ما الذي يشبهه العمل مع فريق من العلماء الافتراضيين المدعومين بالذكاء الاصطناعي؟

On a quiet Sunday morning in April, pathologist Thomas Montine found himself leading one of the most unusual meetings of his career. Using an online test platform called the Virtual Lab, he assembled…

On a quiet Sunday morning in April, pathologist Thomas Montine found himself leading one of the most unusual meetings of his career. Using an online test platform called the Virtual Lab, he assembled a team of six AI-powered virtual scientists, each given a specific specialty — from neuroscientists to a neuropharmacologist and a medicinal chemist. Montine posed to this digital team the kinds of challenging questions he typically tackles in grant proposals: what potential treatments could work for Alzheimer’s, what knowledge gaps exist, and what barriers stand in the way of progress.

In minutes, he received a sprawling transcript exceeding 10,000 words, starting with a virtual principal investigator greeting the team: “Thank you all for joining this important meeting”.

Montine, who researches cognitive impairment at Stanford University, was experimenting with a growing trend: employing teams of specialized AI chatbots to brainstorm research questions as a lab team would. Proponents of these “co-scientist” systems argue that this approach can speed up hypothesis development and more controversially, generate valuable new research ideas.

AI Teams Enter the Lab

Among the leaders in this field is a Google team that shared early tests of its AI co-scientist project in February, opening the system to selected testers for further refinement. But Google is hardly alone. At Stanford, computational biologists recently unveiled their Virtual Lab platform — the same one Montine was exploring — while researchers at the Shanghai Artificial Intelligence Laboratory are building a similar system named VirSci.

Rick Stevens, a computer scientist at the University of Chicago and Argonne National Laboratory, notes that computational researchers everywhere are setting up networks of AI personas to simulate collaborative discussions. “Really, anyone can do it,” he says.

These systems don’t just simulate conversations. In many cases, the AI agents can browse the web, write code, or interact with external tools, making them examples of “agentic AI” — a term for language models that autonomously execute tasks, though humans typically provide oversight. Such multi-agent systems can focus on complex research problems for extended periods without the distractions or fatigue that humans face. As Stevens puts it, “It’s like having extra colleagues who never get tired and know everything they’ve ever read”.

Experimenting with Virtual Collaborations

To see how these systems perform in practice, Nature invited several scientists to try Stanford’s Virtual Lab and interviewed researchers who tested Google’s co-scientist. Would a group of chatbots behave more like a room full of Nobel laureates or a classroom of undergraduates? Could they generate ideas that were insightful and valuable, or would the outputs be trivial or off-track?

How Virtual Labs Assign Roles

All of these co-scientist systems create roles for their agents, but their setups differ. Stanford’s Virtual Lab, designed by Kyle Swanson in James Zou’s group, comes with two default characters powered by GPT-4o: a principal investigator to lead discussions and a critic to provide constructive feedback. Users can add as many agents as they like, assign them specialties, and control the number of discussion rounds. The system quickly generates a transcript simulating a lab meeting.

Meanwhile, Google’s co-scientist, developed by DeepMind’s Alan Karthikesalingam and Vivek Natarajan, uses a set of six predefined agents with roles such as idea generation, critique, idea refinement, duplication checking, ranking, and meta-review. Rather than assigning specialties, users provide a research goal and optional background information. The agents, powered by Google’s Gemini 2.0 model, then collaborate to deliver a detailed summary report.

These systems can produce lengthy documents — sometimes spanning hundreds of pages — covering hypotheses, literature searches, and research suggestions. As Natarajan puts it, “The co-scientist is like a highly capable scientific partner, able to spot both obvious and subtle connections across research”.

Strengths and Challenges of AI Co-Scientists

While multi-agent discussions can filter out nonsensical or incorrect statements — hallucinations being a known flaw of language models — Stevens suggests that occasional AI “hallucinations” can actually spark creative ideas if reviewed carefully by experts.

Evidence indicates that multiple agents improve the quality of outputs compared with a single chatbot. For example, Zou’s team found that adding a critic agent raised GPT-4o’s performance on graduate-level science tests, improving its answers in tasks like designing radiotherapy treatment plans.

Google compared the multi-agent system to solo chatbots by having experts evaluate outputs for novelty and impact. The multi-agent responses scored slightly higher in both categories.

Research teams are also probing how many agents and discussion rounds produce the best results. The team behind VirSci found eight agents with five conversation turns maximized creativity, while Swanson observed that adding more than three agents or rounds could lead to repetitive or unfocused text.

Firsthand Experiences from Scientists

Stanford’s Gary Peltz, an early user of Google’s system, asked the AI team for ideas on drugs to treat liver fibrosis. The AI’s report highlighted epigenetic changes echoing a grant Peltz had recently submitted and suggested three drugs, two of which showed promise when his lab tested them. Reflecting on the experience, Peltz called the technology transformative: “These LLMs are what fire was for early human societies”.

Yet not everyone agrees on the AI’s originality. Other liver researchers argued the drug suggestions were common knowledge. Peltz acknowledged that the AI prioritized options he might have overlooked, offering fresh perspectives akin to discussions with junior researchers.

Conversations Lacking Human Nuance

Testers using Stanford’s Virtual Lab noted the system’s conversations can feel rigid. Agents speak in numbered lists, are unfailingly polite, and avoid arguments, missing the leaps of intuition or spontaneous sparks typical of real human discussions. Francisco Barriga, a cancer-genome researcher, said the AI team suggested exactly what he would have done when designing mouse experiments but found it lacked the creative back-and-forth of conversations with colleagues. Still, he saw its potential as a resource for students needing quick guidance.

Catherine Brownstein, a geneticist working on rare diseases, uses LLMs for efficiency and broader thinking but warns that expertise is needed to catch AI mistakes. In one case, she was surprised and humbled when the AI suggested involving patients in shaping research priorities, a reminder of her original passion for patient-centered research.

AI as a Catalyst, Not a Replacement

For MIT tissue engineer Ritu Raman, Google’s AI co-scientist helped her think outside her specialty, proposing experimental approaches she might not have considered on her own. The AI’s suggestions gave her confidence to pursue ideas beyond her immediate expertise, underscoring Raman’s observation: “The interpreter is just as important as the algorithm.”

Discussions over the novelty of AI-generated ideas continue. In one example, microbiologist José Penadés and his team found that Google’s AI proposed a hypothesis matching their unpublished conclusions, connecting research dots in a way Penadés considered successful.

Looking Ahead

These multi-agent systems are still emerging, with most tools not yet easily accessible for everyday lab use. But researchers expect rapid developments, including integration with robotic labs for automating experiments. FutureHouse, a startup, recently introduced an autonomous AI discovery system combining literature review with experiment design.

Although some testers felt the AI conversations hinted at deeper reasoning, they agreed that LLMs should remain assistants rather than replacements for scientists. The adoption of AI tools into research seems inevitable, but opinions differ on whether their impact will be revolutionary or simply incremental.

As Peltz notes, the limiting factor may soon be funding rather than ideas: “Science will have no shortage of good hypotheses generated by AI. The real challenge will be finding the resources to test them all”.