AI in Research: The Agents Pick Up the Pipette
Image: "Andrew+ robot" by Pocar19 (CC BY-SA 4.0) via Wikimedia Commons, recoloured.
Welcome to this week's scan of how AI is reshaping research, curated from Nature, Science, PNAS and the leading business journals.
From proposing experiments to running them
Most of AI's contribution to science so far has been computational: analysing data, making predictions, searching literature. As James Zou puts it in a Nature News & Views, an AI system might design a promising molecule in minutes, but testing it in a 'wet lab' can take weeks or months of human effort. This week's Nature paper by Wu and colleagues tackles that gap directly. Their system, AutoDNA, is an 'AI-native' biomolecular laboratory that autonomously orchestrates more than 20 types of instruments across more than 11,000 steps to carry out core nucleic-acid operations: synthesis, transcription, amplification and sequencing. Human intervention is limited to hardware maintenance, such as replenishing reagents and labware.
The results across four applications are concrete. AutoDNA matched expert-optimised yields in enzymatic DNA synthesis, ran concurrent protocols 3.2 times faster, completed a full DNA data-storage cycle with exact recovery of the original digital content, and identified and validated a clinically detected KRAS G13D mutation. The authors argue that earlier autonomous labs stayed confined to narrow, single-objective workflows because they used AI as an 'add-on' tool rather than giving it a global understanding of the experiment; AutoDNA instead co-designs models, experiments and instruments. A separate preprint covered by Nature (not yet peer reviewed) pushes the idea further: agents build a virtual replica of a lab, test protocols there, then direct real robots. In one test they wrote more than 200 commands for a robotic arm to measure the optical properties of nanocrystals. An MIT bioengineer quoted in the piece calls the concept "very cool" but notes that the cameras and sensors the agents need could cost more than setting up a lab manually.
Why it matters: The AutoDNA authors frame their work as a blueprint for 'science-as-a-service'. Even if you never touch a pipette, the lesson travels: when executing a study becomes cheap and automated, the scarce skills shift to choosing the right question and checking what the machine did. The paper's diagnosis also applies well beyond biology. Bolting a chatbot onto an existing workflow is an 'add-on'; the bigger gains came from redesigning the workflow around the AI. That is worth keeping in mind before adding AI to your own survey, coding or analysis pipeline.
More from this week
Measure learning with AI, and without it
Responding to the argument that universities should teach students to spar with AI models and assess how well they do, Christoph Trattner points out a missing measurement: how students perform once the AI is gone. He cites a randomized trial with secondary-school maths students (Bastani et al., PNAS 2025) in which unrestricted chatbot access raised practice scores but lowered performance on a later exam taken without it. An evaluation that stopped at the assisted session would have called that tutor a success, a useful warning for anyone designing AI-in-teaching studies.
Scaling ad experiments with prediction
Brett Gordon, Robert Moakler and Florian Zettelmeyer propose Predicted Incrementality by Experimentation (PIE), which treats ad measurement as a prediction problem: learn the mapping from campaign features to causal effects on a sample of randomized experiments, then apply it to campaigns that were not run as RCTs. Because the RCTs pin down the causal effects, PIE can use post-campaign metrics that would be invalid controls in a causal model. Across 2,226 Meta ad experiments, PIE disagreed with RCT-based decisions in only 8–12% of campaigns, versus 12–20% for standard seven-day last-click attribution.
What drives exploration in children, adults and AI agents
Lidayan and colleagues compare adults, children aged 8 to 12 and AI agents exploring the open-ended game environment Crafter, testing three information-theoretic intrinsic objectives: entropy, information gain and empowerment. All three correlate positively with human exploration progress, with entropy and empowerment more strongly related in children. Across humans and agents, entropy rises fast and then plateaus while empowerment keeps climbing, suggesting a shift from seeking novel states to seeking controllable ones. There is also preliminary evidence that children's private speech about goals may aid exploration.
The hidden limits of t-SNE
t-SNE is a go-to method for visualising high-dimensional data, including the text and image embeddings many researchers now generate with AI models. Rupert Li and Elchanan Mossel provide a mathematical framework for the information it discards, with quantitative results across several scenarios showing how important structural information is lost. If your paper's argument rests on clusters in a t-SNE plot, this is a reason to back them with a second method.
Takeda rebuilds its R&D around AI
George Westerman and David Kiron document how Takeda Pharmaceutical embedded AI into its R&D workflows, moving scientists from broad laboratory experimentation toward data-driven prediction and targeted validation. The case identifies executive accountability, strong data foundations, workforce transformation and a culture of trust and experimentation as critical to scaling AI. It is a useful organisational companion to this week's lab-automation papers and a teachable case for management courses.
That's this week. Forward it to a colleague who's still copy-pasting into ChatGPT, and explore AI tools for your own research at gaiforresearch.com.
Comments