Hallucinations and AI errors could lead to scientific breakthroughs

Artificial intelligence (AI), which can convincingly create false or fictitious information, can become an indispensable tool in the hands of scientists. New drugs, weather forecasting and device inventions are just a few examples of how AI’s ability to generate new, seemingly flawed ideas can change the world of science.

Image source: Copilot

AI models are often criticized for their tendency to generate false information, passing it off as facts – so-called “hallucinations”. These believable fakes not only confuse chatbot users, but also lead to lawsuits and errors in medical records. For example, last year, a false claim from a Google chatbot caused the market value of one company to collapse by $100 billion. However, according to The New York Times, researchers recently discovered that AI “hallucinations” can also be surprisingly useful.

It turns out that smart LLMs (large language models) can generate incredible ideas that help scientists fight cancer, develop new drugs, create medical devices, study weather phenomena, and even win a Nobel Prize. “This gives scientists new ideas that they might never have come up with on their own,” explains Amy McGovern, head of the NSF AI Institute in the US for climate and weather forecasting.

Although science is traditionally associated with logic and analytics, its beginnings are often associated with intuition and bold assumptions. Philosopher and scientific methodologist Paul Feyerabend once described this stage as “anything goes.” Errors in AI models enliven this creative process, accelerating the discovery and testing of new ideas. What once took years can now be done in days, hours or even minutes. Massachusetts Institute of Technology (MIT) professor James J. Collins recently praised hallucinations for accelerating his research into new antibiotics as AI models created entirely new molecules.

«AI hallucinations occur when scientists train generative computer models on a particular topic and then allow the machines to process the information. The results can be both absurd and ingenious. For example, in October 2023, the Nobel Prize in Chemistry was awarded to David Baker for his research in the field of proteins. His approach, which uses AI to create entirely new proteins, was previously considered nearly impossible. Baker noted that AI generation has become the basis for developing “proteins from scratch.” “We have created 10 million new proteins that do not exist in nature,” says the scientist. At the same time, his laboratory has already received about 100 patents, including technologies for treating cancer and combating viral infections.

However, the term “hallucinations” is controversial. Many people prefer to call AI results guesses or promising ideas because they are often based on real scientific data. For example, Caltech professor Anima Anandkumar believes that using the term “hallucinations” can be misleading and notes that scientists try to avoid it.

It’s worth noting that some experts are concerned that the pace of scientific discovery has slowed in recent decades. However, AI proponents argue that new capabilities can take science to another level. In particular, Baker and his colleagues envision a future in which protein catalysts harness solar energy, modernize factories, and help create a sustainable world.

Other scientists also share this optimism. “It’s incredible how quickly things are moving,” says Ian C. Haydon, a member of Baker’s team. And Pushmeet Kohli from DeepMind emphasizes that AI is capable of unexpected but ingenious moves. “We clearly have an amazing tool that can be creative,” he concluded.

Despite advances, artificial intelligence “hallucinations” remain a controversial topic. While some scientists see them as a useful tool, others are wary of their negative consequences, such as errors in medical data. However, the scientific community recognizes that the potential of AI to accelerate scientific discoveries is enormous.