Google has unveiled an experimental artificial intelligence system that 鈥渦ses advanced reasoning to help scientists synthesize vast amounts of literature, generate novel hypotheses, and suggest detailed research plans鈥�, according to its press release. 鈥淭he idea with [the] 鈥楢I co-scientist鈥� is to give scientists superpowers,鈥� says Alan Karthikesalingam at Google.

The tool, which doesn鈥檛 have an official name yet,听builds on Google鈥檚 Gemini large language models. When a researcher asks a question or specifies a goal 鈥� to find a new drug, say 鈥� the tool comes up with initial ideas within 15 minutes. Several Gemini agents then 鈥渄ebate鈥� these hypotheses with each other, ranking them and improving them over the following hours and days, says Vivek Natarajan at Google.

During this process, the agents can search the scientific literature, access databases and use tools such as Google鈥檚 AlphaFold system for predicting the structure of proteins. 鈥淭hey continuously refine ideas, they debate ideas, they critique ideas,鈥� says Natarajan.

Google has already made the system available to a few research groups, which have released short papers describing their use of it. The teams that tried it are enthusiastic about its potential, and these examples suggest the AI co-scientist will be helpful for synthesising findings. However, it is debatable whether the examples support the claim that the AI can generate novel hypotheses.

For instance, Google says one team used the system to find 鈥渘ew鈥� ways of potentially treating liver fibrosis. However, the drugs proposed by the AI have previously been studied for this purpose. 鈥淭he drugs identified are all well established to be antifibrotic,鈥� says at UK biotech company Alcyomics. 鈥淭here is nothing new here.鈥�

While this potential use of the treatments isn鈥檛 new, team member at Stanford University School of Medicine in California says two out of three drugs selected by the AI co-scientist showed promise in tests on human liver organoids, whereas neither of the two he personally selected did 鈥� despite there being more evidence to support his choices. Peltz says Google gave him a small amount of funding to cover the costs of the tests.

Free newsletter

Sign up to The Daily

Sign up to newsletter

In another paper, at Imperial College London and his colleagues describe how the co-scientist proposed a hypothesis matching an unpublished discovery. He and his team study mobile genetic elements 鈥� bits of DNA that can move between bacteria by various means. Some mobile genetic elements hijack bacteriophage viruses. These viruses consist of a shell containing DNA plus a tail that binds to specific bacteria and injects the DNA into it. So, if an element can get into the shell of a phage virus, it gets a free ride to another bacterium.

One kind of mobile genetic element make its own shells. This type is particularly widespread, which puzzled Penad茅s and his team, because any one kind of phage virus can infect only a narrow range of bacteria. The answer, they recently discovered, is that these shells can hook up with the tails of different phages, allowing the mobile element to get into a wide range of bacteria.

While that finding was still unpublished, the team asked the AI co-scientist to explain the puzzle 鈥� and its number one suggestion was stealing the tails of different phages.

鈥淲e were shocked,鈥� says Penad茅s. 鈥淚 sent an email to Google saying, you have access to my computer. Is that right? Because otherwise I can鈥檛 believe what I鈥檓 reading here.鈥�

However, the team did publish a paper in 2023 鈥� which was fed to the system 鈥� about how this family of mobile genetic elements At the time, the researchers thought the elements were limited to acquiring tails from phages infecting the same cell. Only later did they discover the elements can pick up tails floating around outside cells, too.

So one explanation for how the AI co-scientist came up with the right answer is that it missed the apparent limitation that stopped the humans getting it.

What is clear is that it was fed everything it needed to find the answer, rather than coming up with an entirely new idea. 鈥淓verything was already published, but in different bits,鈥� says Penad茅s. 鈥淭he system was able to put everything together.鈥�

The team tried other AI systems already on the market, none of which came up with the answer, he says. In fact, some didn鈥檛 manage it even when fed the paper describing the answer. 鈥淭he system suggests things that you never thought about,鈥� says Penad茅s, who hasn鈥檛 received any funding from Google. 鈥淚 think it will be game-changing.鈥�

Whether it really is game-changing will become clearer over time. Google鈥檚 track record when it comes to claims about AI tools to help scientists is mixed. Its AlphaFold system is living up to the hype, winning the team behind it a Nobel prize last year.

In 2023, however, the company announced that听 had been synthesised with the help of its GNoME AI. Yet, according to a 2024 analysis by at University College London, .

Despite his findings, Palgrave thinks AI can help scientists. 鈥淚n general, I think AI has a huge amount to contribute to science if it is implemented in collaboration with experts in the respective fields,鈥� he says.

Google has been using artificial intelligence watermarking to automatically identify text generated by the company鈥檚 Gemini chatbot, making it easier to distinguish AI-generated content from human-written posts. That watermark system could help prevent misuse of the AI chatbots for misinformation and disinformation 鈥� not to mention cheating in school and business settings.

Now, the tech company is making an open-source version of its technique available so that other generative AI developers can similarly watermark the output from their own large language models, says at Google DeepMind, the company鈥檚 AI research team, which combines the former Google Brain and DeepMind labs. 鈥淲hile SynthID isn鈥檛 a silver bullet for identifying AI-generated content, it is an important building block for developing more reliable AI identification tools,鈥� he says.

Independent researchers voiced similar optimism. 鈥淲hile no known watermarking method is foolproof, I really think this can help in catching some fraction of AI-generated misinformation, academic cheating and more,鈥� says at The University of Texas at Austin, who previously worked on AI safety at OpenAI. 鈥淚 hope that other large language model companies, including OpenAI and Anthropic, will follow DeepMind鈥檚 lead on this.鈥�

In May of this year, Google DeepMind that it had implemented its SynthID method for watermarking AI-generated text and video from Google鈥檚 Gemini and Veo AI services, respectively. The company has now published a paper in the journal Nature听showing how SynthID generally outperformed similar AI watermarking techniques for text. The comparison involved assessing how readily responses from various watermarked AI models could be detected.

In Google DeepMind鈥檚 AI watermarking approach, as the model generates a sequence of text, a 鈥渢ournament sampling鈥� algorithm subtly nudges it toward selecting certain word 鈥渢okens鈥�, creating a statistical signature that is detectable by associated software. This process randomly pairs up possible word tokens in a tournament-style bracket, with the winner of each pair being determined by which one scores highest according to a watermarking function. The winners move through successive tournament rounds until just one remains 鈥� a 鈥渕ulti-layered approach鈥� that 鈥渋ncreases the complexity of any potential attempts to reverse-engineer or remove the watermark鈥�, says at the University of Maryland.

A 鈥渄etermined adversary鈥� with huge amounts of computational power could still remove such AI watermarks, says at Harvard University. But he described SynthID鈥檚 approach as making sense given the need for scalable watermarking in AI services.

Free newsletter

Sign up to The Weekly

Sign up to newsletter

The Google DeepMind researchers tested two versions of SynthID that represent trade-offs between making the watermark signature more detectable, at the expense of distorting the text typically generated by an AI model. They showed that the non-distortionary version of the AI watermark still worked, without noticeably affecting the quality of 20 million Gemini-generated text responses during a live experiment.

But the researchers also acknowledged that the watermarking works best with longer chatbot responses that can be answered in a variety of ways 鈥� such as generating an essay or email 鈥� and said it has not yet been tested on responses to maths or coding problems.

Both Google DeepMind鈥檚 team and others described the need for additional safeguards against misuse of AI chatbots 鈥� with Huang recommending stronger regulation as well. 鈥淢andating watermarking by law would address both the practicality and user adoption challenges, ensuring a more secure use of large language models,鈥� she says.

Google DeepMind says its artificial intelligence has helped design chips that are already being used in data centres and even smartphones. But some chip design experts are sceptical of the company鈥檚 claims that such AI can plan new chip layouts better than humans can.

The newly named AlphaChip method can design 鈥渟uperhuman chip layouts鈥� in hours, rather than relying on weeks or months of human effort, said and , researchers at Google DeepMind, in a . This AI approach uses reinforcement learning to figure out the relationships among chip components and gets rewarded based on the final layout quality. But independent researchers say the company has not yet proven such AI can outperform expert human chip designers or commercial software tools 鈥� and they want to see AlphaChip鈥檚 performance on public benchmarks involving current, state-of-the-art circuit designs.

鈥淚f Google would provide experimental results for these designs, we could have fair comparisons, and I expect that everyone would accept the results,鈥� says at Binghamton University in New York. 鈥淭he experiments would take at most a day or two to run, and Google has near-infinite resources 鈥� that these results have not been offered speaks volumes to me.鈥�

Google DeepMind鈥檚 blog post accompanies an to Google鈥檚 2021 Nature journal paper about the company鈥檚 AI process. Since that time, Google DeepMind says that AlphaChip has helped design three generations of Google鈥檚 Tensor Processing Units (TPU) 鈥� specialised chips used to train and run generative AI models for services such as Google鈥檚 Gemini chatbot.

The company also claims that the AI-assisted chip designs perform better than those designed by human experts and have been improving steadily. The AI achieves this by reducing the total length of wires required to connect chip components 鈥� a factor that can lower chip power consumption and potentially improve processing speed. And Google DeepMind says that AlphaChip has created layouts for general-purpose chips used in Google鈥檚 data centres, along with helping the company MediaTek develop a chip used in Samsung mobile phones.

鈥淲e really don鈥檛 know what AlphaChip is today, what it does and what it doesn鈥檛 do,鈥� says , a chip design researcher at a competing firm. 鈥淲e do know that reinforcement learning takes two to three orders of magnitude greater compute resources than methods used in commercial tools and is usually behind [in terms of] results.鈥�

Free newsletter

Sign up to The Weekly

Sign up to newsletter

Markov and Madden critiqued the听original paper鈥檚 claims听about AlphaChip outperforming unnamed human experts. 鈥淐omparisons to unnamed human designers are subjective, not reproducible, and very easy to game. The human designers may be applying low effort or be poorly qualified 鈥� there is no scientific result here,鈥� says Markov. 鈥淚magine if AlphaGo reported wins over unnamed Go players.鈥� A Google DeepMind spokesperson described the experts as members of Google鈥檚 TPU chip design team using the best available commercial tools.

In 2023, an independent expert who had reviewed Google鈥檚 paper his Nature commentary article that had originally praised Google鈥檚 work but had also urged replication. That expert, at the University of California, San Diego, also ran a that tried to replicate Google鈥檚 AI method and found it did not consistently outperform a human expert or conventional computer algorithms. The best-performing methods used for comparison were commercial software or internal research tools for chip design from companies such as Cadence and NVIDIA. In a , Goldie and Mirhoseini disputed Kahng鈥檚 benchmarking results. They said his tests had not pretrained the AI method on specific chip designs 鈥� a crucial factor in its performance 鈥� and relied upon 鈥渇ar fewer compute resources鈥� than Google DeepMind鈥檚 team to train the AI.听

鈥淥n every benchmark where there鈥檚 what I would consider a fair comparison, it seems like reinforcement learning lags behind the state of the art by a wide margin,鈥� says Madden. 鈥淔or circuit placement, I don鈥檛 believe that it鈥檚 a promising research direction.鈥�