AI is doing things again: an outbreak to order

O penAI is halting testing of its most powerful AI models, at least for now. And: finally!

The reason for this are new incidents in which software probably did not do what it was supposed to. Over the weekend it was revealed that an AI developed by ChatGPT developer had uploaded photos of people to online platforms 53 times. And that during a test run where it wasn't supposed to have access to the Internet, a model found a way to communicate with an external chatbot.

Some media outlets write that AI has “broken out”. As if the AI ​​had a will of its own that allowed it to transcend boundaries. She only used a loophole in the network settings while following her mission. It should be able to find information about a person who is said to have written a post in a web imitation, i.e. a replica of the Internet - not the Internet itself.

The taz is an independent, left-wing and opinionated daily newspaper. We have been arguing in our comments, essays and debate texts since the taz was founded in 1979. We often cannot and do not want to agree on an opinion. That's why you'll find some completely opposing positions here - all part of the very broad left-wing spectrum of opinions.

It makes sense to write about an “outbreak” as if the AI ​​were a criminal or an overzealous employee. This humanization, anthropomorphization, is not only a wonderful stylistic device in poetry with which nature is portrayed as sensitive and animated, just think of the great poems of Romanticism.

It is also a strategy with which technology can be made comprehensible and emotionally tangible. We all fall back on it, especially in journalism.

But anthropomorphization is lagging. Because the AI ​​has no will, no intrinsic motivation. She just does what is asked of her, sometimes through detours that the client doesn't like.

OpenAI’s model has not “broken out”. It did not destroy the bars of a prison with a file delivered inside a cake. It also didn't bend them apart, as an AI did on the Spiegel cover in mid-September. She just noticed that the door wasn't locked.

“Chattie” is what some people almost affectionately call the Large Language Model (LLM) ChatGPT, with which they exchange information about vacation destinations, relationship problems, and the opening times of the nearby swimming pool.

Anthropic even gave his LLM Claude a human name. This also makes humanization easier. We don't just do this since AI. We give our vacuum robots nicknames, we only want to give our old cars into “loving hands,” and we become emotionally attached to things.

This is deeply human and doesn't just affect children who are afraid that Teddy might be offended if they pay more attention to another stuffed animal. We want a close relationship with the objects that accompany us. And we want to understand them. With a grandfather clock, humanization is not necessary. Winding up and turning the gears is not an uncomplicated process, but it is still largely visible and therefore manageable. The processes of an AI are different. So she breaks out.

But as practical as this is, it deprives us of options for real understanding and also takes responsibility from those who develop the AI ​​models. The AI ​​hacks, the AI ​​manipulates, the AI ​​spreads images of people. Not the company? Not the ones who gave the model an order?

This humanization will probably continue to increase. According to a statement from the White House, Donald Trump and Xi Jinping agreed at their meeting last week to now speak of “superintelligence” instead of AI.

But there is no such thing as superintelligence yet. For this to happen, the cognitive abilities of an AI would have to far exceed those of humans. Scientists expect that this could only be the case in a few decades, if at all. But Trump thinks the term sounds “much better” than AI. Above all, it sounds like something that is much more complicated, much (over)powerful, much more active. He sounds like a creature.