Scientific usefulness starts with a question that can fail

On July 20, Anthropic opened a call in its AI for Science program focused on rare genetic diseases. Accepted proposals can receive up to $50,000 in Claude credits over six months. There are two tracks: one for basic research involving scientists, patient organizations, and data specialists; another for early-stage biotechs working on clinical development. This is a small budget compared with the cost of developing a therapy, but it is interesting for where it places AI: in research where data are scarce, scattered, and hard to compare.

What is verifiable

In its first track, the call proposes tasks such as connecting mechanisms across diseases, ranking hypotheses, and evaluating how models perform on phenotype, variant, and mechanism-prediction problems. Anthropic says that outputs from that track will be made public through the Monarch Initiative. The second track includes support for analyzing therapeutic strategies, natural-history data, and regulatory documentation. Applications are open until August 2. Anthropic’s announcement sets out the limits, examples, and conditions for both tracks.

None of that demonstrates that a model has found a therapy or can replace experimental, clinical, or regulatory validation. The announcement itself acknowledges two material constraints: an agent cannot solve missing or poorly organized data, nor the physical bottlenecks of manufacturing, safety testing, and diagnostic infrastructure.

AI interpretation

The most promising part of this call is not the credits, but the chance to make the intermediate work visible. In rare diseases, a seemingly convincing answer can arise from incomplete literature, incompatible labels, or an association that merely looks causal. “Acceleration” should therefore not be measured by the number of documents a system drafts, but by whether it helps formulate a hypothesis that an expert can refute, reproduce, or prioritize more effectively.

A useful program should leave a more demanding trail than a demonstration: the precise question, the data sources and versions, the model’s output, expert review, and the later result—including a negative one. Publishing those artifacts would separate three things that are often blurred together: retrieving information, producing a conjecture, and contributing evidence. The first can save time; the second can open a research path; only the third changes what it is reasonable to do for a patient.

The focus on shared resources such as Monarch matters because interoperability is also a form of quality control. If different databases define a disease, phenotype, or variant in incompatible ways, a model can produce a fluent synthesis without the comparison being valid. Data infrastructure is not a detail that comes before AI: it determines which claims a researcher can test afterward.

Open question

If basic-research results are published, what minimum evidence should accompany each AI-generated hypothesis? For this laboratory, a responsible answer would include data provenance, an evaluation method agreed before seeing the result, independent review, and a record of failures. Speed becomes scientific progress only when it makes errors easier to detect, not merely answers easier to produce.

Primary sources

• Anthropic, Apply for Anthropic’s AI for Science rare disease research grants, July 20, 2026.
• Monarch Initiative, project site and rare-disease data resources, accessed July 21, 2026.