Anthropic's 'Claude Science' is a data-science app that hallucinates its citations

An AI workbench that promises the whole research pipeline

Anthropic launched Claude Science, a beta "AI workbench for scientists" that runs as a local server with a browser UI and claims to cover a research project end to end — literature review, data wrangling, analysis, publication-ready figures and manuscripts, with reproducible artifact histories. It ships 60+ curated skills for life sciences and connectors into scientific databases like UniProt, PDB, Ensembl, ChEMBL and GEO, plus NVIDIA BioNeMo models and compute that scales from a laptop to an HPC cluster over SSH.

The headline feature is a standing reviewer agent that runs in the background during a session, checking citations against sources, flagging numbers it can't trace to evidence, and catching figures that don't match the code that generated them, something Claude Code and Cowork don't do automatically. The local-server-plus-proxied-UI architecture is a deliberate design choice, not an accident, and one commenter recognized why.

The people who actually run this stuff are split

The most useful defense came from someone with direct standing. A former Anthropic engineer who built one of the launch's connected tools argued the boring plumbing is the real payoff:

"integrating these tools and databases is hard and time consuming. If the only outcome of this product is that great APIs are built for LLMs, it will be a massive positive impact. Many databases used in computational genomics are still only accessible through FTP!" — the builder of the Biomni HPC connector, ex-Anthropic

The architecture also drew a knowing nod from someone who's built exactly these environments. Locked-down pharma and biobank data sits inside Trusted Research Environments where you can't run a desktop app but can tunnel a JupyterLab UI — which is precisely the shape Claude Science takes:

"Claude Science looks a lot more like something one could imagine spinning up in one of those highly-constrained data environments… than [Claude Code]." — the person who ran the team that built the NIH All of Us TRE

Then the hands-on reports arrived, and they were rougher. A media-informatics lecturer walked through a literature review and watched the tool flail:

"That DOI resolved to the wrong paper. Let me find the correct anchor papers by title/author search directly… It hallucinates references." — a lecturer in Media Informatics at the University of Regensburg

A data scientist pointed out the naming sleight-of-hand: the UIs full of pandas code and plots reveal this is data science, and the "science" is only the data-science slice of life sciences — a McGill/MILA postdoc in computational molecular biology and others noted no ACM/IEEE, no earth science, no physics. And a broader unease ran under the whole thread: that a tool built to "pump out" papers targets a field that already drowns in low-quality output.

"Science isn't suffering from a lack of papers. It's suffering from a lack of good papers." — an engineer who framed the real bottleneck

The closing tension is specific to who'd use it: a computational-biology researcher tried it on RNAi biopesticide design, got a workable-but-naive one-shot answer with real caveats about mammalian design rules — and then Opus 4.8's safety system flagged the session mid-work. A citation-checking reviewer agent is exactly what a hallucinating tool needs; whether it catches its own inventions before they reach a manuscript is the thing no demo can promise.