Anthropic trial: 77% of analysts echoed their agent's arguments, only 15% credited it
paulnovosad · x · 2026-10-28
Anthropic researcher Alex Tamkin shared results from a randomized trial with 50 professional data analysts delegating work to an AI agent.
- When asked afterward why they made each call, 77% repeated the agent's argument in at least one answer, sometimes word for word
- Only 15% credited the agent for the reasoning
- About half called the final recommendation more their own than the agent's
Side-by-side examples show task-time agent quotes nearly verbatim in participants' post-hoc explanations. The study exposes an illusion of idea ownership when experts delegate to agents, with direct implications for how judgment shifts in human-AI collaboration.
More from AGI Musings
- Jeff Clune traces Genie's arc: from a 'crazy' 2019 idea to Hollywood-ready world model — databoydg · 2026-10-28
- Claude Science seen as Anthropic's bid to own life sciences' economic version control — teortaxesTex · 2026-10-28
- Classic coding interviews are a red flag: hiring must change for the AI era — omarsar0 · 2026-10-28
- The rise of the lone scientist under agentic AI — Symbiot10000 · 2026-10-28
- Can AI search automatically discover cross-field math connections like Fermat's proof? — IgorCarron · 2026-10-28
- Kevin Naughton: losing your job to AI is about mindset, not skillset — KevinNaughtonJr · 2026-10-28