top of page

AI in Research: When the Machine Reviews the Work

2 days ago
4 min read

Image: "Angibous-Esnault en correction de manuscrit" by Jogd75 (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 and the leading journals.

AI is learning to inspect the work: papers and other AIs

The most provocative paper of the fortnight comes from Harvard's Institute for Quantitative Social Science. In AI Article Reviews Dominate Human Peer Reviews at Their Best (working paper, 28 Oct), Gary King and Steven Worthington took 134 human reviews of 124 papers submitted to rapidPeer, a service that gives faculty high-quality double-blind reviews of drafts before journal submission, and generated a matching review of each paper with Anthropic's Claude Opus. The authors report that the AI reviews came out ahead *by every measure they could construct*: an AI judge preferred the AI review in every pair, and three former editors of major journals, judging blind, chose it in about 87% of their judgments, as Times Higher Education reports. The AI reviews contained no fabricated sources and about a third the rate of incorrect claims of the human ones.

The same idea of AI as inspector is spreading beyond manuscripts. Nature reports (13 Nov) on a preprint describing an AI 'mechanic' that can look inside other neural networks and modify them, which could make AI systems easier to target and debug. Researchers quoted in the piece are cautious: automating parts of interpretability is not the same as explaining how a network works, and whether understanding models really makes AI safer is still largely untested. King and Worthington are just as explicit about their limits. Their paper has not itself been peer reviewed, they did not measure whether the AI advice actually improved the papers, and they did not study what happens when a whole field switches to AI reviewing. That includes the risk that models nudge everyone towards more conventional ideas.

Why it matters: The authors' most realistic policy option is that some journals skip routine human review. An editor would instead weigh a dossier: the paper, an AI review from the journal's official prompt, and the authors' response. Whether or not your journals go that far, the practical step is already open to you. Run a strong model over your own draft before submission, and treat its review as you would a sharp colleague's: useful, fallible, and no substitute for your judgement about what is new.

More from this week

Your LLM coding choice can flip the sign of your regression

Joan Timoneda and Sebastián Vallejo Vera tackle the reproducibility problem of research that uses LLMs to code text: outputs are probabilistic and vary across models and versions. Using three datasets, they show that measurement choice can reverse the sign of downstream regression coefficients. Design-based corrections using a gold-labelled subsample restore validity but cost precision, and their main contribution is a stacked ensemble of open-weight models that improves precision while keeping estimates valid. It comes with a standardised prompt script and open data and code, and is a chapter of a forthcoming Oxford University Press book.

The four jobs LLMs now do in social science

Fengli Xu, James Evans and colleagues review how LLMs now serve four distinct roles across sociology, political science, psychology, anthropology and economics: subjects that simulate human behaviour, assistants that automate analysis, collaborators that generate hypotheses, and evaluators that audit research quality. They also stress the limits: LLMs struggle with causal inference, raise privacy and misinformation concerns, and their ability to reproduce social complexity is uneven and unproven. A useful map if you are deciding where LLMs belong in your own pipeline.

A 'machine psychology' for language models

Thilo Hagendorff, Eric Schulz and co-authors argue that experiments inspired by psychology are a productive way to understand LLM behaviour and reasoning, going beyond performance benchmarks to computational insight into emergent abilities and behavioural patterns. The article reviews existing work, synthesises best practices, and spells out the caveats of applying methods designed for humans to machines. It is directly relevant to behavioural researchers who now treat models as subjects of study.

AlphaGenome 2 predicts where DNA variants act

Google DeepMind's AlphaGenome 2 is a conditional model that predicts how a DNA sequence and its variants affect a specific cell type, tissue or state. It was trained on more than 100 million individual cell profiles across 68 tissues. On an immune-cell benchmark covering 28 cell types, DeepMind reports that it identified causal variants in 30% of cases versus 16% for earlier approaches. It is available through an API for non-commercial research and is intended for research use only, not clinical diagnosis. Nature's news team reports that scientists are impressed.

Classifiers that survive survey drift

Nathan Justin, Phebe Vayanos and colleagues study how to learn classification trees that stay robust when the data seen at deployment differ from the training data. The problem is common in high-stakes settings such as public health and social work, where self-reported survey data are sensitive to question framing, timing, place and how comfortable respondents feel. Interpretable models that hold up under such shifts matter for anyone turning survey data into decisions.

That's this week. Forward it to a colleague who is about to submit a paper, and explore AI tools for your own research at gaiforresearch.com.

Comments


This initiative is supported by the following organizations:

  • Twitter
  • LinkedIn
  • YouTube
logo_edited.png
bottom of page