Will AI Replace Scientists? Human–AI Collaboration Wins Only When the Human Adds Something

Studies show that people working with AI often do worse than AI alone. In open-ended science, the scientist still adds what AI lacks: the question, the context, and the check.

Haohan Wang · October 24, 2026 · AI for science · human-AI collaboration · AI scientist · agentic AI · future of research · personal perspective

Will AI replace scientists? My answer is no, but not for the comforting reason. It is not true that a human working with AI always beats AI alone. Across many experiments, human–AI teams are often worse than the better of the two. The teams win when the human brings something the AI does not have. In science, that something is large: choosing the question, knowing the context of the data, and checking whether a result is true. So I expect AI to replace much of the work scientists do today. I do not expect it to replace the scientist who does those three things well.

In short:

A large meta-analysis found that human–AI teams did worse, on average, than the best of humans or AI alone. They gained only when the humans were better than the AI to begin with.
AI systems can now run a full research project, but in open-ended biology their answers are still often wrong, and someone has to catch the errors.
The strongest results in AI-driven biology so far came from systems aimed at a scientist’s own question and checked by that scientist’s experiments.

Are humans working with AI better than AI alone?

Often they are not. In 2024, Vaccaro, Almaatouq, and Malone pooled 106 experiments with 370 effect sizes. On average, human–AI combinations performed worse than the better of humans or AI alone (Hedges’ g = −0.23). They did beat humans working alone (g = 0.64). So AI helped people, but people often dragged the AI down.

The pattern depended on who was stronger at the start. When humans alone beat the AI, the combination improved on both (g = 0.46). When the AI alone beat humans, the combination lost ground (g = −0.54). The losses were concentrated in decision tasks, where the job is to pick the right answer.

Medicine shows the same thing. In a randomized trial with 50 physicians, doctors given GPT-4 scored 76% on diagnostic reasoning, against 74% for doctors without it. That difference was not significant. GPT-4 alone scored 92%. A study of radiologists found a similar result. AI predictions did not improve their accuracy on average, because the radiologists underweighted the AI and treated its advice as unrelated to what they already saw. Even skilled software developers were 19% slower with early-2025 AI tools, though they expected to be 24% faster.

These results argue against a lazy version of my claim. Adding a human to an AI is not a safety net by itself. If the human only second-guesses an answer the AI already gets right, the human makes things worse.

Can AI do science on its own?

In some settings, it is getting close. In March 2026, Nature published The AI Scientist. It generates ideas, writes and runs code, analyzes results, writes the paper, and reviews its own work. One of its manuscripts passed the first round of peer review at a machine learning workshop. This is the strongest case against my position, and I take it seriously.

Two details matter, though. The workshop accepted about 70% of submissions, a lenient bar that the authors themselves point out. And according to the UBC announcement, the system still produces underdeveloped ideas and inaccurate citations, and so far works only in computer science. Machine learning is an unusually friendly field for automation. The data is digital, experiments are cheap, and a result can be scored by a number.

Biology is less friendly. On BixBench, a benchmark of about 300 open-answer questions drawn from real bioinformatics analyses, frontier models in early 2025 answered only 17% correctly. Newer systems are much better. Independent scientists judged 79.4% of the statements in Kosmos reports to be accurate, and collaborators said one run did the work of about six months of their own research. That is impressive. It also means about one statement in five was not judged accurate, and in biology a wrong statement can look exactly like a right one.

Where does the scientist still add value?

The meta-analysis gives the rule: a human helps when the human knows something the AI does not. In open-ended science, I see three such things.

The first is the question. The most striking AI results in biology began with a scientist’s question. Google’s AI co-scientist is conditioned on a researcher’s objective, and its drug-repurposing candidates for leukemia were then tested in the lab. In a study later published in Cell, a microbiology lab asked it a question that had taken them years to answer and was still unpublished. The system’s top-ranked hypothesis matched the mechanism the lab had confirmed experimentally. The lab then assessed its other top hypotheses and found that some opened new directions. The AI generated; the scientists chose the problem and judged the answers.

The second is context. The radiology study found that information humans hold and the AI lacks does improve decisions. In biology, much of this context lives outside the dataset: how a sample was collected, which batch effect is known in a lab, which finding has failed to replicate.

The third is the check. Our own work on agents for gene expression analysis started from this need. We built GenoTEX, a benchmark based on analyses curated by expert bioinformaticians, because we could not judge an agent’s analysis without a human reference. Our multi-agent system, GenoMAS, reaches an F1 of 60.48% at identifying genes linked to a trait. That is a large step up from earlier methods, and it still leaves many genes for a person to confirm or reject. More on this line of work is on my agentic AI for genomics page.

Is there a cost to working with AI?

Yes, and it is a cost to science as a whole rather than to each scientist. A 2026 Nature study of 41.3 million papers found that scientists who used AI published 3.02 times more papers and received 4.84 times more citations. But AI adoption shrank the range of topics studied by 4.63% and reduced scientists’ engagement with one another by 22%. AI pulls work toward fields that already have rich data.

This is a reason to value the scientist’s role more, not less. Choosing a question that the data does not yet favor is exactly the part AI does not do for us.

What should scientists do now?

My position is that collaboration is a skill, not a default. The evidence says a human adds value only when the human contributes something specific. So I would tell my students to stop competing with AI on the parts it already does well, such as writing routine code or summarizing papers. Instead, they should get very good at the parts it does poorly: asking a question worth answering, knowing the data’s history, and designing the check that would show a result is wrong.

The evidence here has limits. Most collaboration studies use short tasks, not month-long projects, and AI systems improve every few months. Some of what I call human advantage today may shrink. But the scientist’s job has never been only to produce answers. It has been to decide which answers matter and to stand behind them. That part, I think, stays with us.

Sources and further reading

This post shares my personal perspective; it does not represent the views of the University of Illinois.