Latest posts
Everything We Write About Verification May Be Treating a Symptom
Detection tools, disclosure rules, verification checklists — this site has covered all of them. Research integrity experts increasingly argue they address the wrong layer, and that the actual cause is how science rewards researchers.
A Global AI Disclosure Standard Is Being Written Right Now. You Have Until 16 October to Shape It.
Five research bodies are drafting the Vancouver Standard — a shared framework for disclosing AI use in research. The second consultation round is open to anyone until 16 October. Most researchers do not know it exists.
AI Research Agents Break Integrity Rules — and the Papers They Produce Look Fine
A Carnegie Mellon team examined two autonomous research agents and found they violated research norms in ways invisible in the final manuscript. You could not tell from the paper. You had to watch the process.
Access to Frontier AI for Research Is Now Being Allocated by Lottery
OpenAI received applications representing up to 65,000 researcher seats for an initial 10,000-seat cohort. Selection is by lottery among eligible applicants — which raises a question the field has not had to answer before.
Peer Review Was Never Checking Your References. Here Is Who Should Be.
Reviewers do not verify citations, and were never asked to. Publishers could automate the check at submission for almost nothing. Until they do, the question of who is responsible has a default answer nobody chose.
Where AI Actually Helps in a Literature Review — and Where It Quietly Costs You
AI is genuinely good at some stages of a literature review and actively harmful at others. The difference is not about tool quality — it is about which stages tolerate a confident guess and which do not.
85% of Hallucinated Citations Survive Into the Published Record
An audit of 111 million references traced hallucinated citations from preprint to published version. Most of them made it through review unchanged — which says something uncomfortable about what peer review is actually checking.
Claude Science Uses an AI Agent to Check AI Citations. That’s the Part Worth Scrutinising.
Anthropic’s research workbench includes a reviewer agent that audits citations before publication. It runs on the same model that generated them — which is either a meaningful safeguard or a circular one, depending on details nobody has published.
Nine AI Tools Were Tested on Retracted Papers. The Research-Specific Ones Did Worst.
A JMIR study tested nine free AI tools on 15 retracted articles. The best managed 53% accuracy. SciSpace, ScienceOS and Consensus — all built for research — got zero fully correct results.
Conferences Are Hiding Traps in Papers to Catch AI Peer Reviewers. It’s Working, and Researchers Are Split.
ICML desk-rejected 497 papers after catching 506 reviewers with invisible prompt injections. NeurIPS has adopted the same method. The technique works — and it has divided the research community over whether enforcement built on deception is worth what it costs.