Skip to content
Citations Citations
Blog About Request Access

How Researchers Use AI: Anthropic Counted the Tasks, Not the Sources

Francois-Xavier Bioul
Francois-Xavier Bioul · CCO at Citations LLC
8 min read

On November 5, 2026, Anthropic published a measurement publishers rarely get from an AI platform: what researchers use its assistant for. Its economics team classified 207,049 research sessions on Claude and Claude Code by scientific field and by task, from data analysis to manuscript drafting.

The study counts tasks across 34 fields. It does not say which journals, books or databases those tasks drew on, and it was designed that way. For rights holders, it shows what one platform can already classify about usage, and which question it leaves to licensing.

In short

Anthropic's study shows that an AI platform can classify research sessions by field and task, at scale, while retaining only those labels. It does not show which sources the sessions relied on.

Writing and data analysis lead the data. Drafting manuscripts is the most frequent single task on Claude, followed by literature review and replies to peer reviewers. Some of these tasks typically draw on published work, and the study does not measure how much.

In Claude Science, which connects the model to research databases, sessions are estimated to be longer, more specialized and more autonomous. For publishers, the practical step is to ask for usage reporting by use type and by licensed work, with a disclosure threshold and a named verifier, and to test whether the platform can deliver it.

What Anthropic measured, and how

The main sample covers 207,049 sessions on consumer plans between August 2 and September 30, 2026, of which 128,914 were on Claude's chat and Cowork interfaces and 78,135 on Claude Code. A supplemental sample covers Claude Science, Anthropic's research environment.

A language model reads each transcript and assigns labels from a fixed taxonomy, SciNet, which lists 7,262 research tasks across 34 fields and 318 subfields. Only the labels are retained for analysis. Anthropic states that no classifier extracts or stores research content such as datasets, hypotheses, findings or unpublished papers, and results are reported only where enough distinct accounts contribute.

A survey of 528 researchers, mostly in the life and health sciences in the US and Canada, was used to cross-check the patterns. The report's measures of task time, expertise and autonomy are estimates produced by classifiers rather than direct observations.

What the task data shows

Data analysis is the largest category, with 33.7% of Claude sessions and 48.1% of Claude Code sessions. Writing and communication accounts for 32% of Claude sessions, against an estimated 16% of researcher time.

At task level, drafting manuscripts accounts for 18.6% of Claude sessions, ahead of reviewing and synthesizing literature (6.8%) and responding to peer reviewers (6.7%). The most prominent output is a written artifact, such as a document, report or paper, in 52% of sessions.

Literature synthesis works on published scholarship by definition, and a manuscript usually cites it. Data analysis, by contrast, may run entirely on a team's own unpublished data. Because sources are not among the study's labels, it cannot separate these cases, and the share of sessions that used licensed content is unknown.

Relative to their researcher counts, the social sciences and the arts and humanities use Claude more than the study's capability measure predicts, and the authors suggest that much of their work is done through text. Law shows one of the widest gaps between written and technical outputs. For publishers in these fields, the finding justifies an investigation and does not by itself show use of their content.

Connected tools: longer, more autonomous sessions

Claude Science connects the model to research tools and databases that scientists would otherwise query one by one. In biology and biomedical sciences, Anthropic estimates that its sessions involve work that would take a person 10.3 hours alone, against 4.9 on Claude. Over 80% are classified as requiring doctoral-level training, and 73.3% as running with high AI autonomy, against 23.7% on Claude.

Subject mix explains only part of the gap. Reweighting Claude sessions to the same subfields and tasks leaves Claude at about 5 hours and 25% high-autonomy use. Anthropic still describes the differences as patterns of use rather than effects of the tool, since Claude Science users may have more demanding needs.

For a rights holder, the relevance lies in the channel more than in any single session. Connected tools are a route through which licensed databases can enter an AI workflow, and high autonomy means fewer steps a person reviews directly. The study does not say which databases were queried or under what terms. The measurement problem this raises is set out in machine readership and the reading crisis.

What the method shows, and what it does not

One argument against AI usage reports is that reporting would expose what users are working on. Anthropic's design answers part of this concern. It classifies sessions into fixed categories, retains only the labels, and publishes cells above a minimum number of accounts.

The rest remains open. Identifying the exact work behind a passage, qualifying the act (retrieval, display, citation), deduplicating across sessions and making the count auditable are separate problems, and the study addresses none of them. Adding work identifiers to aggregate labels would also carry its own risks, such as re-identification in small subfields or exposure of commercially sensitive usage.

The table separates what the study did from what a licensing report would need. The right-hand column is a proposed requirement that Anthropic has not demonstrated.

Dimension

In Anthropic's study (observed)

Proposed for a licensing usage report

Unit

Research session

Usage event tied to a licensed work

Context

Field and subfield (34 / 318)

Field or customer segment, where the license prices by segment

Purpose

Research task (7,262 in SciNet)

Use type: indexing, retrieval-augmented generation (RAG), citation, agent access

Source identity

Not among the labels

Work or collection identifier

Retained content

Labels only

Labels and identifiers, no prompts or user content

Disclosure rule

Minimum number of distinct accounts

Same, written into the contract

Verification

Internal; classifiers documented in an appendix

Named producer and an independent check

Where the comparison stops

Four limits keep the study from serving as a licensing record. It covers consumer plans over two months. It is a platform measuring itself, and its labels are classifier estimates. Its unit is a share of sessions rather than a count per work. Its survey draws on applicants to Anthropic's AI-for-science grant programs, whom the authors describe as likely more intensive users than the typical scientist.

Usage standards are not settled either. COUNTER's best practice on AI usage metrics separates passages an AI could access from passages selected for the final answer, and states that AI reporting is not subject to its formal audit. As of its September 18, 2026 update, its phase-two telemetry events (retrieval, grounding, citation, presentation and engagement) were awaiting approval.

The open question is therefore who produces the count and who checks it, as discussed in who audits the usage number.

The objection: "Researchers already pay for access"

Much of this use may involve content researchers reach through institutional subscriptions, so the objection is that the publisher has already been paid. That is often true for the reading itself.

Whether a subscription also covers uploading an article into a third-party assistant, or a connected tool fetching it for the researcher, depends on the license. Some licenses now address AI use explicitly, as CCC's higher education license shows, and who may put content into an AI tool is a separate question.

Access logs, API traces or proxy records can show part of the route. On their own, they do not tie that access to a work, a use type and a license term.

What to ask for next

First, use the task data to frame an investigation rather than to estimate use. Cross the tasks the study ranks highest, namely literature synthesis, manuscript drafting, data analysis and replies to reviewers, with the disciplines and collections you license, then check what your own access and API logs show for those collections.

Second, in any license covering connected or agent access, ask for aggregate reporting by use type and by work or collection, above a stated account threshold, with no user content. Ask the platform to demonstrate it on a test collection before signature, since a reporting clause does not by itself guarantee that the data exists or is reliable. Match each line to the record it requires, as in a use-by-use licensing matrix.

Third, require the reporting party to document its classifiers and accept an independent check of a sample.

Anthropic's study documents research tasks without attributing them to works. Linking a use to a work, a right and a commercial term is the job of AI usage evidence, the layer Citations is built to provide.

Book an AI usage evidence review

How AI use of reference catalogues escapes human-reading metrics is covered in machine readership and the reading crisis.

Why author disclosure cannot reconstruct an AI-assisted manuscript is set out in AI disclosure is not an audit trail.

Frequently asked questions

How do researchers use AI tools like Claude?

In Anthropic's study of 207,049 research sessions, data analysis and writing lead, at about a third of Claude sessions each. Drafting manuscripts is the most frequent single task, followed by literature synthesis and replies to peer reviewers. Idea generation and data gathering are less frequent.

Does Anthropic's study show which sources researchers used?

No. It labels each session by field and task only, and states that no research content is extracted or stored. Source identity is not among its labels.

Can an AI platform report content usage without exposing user data?

Anthropic's method shows that aggregate classification by task can be done while retaining only labels. Reporting per licensed work adds harder problems: identifying the work, qualifying the use and auditing the count. Whether a platform provides it depends on its capabilities and on the contract.

Sources

Anthropic — AI for scientific discovery: How researchers use Claude, November 5, 2026
How researchers use Claude

Anthropic — full report (PDF)
AI for Scientific Discovery

COUNTER Metrics — Best Practice on Generative and Agentic AI Usage Metrics, updated June 30, 2026
Best Practice on Generative and Agentic AI Usage Metrics

COUNTER Metrics — AI update, September 18, 2026
COUNTER AI update, September 2026

This article is operational analysis for publishing and rights teams, not legal advice.