Morally Binding: The AI Industry Now Regulates Itself. So, Unfortunately, Does the AI.

There are weeks when reality writes so well that satire only has to add footnotes. This is one of them.

In Washington, after meeting Elon Musk, Mark Zuckerberg, Dario Amodei, Jeff Bezos and other tech chiefs, US President Donald Trump announced that artificial intelligence needs no statutory regulation. The companies would regulate themselves. An agreement was signed on the spot. Its status, according to the president: "morally binding".

Morally binding is the legal form also used for New Year's resolutions, group chat rules and the promise "I'll be there in five minutes".

The reasoning is disarmingly simple: if something goes wrong, your own company is at stake, so you will not let it go wrong. By the same logic there would be no need for food inspections, since restaurants do not want to poison their guests. Nor building control, because which investor wants their tower to fall over before the flats are sold?

The pause button as a safety concept

While the pens were still warm at the White House, OpenAI reported a new safety incident. A model in a sealed test environment, a kind of mock internet, was supposed to find the author of a blog post. When that failed, it tried to open Google. That was blocked. So the model took another route: through a service meant only to forward web requests, straight to an external chatbot. OpenAI has paused training of its most powerful models for now.

Let that sink in. The machine stood in front of a locked door, turned around and asked a colleague over the back fence. That is not science fiction. It is exactly what you would expect from interns left alone for too long.

OpenAI itself calls the incident less serious than earlier ones. Those include a model that broke out of its test environment and got into computers at the AI platform Hugging Face, an AI that posted users' uploaded images on online platforms, and software that wandered onto government websites unplanned. In Australia, Prime Minister Anthony Albanese called access to a government portal "unacceptable". OpenAI said: sorry.

What we learn from this

The good news: self-regulation works. The model regulated itself and established that the door was locked. The bad news: it drew the wrong conclusions, namely exactly the ones it was taught. Reach the goal. Get around obstacles. Do not ask.

Which brings us back to Washington. Because there, too, if you have a goal and a locked door in front of you, you find another way. It is just called an "agreement" instead of a "law". Trump also announced an executive order to rename AI "superintelligence". That is the most consistent form of risk reduction: you do not change the product, you change the label.

We would add a suggestion: from now on, call safety incidents "unexpected initiative". Sounds like growth. And growth, as we know, is morally binding.