No. 015

Three moves this week to put AI inside the scientific workflow, at three different scales. Anthropic shipped a full research workbench with compute management and a reviewer agent. Google piloted an agentic review framework at STOC and ICML. A bioinformatician built his own pre-submission review skill with Claude Code and Consensus, open-source, for the cost of a prompt file. Meanwhile, NeurIPS embedded hidden prompts in manuscripts to catch reviewers using the same class of tools without permission. The field is simultaneously building AI into the research process and policing it out of the review process, and the line between those two activities is not as clear as anyone pretends.

Research Workbenches

• Claude Science, an AI workbench for scientists, is now available

  Anthropic, June 30 2026

  Anthropic's step from "model with life-sciences connectors" to a full workbench product: 60+ preconfigured skills for genomics, proteomics, structural biology, and cheminformatics, with compute management across local machines, HPC clusters, and cloud GPUs via Modal, plus a reviewer agent that checks citations and calculations, a discounted plan for academic labs, and up to $30K in credits for 50 selected AI-for-science projects.

• Test Driving Claude Science

  Paired Ends, June 30 2026

  The first independent field test: Turner ran a literature review and an IUCN Red List analysis, burned through the 5-hour limit in two prompts plus $50 in overages, and calls it "incredibly cheap relative to the time it would have taken" while noting the system made analytical choices he would not have made without reviewing the generated code.

Automating Review

• Towards Automating Scientific Review with Google's Paper Assistant Tool

  Jayaram et al. (Google), arXiv, June 26 2026

  An agentic framework that examines manuscripts for theoretical soundness and experimental validity, reporting a 34% improvement over zero-shot recall on mathematical errors via inference scaling, with pilot deployments at STOC and ICML that frame AI as augmenting rather than replacing human reviewers.

• A Claude skill for pre-submission peer review

  Paired Ends, June 30 2026

  A DIY Claude Code skill that runs structured pre-submission review by searching Consensus for published literature and producing a Word-doc report covering background accuracy, missing citations, methods, and recommendations, with every search query logged as an audit trail against hallucinated references.

• Now is the Time for AI in Peer Review, and Publishing Policies Need to Recognize This

  The Scholarly Kitchen, June 30 2026

  Frontiers' Director of Research Integrity argues that blanket restrictions on reviewer use of AI are inconsistent with the AI already embedded in publishing workflows for plagiarism detection, image manipulation, and paper-mill identification, and calls for transparent frameworks that establish where AI can support review rather than policies that pretend it is not happening.

Review Integrity

• Scientists decry conference's use of hidden prompts to snare AI peer reviews

  The Transmitter, July 1 2026

  NeurIPS embedded concealed LLM instructions in manuscripts sent to reviewers, designed to trigger telltale phrases if a reviewer feeds the paper to a chatbot, following ICML's earlier move that caught 506 reviewers violating AI-use policies and desk-rejected 497 papers, roughly 2% of submissions.