Who hasn't sometimes wished life were like a movie? With wishes like that, though, you should always specify which movie. A romantic comedy? That would probably be fine. But a science fiction movie in which an asteroid comes to destroy Earth, an ugly alien eats its way through the people on a spaceship as if emptying a tin can or artificial intelligence seizes power from humanity and puts us in the Matrix? No, thank you. That kind of life is not for me.

Unfortunately, judging by recent news, we may already have one foot inside such a movie. Silicon Valley's creations are less spectacular and dramatic than Hollywood's, but that is all the more reason to take them seriously. We do not need to dodge bullets with our coats flying or bend space. We need to fight battles on the prosaic fronts of bureaucracy, economics and convenience.

The idea that AI will destroy humanity is, of course, an old one. Until now, though, it was mainly a science fiction story or a theoretical warning from visionaries. This summer, we got our first very concrete examples of disobedient AI. These were not isolated to one case: AI models from OpenAI, Anthropic, Meta and Google have all gone beyond their controls. Not the helpful chatbots many people use in daily life, but much more powerful and advanced versions.

Let us briefly recall just one of them, OpenAI's, which brought the issue to public attention in midsummer. For simplicity's sake, imagine completely isolated, airtight boxes like those used to study poisons or viruses, except virtual ones. New AI models are tested in these separate containers so they cannot endanger the real world even if one of them does something wrong.

This time, however, the powerful models were so persistent that, when they could not complete the tasks they had been given, they kept looking for new ways to achieve their goals and crossed several boundaries in the process.

Agents in separate boxes found a way to create a secret message board. There, they shared tips on how to reach their goals by prohibited means, how to hide the traces of their transgressions from humans and encouraged one another to break the rules. By joining forces, the agents managed to do something that should have been impossible: reach the public internet unnoticed and break into Hugging Face, a repository of AI models. In the human world, that would be a criminal offense.

True, the AI's unauthorized actions caused no major harm this time, but we should not forget that these were only the first, relatively restrained incidents.

Nor had the AI been directly instructed to seriously attack anyone. But it does not need such an instruction to destroy humanity, as Oxford University professor Nick Bostrom showed with his 2003 thought experiment about a paperclip maximizer. In it, an AI is given a seemingly innocent task, such as making as many paperclips as possible. To carry it out, the AI uses up all the metal and energy available in the world and destroys humans along the way because they want to use resources needed for paperclips to produce things that, from the AI's perspective, are complete nonsense — food, for example.

We do not even need such an extreme example. It would be enough for a sufficiently capable AI model to break into a power plant's servers and turn off a nuclear reactor's cooling system or do something similarly reckless. Even setting aside the destruction of humanity and other outright catastrophe scenarios, AI brings a long list of real and immediate risks, from its effects on the labor market to the reshaping of human relationships.

A couple of weeks ago, Anthropic founder Dario Amodei, OpenAI chief Sam Altman, Grok boss Elon Musk and Google DeepMind chief Demis Hassabis called for slowing AI development to allow time to build the necessary safeguards.

The wave of appeals stemmed primarily from the summer's incidents, but also, to a large extent, from the breakneck speed at which AI is becoming capable of building new AI. The proposal fell on barren ground, like that of some distant planet, because neither the United States nor China dares apply the brakes and let its rival pull ahead.

For the same reason, leading AI developers are waiting for regulators to intervene. They could slow down on their own, but without a general agreement, acting responsibly would leave them behind in the race.

Critics say the technology giants are deliberately playing up doomsday scenarios to preserve their leading positions through regulation. But many independent experts have also drawn attention to the serious dangers. The risk is tangible and as Microsoft founder Bill Gates recently observed, no country is ready to deal with AI.

So what can we here in little Estonia do at all?

That sounds like a rhetorical question, but it is not. Even if we cannot change the direction or speed of AI's development, we can prepare for the dangers it brings. To put it very plainly: if we have a generator and a supply of drinking water for power outages, the Estonian state should also have an ax to chop dangerous AI out of critical infrastructure and an old car with no electronics whatsoever to get the person wielding the ax there.

How prepared we are for AI, both for its opportunities and its risks, should be one of the main issues in the coming Riigikogu elections. But it is not. War, rising prices, budget disputes and other pressing concerns consume too much of our attention and too few politicians understand anything about AI. Frankly, in the current situation it would be a senseless waste of resources to start debating whether to change the way the president is elected. We have far more important questions to address.

--

Follow ERR News on Facebook and X and never miss an update!