Enabling independent research on how people use Claude

Key Findings

• — Each of three external groups analysed roughly 250,000 Claude.ai or Claude Code conversations from April to May 2026 through Anthropic Insights, with aggregate outputs only and a third-party privacy audit.
• — Stanford's SALT Lab found that over half of Claude conversations involved people delegating consequential tasks, defined as work that affects others or is hard to undo, contrary to prior research suggesting people keep such tasks for themselves.
• — Consequential delegation was most likely when people sought professional guidance, particularly on legal or financial questions.
• — In nearly three-quarters of conversations people set the direction while Claude assisted, and they usually adapted its output rather than using it verbatim, but varied in how much they understood or learned from what it produced.
• — Oxford's early results suggest affect patterns in Claude chats resemble those in everyday internet browsing; METR's early results suggest more capable models save users more time; both writeups are pending.
• — Anthropic's review rights were limited to user privacy, information enabling usage-policy violations, confidential information and factual accuracy, and the company states it intends to expand the programme.

Methodology Notes

Pilot programme description with summarised third-party findings; observation window April to May 2026; roughly 250,000 conversations per group; the SALT Lab's own writeup ('Human-AI Collaboration at Scale', Shao, Zhao, Padmakumar, Wang, Yang, 249,834 conversations) is hosted on alphaXiv rather than arXiv as of 2026-09-15. Published 2026-08-26 (date on the page); fetched directly from anthropic.com. Findings are Anthropic's summary of partner analyses of Anthropic's own usage data, with the privacy audit acknowledged in the text.

Sources