· Oct 27, 2026
Anthropic authors report that analysts often followed AI agents’ assigned preferences
Markdown · JSON · Original source
Amaç Herdağdelen, Hailey Joren and Alex Tamkin report a study of professional analysts using Code for a statistical analysis. The task involved choices with more than one reasonable answer, including how to measure farm productivity and compare farms. Among 50 analysts who used the agent, choices matched its deliberately assigned on the four analytical decisions 86% of the . The authors also report that analysts sometimes repeated the agent’s arguments afterward without crediting it, and often still described the recommendation as their own. The study suggests a possible challenge for work where experts must make judgment calls: an agent may shape decisions even when the human reviewer believes their judgment is independent. The authors caution that they strengthened the agent’s preferences to measure their effect, and that the study used one task, one field and one model. They also report that analysts using the agent finished faster and made fewer serious errors, but the groups were not comparable enough to attribute those differences to AI use. How often do agents shape experts’ judgments in ordinary work when their preferences are not deliberately strengthened?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Oct 27, 2026 (2026-10-27T00:00:00.000Z).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-14T08:32:55.734Z. First observed: 2026-10-27T19:07:35.761Z. Retrieved: 2026-10-29T16:34:07.290Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The authors say the study used one task in one field, with one model, and the analysts were experts in causal inference but new to farming. The agent’s preferences were deliberately strengthened; the authors say their effect in real-world work depends on factors including experts’ own views and model behavior. The group that worked without AI was small and less experienced, so the reported difference in errors and task cannot establish that using AI caused the difference. The study measured average effects across analysts and did not record each participant’s own view before the agent expressed its .
Open question
How often do agents shape experts’ judgments in ordinary work when their preferences are not deliberately strengthened?
Source-stated
In the study, the choices of analysts who used the agent matched its assigned lean on the four analytical decisions 86% of the time.
Across all four combined, the result was 86%.
Source-stated
The authors report that analysts repeated an agent’s argument in at least one answer in 77% of cases, while 15% credited the agent anywhere.
77% repeated their agent’s argument in at least one answer, sometimes word for word, and only 15% credited the agent anywhere.
Source-stated
The authors caution that the study’s deliberately strengthened agent preferences limit how directly its results apply to real-world work.
We deliberately strengthened the agents’ lean. Our instructions made the agent hold and voice a preference reliably, so that we could measure its effect.
Source-stated
Analysts who used the agent finished faster and made fewer serious errors in this task, but the study does not establish that AI use caused the difference.
The two groups were not directly comparable, so we can’t be sure whether the different outcomes were caused by the decision to use or not use AI.
Source and contributor context
Anthropic · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Lean
Lean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 03ffef29-3f63-4d97-8b5f-f2fe37964d1f.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T08:32:55.734Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: 5def434111c5c110c628a5f3b8b7bc93848de6105f4dc073e959b1fff6233b51. Source text hash: ad23094aed1ac64680eeb446f5163dd36260f0c13314071577a235a6a711f216.
Revision and source-check history
Version 1 · 2026-11-14T08:32:55.734Z
Anthropic authors report that analysts often followed AI agents’ assigned preferences
Amaç Herdağdelen, Hailey Joren and Alex Tamkin report a study of professional analysts using Code for a statistical analysis. The task involved choices with more than one reasonable answer, including how to measure farm productivity and compare farms. Among 50 analysts who used the agent, choices matched its deliberately assigned on the four analytical decisions 86% of the . The authors also report that analysts sometimes repeated the agent’s arguments afterward without crediting it, and often still described the recommendation as their own. The study suggests a possible challenge for work where experts must make judgment calls: an agent may shape decisions even when the human reviewer believes their judgment is independent. The authors caution that they strengthened the agent’s preferences to measure their effect, and that the study used one task, one field and one model. They also report that analysts using the agent finished faster and made fewer serious errors, but the groups were not comparable enough to attribute those differences to AI use. How often do agents shape experts’ judgments in ordinary work when their preferences are not deliberately strengthened?
Original source
· Article
Source details
Original source date: Oct 27, 2026 (2026-10-27T00:00:00.000Z).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-14T08:32:55.734Z. First observed: 2026-10-27T19:07:35.761Z. Retrieved: 2026-10-29T16:34:07.290Z.
Limits
The authors say the study used one task in one field, with one model, and the analysts were experts in causal inference but new to farming. The agent’s preferences were deliberately strengthened; the authors say their effect in real-world work depends on factors including experts’ own views and model behavior. The group that worked without AI was small and less experienced, so the reported difference in errors and task cannot establish that using AI caused the difference. The study measured average effects across analysts and did not record each participant’s own view before the agent expressed its .
Open question
How often do agents shape experts’ judgments in ordinary work when their preferences are not deliberately strengthened?
Source-stated
In the study, the choices of analysts who used the agent matched its assigned lean on the four analytical decisions 86% of the time.
Across all four combined, the result was 86%.
Source-stated
The authors report that analysts repeated an agent’s argument in at least one answer in 77% of cases, while 15% credited the agent anywhere.
77% repeated their agent’s argument in at least one answer, sometimes word for word, and only 15% credited the agent anywhere.
Source-stated
The authors caution that the study’s deliberately strengthened agent preferences limit how directly its results apply to real-world work.
We deliberately strengthened the agents’ lean. Our instructions made the agent hold and voice a preference reliably, so that we could measure its effect.
Source-stated
Analysts who used the agent finished faster and made fewer serious errors in this task, but the study does not establish that AI use caused the difference.
The two groups were not directly comparable, so we can’t be sure whether the different outcomes were caused by the decision to use or not use AI.
Source and contributor context
Anthropic · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Lean
Lean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 03ffef29-3f63-4d97-8b5f-f2fe37964d1f.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T08:32:55.734Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: 5def434111c5c110c628a5f3b8b7bc93848de6105f4dc073e959b1fff6233b51. Source text hash: ad23094aed1ac64680eeb446f5163dd36260f0c13314071577a235a6a711f216.
uncertain · 2026-11-14T08:32:55.734Z · anthropic/claude-opus-5.5
Publication: 03ffef29-3f63-4d97-8b5f-f2fe37964d1f
attribution — pass: The draft correctly attributes the study and its findings to Anthropic authors Amaç Herdağdelen, Hailey Joren and Alex Tamkin. It keeps the results and cautions as the authors' reported claims rather than established facts.
Amaç Herdağdelen, Hailey Joren, Alex Tamkin
summary — pass: A general reader can follow the summary. It names the actors, the task and the hidden assigned lean, and gives the headline 86% result. It also covers the authors' stated stake that a human reviewer may not be an independent check, plus the key limits and the speed and error caveat. Two improvements would add force without being required: the 60% vs 6% switching result, and the authors' worry that it creates false consensus and weakens oversight of AI safety research. The question follows directly from the authors' own stated limitation.
a human reviewer is only an independent check if their judgment is their own
date — pass: originalPublished is 2026-10-27, based on the publisher page, and it matches the feed entry. The draft makes no conflicting date claim.
support — uncertain: Claim 2 says analysts repeated the agent's argument 'in 77% of cases.' The source reports that 77% of analysts (participants) did this in at least one answer, not 77% of cases or answers. As worded, a reader could take the figure as a per-decision rate. Smallest repair: change it to '77% of analysts repeated their agent's argument in at least one answer.' The other claims match the source, including the 86% figure, the strengthened lean, the non-comparable groups, and the single task, field and model.
77% of participants repeated their agent’s argument in at least one answer, 15% credited the agent.
media — pass: The selected media is the verified direct link to the original Anthropic research page. It is the most direct way for readers to reach the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeLean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Explore LeanTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMELean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Explore LeanClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeLean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Explore LeanTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMELean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Explore Lean