No. 003

Integrity & Trust

• Scientists invented a fake disease. AI told people it was real

  Nature, April 7 2026

  A Swedish researcher planted a fictional eye condition called "bixonimania" in obviously bogus preprints, complete with references to Starfleet Academy. Within weeks, ChatGPT, Gemini, and Copilot were advising users to see an ophthalmologist for it, and a peer-reviewed Springer journal cited the preprint before retracting the paper. The lesson is not that chatbots are gullible; it is that the pipeline from preprint to AI-mediated medical advice has no meaningful filter.

• A retrospective on the ICLR 2026 review process

  ICLR Blog, March 31 2026

  ICLR processed 76,000+ reviews across 13,763 submissions and released its post-mortem. Two LLM detectors flagged suspected AI-generated reviews for area chairs; automated checks caught hallucinated references in submissions, leading to 779 desk rejections for procedural violations. The conference treated detection flags as quality signals rather than automatic penalties, keeping human judgment in the loop.

Research Automation

• PaperOrchestra: a multi-agent framework for automated research paper writing

  arXiv (Google Research), April 8 2026

  Five specialized agents turn lab notes and experimental logs into a submission-ready LaTeX manuscript with literature review, figures, and API-verified citations. In human evaluations it beat autonomous baselines by 50-68% on literature review quality. The authors frame it as assistive, not autonomous: it cannot fabricate results, and its refinement agent is instructed to ignore reviewer requests for data that does not exist in the experimental log.

• Two AI agents for better figures and peer review

  Google Research, April 8 2026

  PaperVizAgent orchestrates five sub-agents to generate publication-ready figures from manuscript text, outperforming human baselines on clarity and aesthetics. ScholarPeer automates peer review by combining a sub-domain historian, a baseline scout that hunts for missing comparisons, and a claim-verification engine. Both are research prototypes, not production tools, but they sketch a future where much of the mechanical work around a paper is delegated.

• DISCO: inventing enzymes for chemistry that nature never explored

  FutureHouse, April 8 2026

  DISCO jointly designs protein sequence and 3D structure around target molecules, producing functional enzymes for reactions with no natural homologs, including B-H insertion at 98% yield. Unlike conventional two-stage design pipelines, it does not require researchers to pre-specify catalytic geometry. This is AI-driven protein engineering moving past remixing known biology.

• UCLA team awarded $5 million DARPA contract to develop AI for math advancement

  UCLA Samueli, April 2026

  DARPA is funding AI systems that can generate and verify mathematical proofs, not just assist with them. The contract signals where defense funding thinks automated research is headed: not writing papers, but doing the reasoning.

What AI Changes

• AI predictions and the expansion of scientific frontiers: evidence from structural biology

  bioRxiv, April 6 2026

  Tracking 245,396 experimental structures in the Protein Data Bank, the authors find that a long-running decline in the study of novel proteins reversed after AlphaFold2's release. The shift is concentrated among papers citing AlphaFold2 and targets with high-confidence predictions. A concrete counter-example to the "AI narrows science" narrative, at least in structural biology.

• Large language models are not the problem

  Nature Astronomy, April 3 2026

  Astrophysicist Hiranya Peiris proposes a blunt test: if an LLM can replicate your scientific contribution, the problem is not the LLM. The argument reframes the authorship anxiety around AI as a question about what counts as a meaningful contribution in the first place. Worth reading alongside the January Nature study showing AI adopters publish 3x more but collectively narrow the field.