam victims never got a vote

He did draw a limit. Altman said he doesn't accept "the really catastrophic risks," including a serious loss of control to AI systems. He also rejected the idea of one lab holding the most powerful models and handing out the benefits, calling it "a completely unacceptable trade-off". An Anthropic spokesperson replied that the rules it backs would apply only to frontier models, meaning the most advanced systems.

So this is not a CEO shrugging at the end of the world. It's a CEO saying everyday harms are the price of wide access. That's a narrower claim, and a more serious one to argue with.

The bad things already have a price tag

The harms Altman named, hacks and scams, are not hypothetical. The FBI's Internet Crime Complaint Center says Americans reported about $20.9 billion in cybercrime losses in 2025, up 26% from the year before. People aged 60 and older reported about $7.7 billion of that, more than a third of the total.

For the first time, the FBI's annual report also counted crimes with an AI link: 22,364 complaints and close to $900 million in losses. Most of that money went to investment fraud. The FBI says scammers are pairing pressure tactics with cloned voices, fake social media profiles, fake ID documents and realistic videos of public figures or loved ones. In other words, AI isn't inventing new crimes so much as making old ones cheaper to run.

Think about who sits behind those numbers. The tactics in the FBI's own findings are cloned voices, fake videos and fake profiles, and older Americans carry more than a third of all reported losses. None of those victims signed up for Altman's trade. The "agency" he wants to protect belongs to the people using the tools, while the losses often land on people who never touched them.

Who decided, and who gets the bill

This is the gap in most of the coverage. Altman says "the world" should accept some harm, but the world didn't hold that vote. A handful of companies did, and the closest thing to a public process so far is a set of voluntary safety standards that AI chiefs signed after a White House lunch with President Trump on September 29, which Trump called "morally binding". Voluntary means a company can step away from them. They come with no liability and no compensation fund.

Compare how we handled the last technology that killed people while making life better. Society did accept car crashes as a cost of mobility. It also built a system to pay for them: mandatory insurance, crash standards, recalls and product liability. Carmakers got a broad licence to sell, and in return they took on real costs when things went wrong. Altman wants the first half of that deal, a lighter touch from regulators. So far, nobody has asked OpenAI to sign the second half.

That's the fair middle position. If OpenAI believes good uses will outnumber bad ones by orders of magnitude, it should be comfortable paying for the bad ones. Some of that money could go into detection, some to victims, and some could come out of lawsuits when its products are shown to have helped cause harm. A company that's confident in the maths of its trade-off shouldn't fear being held to it.

The strongest case for Altman

Altman's side deserves a fair hearing, because parts of it are right. Every general-purpose technology gets misused. Email brought phishing and phones brought robocalls, and nobody seriously thinks we should have banned either. A rule that demanded zero scams would in practice mean zero public access, and that would hand the technology to the few companies or governments big enough to lock it down. That's the outcome Altman warns against when he talks about one lab in San Francisco keeping it all.

He also isn't alone in rejecting doom-first thinking. When Amodei published an essay on September 12 urging the industry to "pace the frontier," Altman replied on X the same day: "I agree with Dario that we need to pace the frontier". The real fight is narrower than the headlines suggest. It's about which ordinary harms justify locking things down.

Critics haven't really answered that point. Zscaler (ZS) CEO Jay Chaudhry told Yahoo Finance that bad things happening is "not acceptable" and that model makers and businesses both need to do better. That sounds good. It also comes from a company that sells protection against exactly these threats, and "try harder" isn't a policy. Others went further. Patrick Gaspard, a distinguished senior fellow at the Center for American Progress and its former president, called Altman's view "freaking insane". That's understandable anger, but anger doesn't tell anyone how to split the costs.

How we'll know who was right

Altman's position gets stronger if AI-linked fraud stays a small slice of total cybercrime, as the roughly $900 million in last year's FBI figures was, while the benefits he promises show up in data we can measure, not just in demos. It gets weaker if next year's FBI report shows AI-linked losses climbing much faster than the overall total, especially among older Americans. It also gets weaker if OpenAI keeps arguing for light rules while opposing any liability for the harm its tools help cause.

The question I'd put to OpenAI is the one its CEO left out. If some bad things are the price of progress, which ones, and how much is OpenAI itself prepared to pay?

So what would you accept? Would a fund that pays AI scam victims change your answer?