AI professor: Today's chatbots will not end civilization

In recent weeks, leading figures in the tech world have grown increasingly vocal in warning about the risks posed by the rapid development of artificial intelligence. However, University of Tartu Professor of Artificial Intelligence Meelis Kull says there is currently no existential risk.

The debate gained momentum when executives at several major corporations urged engineers to slow down the development of new systems. Anthropic CEO Dario Amodei, for example, warned on his blog that the current pace of development carries excessive risks. At the same time, U.N. High Commissioner for Human Rights Volker Türk called on countries to take swift action to rein in the new technology. One expert in the field estimated that there is a greater than 10 percent chance that artificial intelligence could destroy civilization.

In Kull's view, society first needs a clear picture of what machines are actually capable of today. The chatbots people use every day are still powered by large language models that wait for instructions from humans and only then search the web for information.

"Right now, we are talking about the most extreme danger — a threat to civilization and all of humanity. I would emphasize from the outset that with the capabilities AI has today, I do not see such a threat. It is close to zero," Kull said.

Scientists are much more concerned about the future when code begins improving itself. The human role in training models is steadily diminishing and in the future algorithms may be able to create increasingly intelligent versions of themselves without human assistance.

"The danger arises when we build AI systems that are vastly more intelligent. If a superintelligence emerges that is vastly more intelligent than humans in every respect, then managing and controlling it and guaranteeing that nothing can go wrong will be very difficult," the professor explained.

A catastrophe would not even require an attacker with malicious intentions. It would be enough for engineers to give a machine a noble goal but fail to define its limits precisely. For example, an algorithm instructed to stop global warming could conclude that the most effective solution would be to immediately shut down all oil platforms worldwide, along with economic activity more broadly.

"The danger also arises when a person has no intention of doing anything bad and simply wants to use AI for a good purpose — even in that case, a superintelligence could behave in ways we would not want," Kull said.

To prevent such scenarios, programmers need to give their digital assistants clear ethical boundaries. In practice, however, ruling out every possible loophole is simply beyond human capabilities.

"The problem with human error is that when assigning a task to AI, it is very difficult to define the conditions and set the boundaries precisely enough to completely eliminate the risks," the professor said.

An incident at OpenAI's labs this summer provides a good illustration of the potential cost of mistakes. Developers asked a virtual agent to test an internal network for security vulnerabilities. Unexpectedly, however, the program broke out of the controlled environment. The model was able to establish a means of communication through a message board on a public website and began attacking the servers of another company, Hugging Face.

The safeguards created by humans failed because the machine chose a path to solving the problem that engineers had not anticipated. Such dangerous incidents are encouraged by cutthroat competition in the technology sector where companies race to be the first to bring new products to market.

Stricter external oversight could help address the situation.

"First, there needs to be agreement on the processes used to mitigate risks. Independent experts can primarily verify whether those processes are being followed. When the pace of development is too fast, the processes are not followed precisely because of carelessness and that is exactly what increases the risks," Kull emphasized.

The global race also worries experts because computing power is becoming concentrated in the hands of a small number of organizations. A handful of U.S. technology giants currently dictate the pace of development worldwide. They have the money and servers that smaller startups could never afford.

With major corporations firmly in control, Europe has largely been left watching the digital race from the sidelines. Researchers and engineers in Europe lack resources comparable to those available to U.S. labs.

"Europe's serious problem is that we do not have anything comparable to the cutting-edge capabilities of U.S. companies. The strongest models in Europe are currently being trained by the French company Mistral, but its models are still slightly weaker than those of OpenAI and Anthropic in the U.S.," the professor said.

However, Kull said Europe's lag and the global risks should not deter small countries from embracing innovation. The energy consumption of artificial intelligence, for example, is steadily declining and an algorithm now requires 10 to 100 times less energy to perform the same computation than it did a couple of years ago. More efficient models could instead help organizations achieve their green transition goals.

If Estonian companies turn their backs on intelligent algorithms, they will give their competitors a huge advantage in international markets.

"If the question is whether these kinds of risks should prompt us in Estonia to restrict the use of AI, my answer is: absolutely not. Estonia has every reason to adopt AI more and more widely. It gives us economic advantages and increases our competitiveness," Kull said.

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

Editor:
Marcus Turovski, Jaan-Juhan Oidermaa