AI Now Designs Its Own Experiments — But the Proof Is Still at the Bench
- Lille My
- 19 hours ago
- 3 min read
Welcome to this week's scan of how AI is reshaping research — curated from Nature, Science, PNAS and the leading business journals.
AI now designs its own experiments — but the proof is still at the bench
The past two years gave researchers AI that could summarise papers and draft code. This month Nature spotlighted something more ambitious: AI systems that run the full loop of scientific reasoning. In a News & Views article, Olivier Elemento describes two independently developed systems — Google DeepMind's Co-Scientist and FutureHouse's Robin — that generate hypotheses, propose experiments to test them, interpret the results, and then refine their hypotheses in light of the findings. Rather than a chatbot waiting for prompts, the design goal is a collaborator that participates at every step of the discovery cycle. The underlying work is reported in Nature's research pages, including Accelerating scientific discovery with Co-Scientist.
The framing matters as much as the capability. A companion Nature piece, AI tools can speed up thinking, but evidence still comes from the lab bench, insists these agents are meant to *empower* scientists, not replace them: they can compress the slow, cognitive parts of research — surveying a literature, spotting a plausible mechanism, sketching an experimental design — but a hypothesis is only worth as much as the wet-lab or field data that eventually tests it. The bottleneck simply moves. When an AI can propose a hundred promising experiments overnight, the scarce resources become bench time, reagents, human subjects and the judgment to decide which ideas are worth the cost of finding out.
Why it matters: For social scientists, business scholars and marketers, the lesson generalises well beyond biomedicine. Agentic AI is starting to automate the *ideation and design* phase of research — the part many of us treat as the creative core — while leaving validation firmly in human hands. That reshuffles where a scholar adds value: less in generating candidate hypotheses, more in choosing which to pursue, designing credible tests, and guarding against a flood of plausible-but-unverified claims. The labs and departments that win won't be the ones that generate the most hypotheses; they'll be the ones that test the right ones fastest.
More from this week
Can people be trained to spot AI-generated faces?
AI-generated faces are now close to indistinguishable from real photographs, escalating threats to information integrity and security. Algorithmic detectors exist but are opaque and brittle. This PNAS study instead asks whether people themselves can be trained to spot the tell-tale signs of synthetic faces, a human-centred complement to automated deepfake detection that matters wherever trust in images underpins research, journalism and evidence.
Deep learning reads between the lines of company filings
The Management Discussion and Analysis (MD&A) section of a firm's Form 10-K is a rich, narrative window into fiscal health, operations and outlook — but hard to analyse at scale. This MIS Quarterly paper develops a text-based deep learning approach that reads these disclosures to understand company dynamics, illustrating how modern NLP turns unstructured corporate text into signals useful for finance, accounting and IS research.
Neural networks that guide operational decisions
In this Management Science paper, researchers propose deep neural networks for data-driven stochastic optimization: using historical covariates, decisions and costs, a network learns to predict the objective value as a function of a decision, and that predictor is then used to choose better decisions. It is a concrete template for embedding machine learning inside operational decision-making rather than treating prediction and optimization as separate steps.
What happens to ad revenue when the bidders are learning agents
Auction theory predicts revenue equivalence between first- and second-price formats, but that rests on assumptions that fray in real ad markets. This Information Systems Research study examines revenue in first- and second-price display advertising auctions when bidding is done by learning agents, offering platforms and advertisers a more realistic account of how the shift to first-price auctions actually affects what publishers earn.
Is evolvable AI a Darwinian threat? A PNAS debate
In a pointed PNAS exchange, Maarten Boudry argues that AI capable of evolving is still 'domesticated, not feral' and does not yet pose a genuine Darwinian threat. Müller, Steels and Szathmáry reply that the danger of evolvable AI is too grave to be settled by intuition. The back-and-forth is a compact primer on how researchers are reasoning — and disagreeing — about long-run AI risk as a subject of study in its own right.
That's this week. Forward it to a colleague still copy-pasting into a chatbot — and explore AI tools for your own research at gaiforresearch.com.
Comments