Google has announced an experimental artificial intelligence system that uses advanced reasoning to help scientists integrate vast amounts of literature, generate new hypotheses, and propose detailed research plans. “Idea with [the] “AI co-scientists” is about giving scientists a superpower,” says Alan Karthikesalingam of Google.

The tool, which has not yet been officially named, is based on Google’s Gemini major language model. When researchers ask questions or specify goals, they come up with their first idea within 15 minutes, for example, to find a new drug. According to Google’s Vivek Natarajan, several Gemini agents “discuss” these hypotheses with each other, ranking them over the next hours and days, and improving them.

During this process, agents can search the scientific literature, access databases, and use tools such as Google’s AlphaFold system to predict protein structure. “They constantly refine ideas, discuss ideas, criticize ideas,” says Natarajan.

Google has already made the system available to several research groups and has released a short paper explaining its use. The teams who tried it were keen on the possibilities, and these examples suggest that AI co-scientists can help integrate their findings. However, whether the example supports the claim that AI can generate new hypotheses is debatable.

For example, Google says a team used the system to find a “new” method of potentially treating liver fibrosis. However, drugs proposed by AI have been previously studied for this purpose. “It is well established that all identified drugs are anti-fibrotic.” Stephen O’Reilly at the UK biotechnology company Alcyomics. “There’s nothing new here.”

The potential use of this treatment is not new, but team members Gary Peltz at Stanford University School of Medicine in California, two of the three drugs selected by AI co-scientists showed promise in testing for human liver organoids, while the two he selected were both his. There is no growing evidence supporting a choice. Peltz says Google gave him a small amount of money to cover the costs of the test.

In another paper, Jose Penades Imperial College London and his colleagues explain how co-scientists proposed hypotheses that matched unpublished findings. He and his team are studying mobile genetic elements that can move between bacteria – bits of DNA that can move between bacteria – mobile gene elements. Some mobile genetic elements hijack the bacteriophage virus. These viruses consist of a DNA-containing shell and a tail that binds to specific bacteria and injects DNA into IT. Therefore, if the element can enter the shell of a phage virus, you can ride another bacteria for free.

One mobile genetic element creates its own shell. This type was particularly popular and confused Penade and his team. The answer they discovered recently is that these shells can connect with different phage tails, allowing mobile elements to enter a wide range of bacteria.

The discovery was still unpublished, but the team asked AI co-scientists to explain the puzzle. The number one suggestion was to steal a different phage tail.

“We were shocked,” Penades says. “I sent an email to Google. I can access the computer. Is that right? Otherwise, I can’t believe what I’m reading here.”

However, the team released a paper supplied to the system in 2023 – how this family of mobile genetic elements “It steals the tail of a bacteriophage and spreads naturally.” at the time, researchers thought that the elements were limited to obtaining tails from phages that infect the same cell. Only later they discovered that elements can pick up tails floating outside the cell.

So one explanation of how AI co-scientists came up with the correct answer is that they missed the obvious limitation that stopped humans from getting it.

What’s clear is that instead of coming up with a whole new idea, you’re given everything you need to find the answer. “Everything was already public, but it was publicly available on different bits,” Penades says. “The system was able to put it all together.”

The team tried other AI systems already on the market, but none of them came up with an answer, he says. In fact, some people didn’t manage it even when they gave the answer to a paper explaining it. “This system suggests something you’ve never thought of,” says Penades, who hasn’t received funding from Google. “I think it’s going to change the game.”

It becomes clearer over time whether it really changes the game. There’s a mix of Google’s track record when it comes to claiming AI tools to help scientists. Its Alphafold system withstands hype and won the team behind it a Nobel Prize last year.

However, in 2023, the company announced it. Approximately 40 “new materials” It was synthesized with the help of GNOME AI. However, according to the 2024 analysis Robert Palgrave University College London One of the synthesized materials was not actually new.

Despite his discoveries, Palgrave believes that AI can help scientists. “In general, I think AI has a huge amount of contributions to science when implemented in collaboration with experts in their respective fields,” he says.