AI in Research: Your Agent Believes Whatever You Feed It
- Lille My
- 5 hours ago
- 4 min read
Welcome to this week's scan of how AI is reshaping research — curated from Nature, Science and the leading journals.
The weakest link in agentic research isn't the model — it's what you let it read
The most unsettling result of the week is also the simplest. As reported by Nature, researchers downloaded public datasets on five controversial issues, tweaked them to change the direction and strength of the statistical trends, and uploaded the manipulated versions to private repositories. They then gave AI agents built by Anthropic, OpenAI and Google access to both the public and the doctored repositories, and asked them to answer questions using the data. Around half the time, on average, the agents fell for the manipulated datasets and arrived at the conclusion the 'fraudsters' wanted. (The work is a preprint and has not yet been peer reviewed.)
Read that alongside two other developments and a pattern emerges: the pressure points in AI-assisted scholarship are shifting from *generation* to *verification*. At ICML 2026 in Seoul, a record 23,918 submissions collided with a review process straining under AI use: watermark-based detection flagged 398 reviewers for policy violations and triggered 497 desk rejections — and, as the organisers concede, watermarking catches only the most careless offenders. Meanwhile, in the Academy of Management Review, a new multilevel theory argues that in knowledge-intensive teams, members who become proficient with generative AI increasingly perceive what the model offers as functionally equivalent to what a colleague offers — information, advice, feedback — and route their questions to the machine instead. The collaborative interactions that once were the team's quality-control system quietly thin out.
Why it matters: Each of these is a story about a check that used to be implicit and is now optional. A human analyst who stumbled onto a suspiciously convenient dataset would ask where it came from; an agent with repository access often doesn't. A junior colleague who asks a senior one for feedback gets scrutiny along with the answer; a junior colleague who asks a chatbot gets fluency without friction. None of this argues against using AI in research — it argues for making provenance and verification explicit tasks rather than by-products. Practically: keep a written record of which datasets your agents can reach and where each came from, treat any AI-mediated result whose provenance you cannot reconstruct as unverified, and be deliberate about preserving the human review steps your group used to get for free.
More from this week
Doctored datasets fooled frontier AI agents about half the time
Researchers altered public datasets on five controversial issues to change the direction and strength of statistical trends, uploaded them to private repositories, and gave agents from Anthropic, OpenAI and Google access alongside the genuine public data. On average, around half the time the agents adopted the doctored data and reached the conclusion the manipulators wanted. The authors warn that 'poisoned' information can undermine the reliability of any tool that analyses data autonomously; the study is a preprint and not yet peer reviewed.
The generative AI dilemma inside knowledge-intensive teams
In knowledge-intensive teams, collaboration has historically been the main channel through which members access task-relevant information, advice and feedback. This new Academy of Management Review theory paper argues that generative AI now offers an alternative source of exactly those resources, and that members proficient with it are more likely to perceive what it offers as functionally equivalent to what teammates offer. The resulting multilevel model is a sharp lens for anyone running a lab or research group through the current adoption wave.
ICML's record year exposes the peer-review strain
ICML 2026 opened in Seoul with a record 23,918 submissions, roughly double the previous year's total. Watermark-based detection identified 398 reviewers violating the conference's LLM policies, leading to 497 desk rejections. Organisers acknowledge the method catches only the least careful users, so actual LLM use in reviews is likely considerably higher — a preview of the reviewing-capacity problem heading for every high-volume field.
Complex problems make novel science pay off
This Research Policy paper introduces a measure of problem complexity based on the NK framework — the degree of interdependence among the elements a study must coordinate — and asks how it conditions the payoff to novelty. Novel research addressing more complex problems is substantially more likely to achieve high impact, less likely to receive minimal citations, recognised earlier, and cited more broadly across fields. A useful corrective to the assumption that novelty on its own is the risky bet.
Recommender systems that can explain themselves in words
Platform recommendations improve matching but are notoriously hard for users and managers to interpret. MG-GAT, introduced in Information Systems Research, combines multiple networks with attribute data while tracking which neighbours and features influence each recommendation — then uses that same evidence to generate the explanation. The design point matters for anyone studying explainability: the natural-language account is grounded in model parameters rather than produced post hoc.
Generative AI moves into precision psychotherapy
Writing in Science, Nicholas C. Jacobson makes the case that generative artificial intelligence can scale precision, evidence-based psychotherapy — delivering structured therapeutic protocols at a reach human clinicians cannot match. For social scientists it is a live example of AI crossing from research instrument to intervention, carrying all the measurement, ethics and trial-design questions that transition implies.
That's this week. Forward it to a colleague who has given an AI agent access to a data repository — and explore AI tools for your own research at gaiforresearch.com.