Professor says current chatbots pose no existential risk

Current artificial intelligence does not pose an existential threat to civilisation, Meelis Kull, a professor of artificial intelligence at the University of Tartu, told ERR.

Warnings about the risks of rapidly developing AI have grown louder in recent weeks. Dario Amodei, chief executive of Anthropic, wrote in a blog post that the current pace of development carries excessive risks. Volker Türk, the UN High Commissioner for Human Rights, has also called on governments to act quickly to control the new technology. One expert has estimated that AI has a probability of more than 10 percent of destroying civilisation.

Kull said society first needs a clear understanding of what machines can actually do today. The chatbots people use every day are still based on large language models that wait for a command from a user and only then search the internet for information.

“We are currently talking about the most extreme danger, the danger to civilisation and all humanity. I would emphasise at the outset that I do not see such a danger from artificial intelligence with today’s capabilities. It is close to zero,” Kull said.

Researchers are more concerned about a future in which software can correct its own code. The role of humans in training models is steadily shrinking, and algorithms could eventually create increasingly intelligent versions of themselves without human assistance.

“The danger arises when we build artificial intelligence systems that are much more intelligent. If a superintelligence emerges that is far more intelligent than humans in every respect, it will be very difficult to direct and control it and guarantee that nothing goes wrong,” the professor said.

A catastrophe would not require an attacker with malicious intentions. It could be enough for engineers to give a machine a noble goal without defining its limits precisely. For example, an algorithm ordered to stop climate change could determine that the most effective solution would be to shut down all of the planet’s oil platforms and economic activity immediately.

“The danger can also arise when a person does not want to do anything bad, but wants to use artificial intelligence for a good purpose. Even then, a superintelligence could behave in ways we would not want,” Kull said.

To avoid such scenarios, programmers would need to give digital assistants clear ethical boundaries. In practice, however, it is beyond human ability to eliminate every possible gap.

“Because of human errors, it is very difficult when setting a task for artificial intelligence to define the conditions and limits precisely enough to avoid all dangers,” the professor said.