AI in Research: The Machine That Predicts Your Experiment — Before You Run It
- Lille My
- 11 minutes ago
- 3 min read
Welcome to this week's scan of how AI is reshaping research — curated from Nature, Science and the leading journals.
AI is becoming a research instrument — one that both predicts and performs experiments
For social scientists, the headline result this week is unsettling in the best way. Writing in Nature, Ashwini Ashokkumar, Luke Hewitt, Isaias Ghezae and Robb Willer assembled 70 pre-registered, nationally representative survey experiments — 476 treatment effects across 105,165 U.S. participants — and asked GPT-4 to simulate how representative samples of Americans would respond to each stimulus. The model's predicted treatment effects correlated with the real ones at r = 0.85, matching or beating panels of human forecasters. Crucially, accuracy held (r = 0.90) for unpublished studies that could not have leaked into training data, though the model tended to overestimate effect sizes. A companion Nature commentary argues for judicious use of LLMs to triage and pre-test designs before fielding them.
If that paper shows AI predicting experiments, a second one shows AI running them. In Science, Kexin Huang and colleagues introduce Biomni, a general-purpose biomedical agent whose action-discovery component mined tools, databases and protocols from thousands of publications across 25 domains to build a unified environment in which it autonomously executes diverse research tasks. Read together, the two papers sketch a continuum: a model that can cheaply forecast which hypotheses are worth testing, feeding an agent that can carry parts of the testing out — compressing the loop between idea and evidence.
Why it matters: An instrument that predicts your results before you collect data is genuinely useful for prioritising studies, powering analyses and stress-testing designs — but it is also a temptation to skip the messy, surprising empirical work that overturns priors. The r = 0.85 headline is an average over well-trodden survey paradigms; the questions your field most needs answered are often the ones an LLM has the least to say about. Use these tools to decide what to run, not as a substitute for running it — and treat any 'simulated participants' as a hypothesis generator, never as data.
More from this week
'Humanizer' tools can erase the signs of AI-written text
Nature reports on a class of 'humanizer' tools that paraphrase and restyle AI-generated text specifically to defeat the detectors journals and universities rely on. For researchers and editors, the arms race matters: it undercuts detection-based integrity checks and pushes the burden back onto disclosure norms, provenance and trust. The piece is a useful reality check for anyone counting on automated AI-text detection to police submissions.
Are we becoming 'reverse centaurs'?
In a Science review of Cory Doctorow's 'The Reverse Centaur's Guide to Life After AI,' Tim Wu unpacks the book's central worry: a 'reverse centaur' is a worker reduced to feeding and cleaning up after an automated system rather than being augmented by it. For academics weighing how to fold AI into research and teaching workflows, it is a sharp reminder that the design of the human-AI division of labour, not the model's raw capability, determines who is really in charge.
A foundation model for cells
Researchers present the Universal Cell Embedding (UCE), a foundation model trained by self-supervised learning on a large corpus of single-cell data to place cells from many types and species into a shared representation space. Like language-model embeddings, it aims to be a general-purpose substrate other analyses can build on — offering social and behavioural scientists a concrete template for what 'foundation models' look like outside of text.
A neural network built inside a memory chip
Science reports on an artificial neural network implemented directly within a computer memory chip — 'computing in a memory with physics' — that reconstructs patterns of human cortical activity with high accuracy in real time. By doing the computation where the data already sits, such neuromorphic hardware points toward far more energy-efficient AI, a growing concern as the compute and electricity costs of large models climb.
Humanoid robots take a first step into surgery
A Nature feasibility study reports early in vivo tests of humanoid robots performing surgical tasks — an incremental but concrete probe of whether general-purpose robotic platforms, rather than bespoke surgical rigs, can operate in a living body. It is a reminder that the same embodied-AI advances driving warehouse and lab automation are edging into high-stakes clinical settings, where the bar for reliability is far higher.
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