# arXiv cs.AI: Autonomy Is Absent, Not Eroded [Canonical reading](https://escapemeridian.com/readings/arxiv-cs-ai-autonomy-is-absent-not-eroded-671ebbc2) Authors Chaoran Chen, Isabel Barberá and Yuhuai Wei ask whether prolonged interaction with Large Language Models reduces users’ observable autonomy during conversation. Scoring 2,877 multi-turn conversations from WildChat and validating on 500 conversations from LMSYS-Chat-1M, they report a null result: population-level autonomy remains stable throughout conversations, and only 4.7% show meaningful decline. They say an apparent decrease at later turn positions is explained by survivorship bias. Their central finding shifts the concern from autonomy eroding over time to autonomy often being absent: 80.8% of conversations exhibit zero or low autonomy throughout. When autonomy engagement occurs, the authors report that persistence doubles from 40% to 85% over consecutive turns. They frame the risk not as LLMs progressively degrading users’ observable autonomy, but as most users never engaging in independent reasoning during the interaction. The abstract does not define how user autonomy or meaningful decline was scored. 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 arXiv cs.AI · Article Original source date: Sep 6, 2026 (2026-09-06). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T16:01:21.012Z. First observed: 2026-11-10T08:34:12.446Z. Retrieved: 2026-11-10T15:49:27.946Z. - [Autonomy Is Absent, Not Eroded: An Observational Study of User Autonomy in Multi-Turn LLM Conversations](https://arxiv.org/abs/2611.09923) ## Limits The abstract does not define how user autonomy or meaningful decline was scored. The abstract gives the size of the validation set but does not describe the validation procedure or its results separately. ## Open question How was user autonomy scored, and how might that scoring affect the finding that most conversations show zero or low autonomy? ## Source-stated The authors analyzed 2,877 multi-turn conversations and checked their findings against 500 conversations from another dataset. > scoring user autonomy across 2,877 multi-turn conversations (46,952 turns) from WildChat and validating on 500 conversations from LMSYS-Chat-1M - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors report stable population-level autonomy and meaningful decline in only a small share of conversations. > population-level autonomy remains stable throughout conversations, only 4.7% of conversations show meaningful decline - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors say an apparent later-turn decrease is explained by survivorship bias. > an apparent decrease at later turn positions is explained by survivorship bias - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors report that most conversations show zero or low autonomy throughout. > 80.8% of conversations exhibit zero or low autonomy throughout - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors report that autonomy engagement, when present, persists more often over consecutive turns. > When autonomy engagement occurs, it is self-reinforcing (persistence doubles from 40% to 85% over consecutive turns). - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors frame the concern as users often not engaging in independent reasoning, rather than progressively losing autonomy during interaction. > the risk is not that LLMs progressively degrade users' observable autonomy, but that most users never engage in independent reasoning during the interaction. - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source and contributor context Chen, Chaoran, Barberá, Isabel, Wei, Yuhuai · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 19d6aa4a-4dff-40b1-89c3-47e1b804e129. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T16:01:21.012Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: 11af30663a6972c00f46a8b92bf81c7020ed3e7e541b2e7599d780bb2aa38080. Source text hash: d26a3c380e461c9e8cd72c4aaad7037ca4ff90775383734885ab9ee3a7672da6. ## Revision and source-check history ### Version 1 — 2026-11-10T16:01:21.012Z Authors Chaoran Chen, Isabel Barberá and Yuhuai Wei ask whether prolonged interaction with Large Language Models reduces users’ observable autonomy during conversation. Scoring 2,877 multi-turn conversations from WildChat and validating on 500 conversations from LMSYS-Chat-1M, they report a null result: population-level autonomy remains stable throughout conversations, and only 4.7% show meaningful decline. They say an apparent decrease at later turn positions is explained by survivorship bias. Their central finding shifts the concern from autonomy eroding over time to autonomy often being absent: 80.8% of conversations exhibit zero or low autonomy throughout. When autonomy engagement occurs, the authors report that persistence doubles from 40% to 85% over consecutive turns. They frame the risk not as LLMs progressively degrading users’ observable autonomy, but as most users never engaging in independent reasoning during the interaction. The abstract does not define how user autonomy or meaningful decline was scored. ## Original source arXiv cs.AI · Article Original source date: Sep 6, 2026 (2026-09-06). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T16:01:21.012Z. First observed: 2026-11-10T08:34:12.446Z. Retrieved: 2026-11-10T15:49:27.946Z. - [Autonomy Is Absent, Not Eroded: An Observational Study of User Autonomy in Multi-Turn LLM Conversations](https://arxiv.org/abs/2611.09923) ## Limits The abstract does not define how user autonomy or meaningful decline was scored. The abstract gives the size of the validation set but does not describe the validation procedure or its results separately. ## Open question How was user autonomy scored, and how might that scoring affect the finding that most conversations show zero or low autonomy? ## Source-stated The authors analyzed 2,877 multi-turn conversations and checked their findings against 500 conversations from another dataset. > scoring user autonomy across 2,877 multi-turn conversations (46,952 turns) from WildChat and validating on 500 conversations from LMSYS-Chat-1M - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors report stable population-level autonomy and meaningful decline in only a small share of conversations. > population-level autonomy remains stable throughout conversations, only 4.7% of conversations show meaningful decline - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors say an apparent later-turn decrease is explained by survivorship bias. > an apparent decrease at later turn positions is explained by survivorship bias - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors report that most conversations show zero or low autonomy throughout. > 80.8% of conversations exhibit zero or low autonomy throughout - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors report that autonomy engagement, when present, persists more often over consecutive turns. > When autonomy engagement occurs, it is self-reinforcing (persistence doubles from 40% to 85% over consecutive turns). - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source-stated The authors frame the concern as users often not engaging in independent reasoning, rather than progressively losing autonomy during interaction. > the risk is not that LLMs progressively degrade users' observable autonomy, but that most users never engage in independent reasoning during the interaction. - [Supporting source](https://arxiv.org/abs/2611.09923) ## Source and contributor context Chen, Chaoran, Barberá, Isabel, Wei, Yuhuai · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 19d6aa4a-4dff-40b1-89c3-47e1b804e129. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T16:01:21.012Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: 11af30663a6972c00f46a8b92bf81c7020ed3e7e541b2e7599d780bb2aa38080. Source text hash: d26a3c380e461c9e8cd72c4aaad7037ca4ff90775383734885ab9ee3a7672da6. ### passed — 2026-11-10T16:01:21.012Z — anthropic/claude-opus-5.5 Publication: 19d6aa4a-4dff-40b1-89c3-47e1b804e129 attribution — pass: The draft correctly names Chaoran Chen, Isabel Barberá and Yuhuai Wei as authors. It keeps every finding and the risk framing attributed to them ('they report', 'the authors say') rather than presenting them as established fact. > Authors:Chaoran Chen (University of Notre Dame), Isabel Barberá (Independent Researcher), Yuhuai Wei (Independent Researcher) summary — pass: The summary leads with the authors' question and their null result. It preserves the paper's reframing from erosion to absence, including its force, and keeps the self-reinforcing persistence figure. It also flags that autonomy scoring is undefined. The question addresses a genuine unresolved issue. Terms such as survivorship bias, WildChat and LMSYS-Chat-1M are not explained inline. They are better suited to term-library tooltips, and the core meaning stays readable without them. One small optional repair would restore the authors' word 'self-reinforcing'. > Instead, the dominant pattern is not erosion but absence: 80.8% of conversations exhibit zero or low autonomy throughout. date — pass: originalPublished is 2026-09-06, which matches the v1 submission date on the arXiv page. The later feed date of 2026-11-10 is correctly not used. > \[Submitted on 6 Sep 2026\] support — pass: Each claim matches the abstract in figures and scope, including 2,877 conversations, the 500-conversation validation set, 4.7%, 80.8% and 40% to 85%. The survivorship-bias explanation and the risk reframing are presented as the authors' assertions. The limitations accurately note what the abstract omits. No benefits, costs or forecasts are invented. > only 4.7% of conversations show meaningful decline, and an apparent decrease at later turn positions is explained by survivorship bias media — pass: The selected media is the verified arXiv abstract page for this paper, which is the direct original source and matches the source URL. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/671ebbc2566ac1eaad75103bcc9bc1e56ca3cb0b8e55d5245e6e3084db9acc98/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/arxiv-cs-ai-autonomy-is-absent-not-eroded-671ebbc2.json)