‘AI red lines’ campaign launches at U.N. Here’s what they might be.

The AI ​​Red Lines initiative was launched on Tuesday at the United Nations General Assembly – the ideal place for a very non -specific declaration.

More than 200 laureates of the Nobel Prize and other experts in artificial intelligence (including the co-founder of Openai Wojciech Zaremba), as well as 70 organizations that deal with AI (including Google Deepmind and Anthropic), have signed a letter of call for “world red lines to prevent the risks of unacceptable AI”. However, he was marked as much by what he did not say that what he did.

“AI systems have already presented a deceptive and harmful behavior, and yet these systems are granted more autonomy,” said the letter, fixing a deadline of 2026 so that its recommendation is implemented: “An international agreement on clear and verifiable red lines is necessary to prevent universally unacceptable risks”.

Fairly fair, but What Red lines, exactly? The letter only indicates that these parameters “should rely on and apply existing world executives and voluntary commitments of companies, ensuring that all advanced AI providers are responsible for shared thresholds”.

The lack of details may be necessary to maintain a very loose coalition of signatories together. They include Alarmists of AI like Geoffrey Hinton, 77, the so-called “sponsor of the AI ​​who has spent the last three years predicting various forms of misfortune of the imminent arrival of AG (general artificial intelligence); the list also includes skeptics of AI like the cognitive scientist Gary Marcus, who has spent the last three years to tell us that No come soon.

What could they all agree on? Moreover, that governments could already be connected to AI, mainly the United States and China, and trust themselves to implement? Good question.

This tweet is currently not available. It can be loaded or has been deleted.

The most concrete answer by a signatory probably came from Stuart Russell, professor of computer veteran at UC Berkeley, following a previous attempt to talk about the red lines at the World Safety of the IA 2023. In an article entitled “Make IA Safe or Making Safe IA?” Russell wrote that AI companies offer “attempts after reducing unacceptable behavior once AI system has been built”. He contrasting that with the approach of red lines: ensuring an integrated safety in the design from the start, and the “unacceptable behavior” will not be possible in the first place.

“It should be possible for developers to say, with great confidence, that their systems will not present harmful behavior,” wrote Russell. “An important side effect of red line regulations will be to considerably increase the safety engineering capacities of developers.”

In his article, Russell has gone up to four examples of red line: AI systems should not try to reproduce; They should never try to enter other computer systems; They should not be allowed to give instructions on the manufacture of bio-armes. And their production should not allow “false and harmful statements on real people”.

From the point of view of 2025, we could add red lines that treat current threats from AI psychosis and AI chatbots that could be manipulated to give advice on suicide.

We can all agree on this, right?

The problem is that Russell also believes that no large tongue model (LLM) is “capable of demonstrating compliance”, even with its four minimum red line requirements. For what? Because these are predictive word engines that fundamentally do not understand what they say. They are not able to reason, even on basic logical puzzles, and more and more “hallucinate” responses to satisfy their users.

The real safety of the AI ​​red line, no doubt, would mean that none of the current AI models would be authorized on the market. It doesn’t bother Russell; As he points out, we do not care that conformity is difficult with regard to medicine or nuclear energy. We regulate whatever the result.

But the idea that AI societies will simply be voluntarily closed their models until they can prove to regulators that no damage will return to users? It is a larger hallucination than anything that Chatgpt can find.