Kong, Dingwen, Lyu, Liang, Amin, Saurabh, Bennouna, Amine, Huttenlocher, Daniel, Ozdaglar, Asuman · Nov 4, 2026
AI-Assisted Classification under Correlation Neglect and Trust
Markdown · JSON · Original source
In their paper, Dingwen Kong, Liang Lyu, Saurabh Amin, Amine Bennouna, Daniel Huttenlocher and Asuman Ozdaglar argue that whether AI improves a person’s decision depends not only on AI accuracy, but also on how much its evidence overlaps with the person’s and how the person combines both signals. They model a worker making a binary decision using a personal signal and an AI recommendation, with two behavioral factors: awareness of evidence overlap and trust in the AI. Their analysis maps conditions for complementarity, impairment and automation. If a worker neglects shared evidence, a more capable AI can reduce accuracy by causing that evidence to be counted twice. Trust also changes the outcome: undertrust may offset correlation neglect, so the trust level associated with highest accuracy depends on deployment parameters. The authors say their empirical method pairs the same workers with different AIs to estimate the two factors and separate accuracy losses associated with correlation neglect and undertrust. In their reported ChatBench results, AI assistance raises accuracy by 21 points, but workers use only about a quarter of the Bayes-rational weight on AI information. They report different outcomes for GPT-4o and Llama-3.1-8b, and say trust calibration recovers about 2.6 points for GPT-4o but almost nothing for Llama-3.1-8b.
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
cs.AI · Article
Source details
Original source date: Nov 4, 2026 (2026-11-04).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T17:32:09.991Z. First observed: 2026-11-05T05:06:39.166Z. Retrieved: 2026-11-09T13:09:40.429Z.
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 supplied evidence is the paper abstract; it does not provide the detailed experimental setup or the context needed to interpret the reported accuracy-point changes beyond the statements quoted here.
Open question
How do the reported accuracy effects change across the deployment parameters that determine evidence overlap and workers’ trust?
Source-stated
The authors argue that AI’s effect on human decision-making depends on evidence overlap and how people combine their own signal with the AI recommendation, not only on AI accuracy.
the answer depends not just on AI accuracy but also on how much its evidence overlaps with the human's and how the human combines the two
Source-stated
Their model identifies awareness of evidence overlap and trust in the AI as two behavioral factors, and analyzes regions of complementarity, impairment and automation.
We develop a human behavioral model with two primitives: her awareness of the overlap and her trust in the AI.
Source-stated
The authors say that when a worker neglects overlap, a more capable AI can reduce accuracy because the worker double-counts shared evidence.
A worker who neglects the overlap (correlation neglect) can be made less accurate by a more capable AI, double-counting the shared evidence.
Source-stated
The authors report that undertrust may compensate for correlation neglect, and that the trust level maximizing accuracy depends on deployment parameters.
undertrust may compensate for correlation neglect, so that the trust level maximizing accuracy depends on the deployment parameters
Source-stated
For ChatBench, the authors report an accuracy increase of 21 points with AI assistance, while workers apply about a quarter of the Bayes-rational weight to AI information.
On ChatBench, AI assistance raises accuracy by 21 points, yet workers apply only about a quarter of the Bayes-rational weight to AI information
Source-stated
The authors report GPT-4o in the automation region and Llama-3.1-8b at the complementarity boundary; trust calibration recovers about 2.6 points for GPT-4o and almost nothing for Llama-3.1-8b.
GPT-4o falls in the automation region and Llama-3.1-8b stays at the complementarity boundary; and calibrating trust recovers about 2.6 points for GPT-4o but almost nothing for Llama-3.1-8b.
Source and contributor context
Kong, Dingwen, Lyu, Liang, Amin, Saurabh, Bennouna, Amine, Huttenlocher, Daniel, Ozdaglar, Asuman · 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.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 6022e637-6f77-4081-8bf9-a0cf2bff8a8d.
Prepared by gpt-6-luna; stream-draft-repair-v1; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T17:32:09.991Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: b8c05837a3a32ca75a12d2d6b1a22619fb5da3fdf3c0bac428ad1714dbfa08ee. Source text hash: 7693172642c46c4c36ce4fe8b8009162eca33801575394e22602159eb75982c1.
Revision and source-check history
Version 1 · 2026-11-10T17:32:09.991Z
AI-Assisted Classification under Correlation Neglect and Trust
In their paper, Dingwen Kong, Liang Lyu, Saurabh Amin, Amine Bennouna, Daniel Huttenlocher and Asuman Ozdaglar argue that whether AI improves a person’s decision depends not only on AI accuracy, but also on how much its evidence overlaps with the person’s and how the person combines both signals. They model a worker making a binary decision using a personal signal and an AI recommendation, with two behavioral factors: awareness of evidence overlap and trust in the AI. Their analysis maps conditions for complementarity, impairment and automation. If a worker neglects shared evidence, a more capable AI can reduce accuracy by causing that evidence to be counted twice. Trust also changes the outcome: undertrust may offset correlation neglect, so the trust level associated with highest accuracy depends on deployment parameters. The authors say their empirical method pairs the same workers with different AIs to estimate the two factors and separate accuracy losses associated with correlation neglect and undertrust. In their reported ChatBench results, AI assistance raises accuracy by 21 points, but workers use only about a quarter of the Bayes-rational weight on AI information. They report different outcomes for GPT-4o and Llama-3.1-8b, and say trust calibration recovers about 2.6 points for GPT-4o but almost nothing for Llama-3.1-8b.
Original source
cs.AI · Article
Source details
Original source date: Nov 4, 2026 (2026-11-04).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T17:32:09.991Z. First observed: 2026-11-05T05:06:39.166Z. Retrieved: 2026-11-09T13:09:40.429Z.
Limits
The supplied evidence is the paper abstract; it does not provide the detailed experimental setup or the context needed to interpret the reported accuracy-point changes beyond the statements quoted here.
Open question
How do the reported accuracy effects change across the deployment parameters that determine evidence overlap and workers’ trust?
Source-stated
The authors argue that AI’s effect on human decision-making depends on evidence overlap and how people combine their own signal with the AI recommendation, not only on AI accuracy.
the answer depends not just on AI accuracy but also on how much its evidence overlaps with the human's and how the human combines the two
Source-stated
Their model identifies awareness of evidence overlap and trust in the AI as two behavioral factors, and analyzes regions of complementarity, impairment and automation.
We develop a human behavioral model with two primitives: her awareness of the overlap and her trust in the AI.
Source-stated
The authors say that when a worker neglects overlap, a more capable AI can reduce accuracy because the worker double-counts shared evidence.
A worker who neglects the overlap (correlation neglect) can be made less accurate by a more capable AI, double-counting the shared evidence.
Source-stated
The authors report that undertrust may compensate for correlation neglect, and that the trust level maximizing accuracy depends on deployment parameters.
undertrust may compensate for correlation neglect, so that the trust level maximizing accuracy depends on the deployment parameters
Source-stated
For ChatBench, the authors report an accuracy increase of 21 points with AI assistance, while workers apply about a quarter of the Bayes-rational weight to AI information.
On ChatBench, AI assistance raises accuracy by 21 points, yet workers apply only about a quarter of the Bayes-rational weight to AI information
Source-stated
The authors report GPT-4o in the automation region and Llama-3.1-8b at the complementarity boundary; trust calibration recovers about 2.6 points for GPT-4o and almost nothing for Llama-3.1-8b.
GPT-4o falls in the automation region and Llama-3.1-8b stays at the complementarity boundary; and calibrating trust recovers about 2.6 points for GPT-4o but almost nothing for Llama-3.1-8b.
Source and contributor context
Kong, Dingwen, Lyu, Liang, Amin, Saurabh, Bennouna, Amine, Huttenlocher, Daniel, Ozdaglar, Asuman · 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.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 6022e637-6f77-4081-8bf9-a0cf2bff8a8d.
Prepared by gpt-6-luna; stream-draft-repair-v1; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T17:32:09.991Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: b8c05837a3a32ca75a12d2d6b1a22619fb5da3fdf3c0bac428ad1714dbfa08ee. Source text hash: 7693172642c46c4c36ce4fe8b8009162eca33801575394e22602159eb75982c1.
uncertain · 2026-11-10T17:32:09.991Z · anthropic/claude-opus-5.5
Publication: 6022e637-6f77-4081-8bf9-a0cf2bff8a8d
attribution — pass: The six authors are named exactly as listed on the arXiv page. Arguments and results are consistently framed as the authors' claims or reported findings, not as established fact.
Authors:Dingwen Kong, Liang Lyu, Saurabh Amin, Amine Bennouna, Daniel Huttenlocher, Asuman Ozdaglar
summary — uncertain: The summary is faithful but hard to follow without specialist background. Central terms are left unexplained: complementarity, impairment and automation regions; Bayes-rational weight; ChatBench; undertrust. They need tooltips or one-line glosses. Smallest repair 1: restore the specific GPT-4o automation and Llama-3.1-8b complementarity-boundary placements instead of the vague 'different outcomes'. Smallest repair 2: restore the practical split the abstract draws, with one cost tied to reducing overlap on the AI side and the other to calibrating trust on the human side. That split is the source's actionable stake. The question is substantive and open.
one tied to reducing overlap on the AI side, the other to calibrating trust on the human side
date — pass: originalPublished 2026-11-04 matches the arXiv v1 submission date. It is not the later feed date.
[Submitted on 4 Nov 2026]
support — pass: Every claim is source-stated, and each excerpt is verbatim from the abstract. Hedges such as 'may compensate' and 'can be made less accurate' are preserved. No unsupported benefits, costs or forecasts are added. The limitation accurately notes that only the abstract is available.
undertrust may compensate for correlation neglect, so that the trust level maximizing accuracy depends on the deployment parameters
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to 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.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv