Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 71 of the 76 references it cites were found in a registry or an index; 23 of the 38 sources its claims cite were checked for support.

Improving Human-AI Collaboration With Descriptions of AI Behavior

Open access
Cabrera, Ángel Alexander, Perer, Adam, Hong, Jason I. · 2023 · Proceedings of the ACM on Human-Computer Interaction

Abstract

People work with AI systems to improve their decision making, but often under- or over-rely on AI predictions and perform worse than they would have unassisted. To help people appropriately rely on AI aids, we propose showing them behavior descriptions, details of how AI systems perform on subgroups of instances. We tested the efficacy of behavior descriptions through user studies with 225 participants in three distinct domains: fake review detection, satellite image classification, and bird classification. We found that behavior descriptions can increase human-AI accuracy through two mechanisms: helping people identify AI failures and increasing people's reliance on the AI when it is more accurate. These findings highlight the importance of people's mental models in human-AI collaboration and show that informing people of high-level AI behaviors can significantly improve AI-assisted decision making.

Overview

CoChat Summary

TL;DR

This paper tests whether showing people short, subgroup-level descriptions of an AI’s behavior helps them collaborate better with the AI. In three Amazon Mechanical Turk classification studies with 225 participants, behavior descriptions improved performance in two tasks, but did not change participants’ reported trust, helpfulness, or willingness to reuse the AI.

Key findings

  • Across the three tasks, behavior descriptions significantly improved overall accuracy in 2 of 3 domains: reviews and birds; the satellite task showed no significant overall improvement.
  • In the reviews task, accuracy differed across conditions with F(2,64) = 9.18, p < 0.001; the only significant pairwise difference was between No AI and AI + BD (p < 0.001, 95% CI [0.06, 0.22]).
  • In the birds task, accuracy differed across conditions with F(2,65) = 3.98, p = 0.023; the only significant pairwise difference was between No AI and AI + BD (p = 0.048, 95% CI [0.00, 0.16]), and this produced complementary human-AI accuracy.
  • The proposed mechanism in H2 was only partially supported: in reviews, there were no accuracy differences for the behavior-description groups, while in birds, group 2 showed a significant difference (F(2,65) = 6.81, p = 0.002) and supported H2; overall, increased accuracy also came from greater reliance on the AI in the main group.

Methods

  • Three between-subjects human-subject experiments were run on Amazon Mechanical Turk with 225 U.S. participants total, across fake review detection, satellite image classification, and bird classification.
  • Each participant labeled 30 instances: 20 from the main group and 5 each from two behavior-description groups; in the AI + BD condition, participants saw a behavior description for those subgroup instances.
  • Main analyses used ANOVA with Tukey post-hoc tests for accuracy, Mann-Whitney U tests with Bonferroni corrections for Likert items, and linear models to test learning effects.

Limitations

  • The paper states that participants were Amazon Mechanical Turk workers with limited domain expertise, so results may differ for domain experts/professionals; the authors explicitly say future studies should test that.
  • The study was a controlled simulated setting, so effects may differ in real-world consequential settings such as radiology; the paper explicitly notes this.
  • The paper says it tested one specific type of behavior description—subgroup accuracy—so the generality to other behavior-description formats is not established; the text makes evident this is a design-scope limitation.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3579612

Cites

71 of 71 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Don Norman's mental model framework [43] describes how designers use their mental models to implement systems. End-users then interact with the systems and develop their own mental models of how they believe the systems work. While this process is similar for AI models, a key difference is that an AI is not a direct representation of a developer's intent, but a stochastic model learned from data. This means that (1) AI developers themselves have to make sense of what an AI system has learned through testing and iteration. Subsequently, they can encode these insights as (2) behavior descriptions , details of how an AI performs on subgroups of instances, that can be shown to end-users to improve human-AI collaboration.
Fig. 1 Don Norman's mental model framework [43] describes how designers use their mental models to implement systems. End-users then interact with the systems and develop their own mental models of how they believe the systems work. While this process is similar for AI models, a key difference is that an AI is not a direct representation of a developer's intent, but a stochastic model learned from data. This means that (1) AI developers themselves have to make sense of what an AI system has learned through testing and iteration. Subsequently, they can encode these insights as (2) behavior descriptions , details of how an AI performs on subgroups of instances, that can be shown to end-users to improve human-AI collaboration.From the PDF
Fig. 2 Experimental setup . Each participant was randomly assigned to a condition and dataset (3x3 between-subjects study, 25 participants per condition, 225 participants total). For each dataset, participants saw 30 instances, 20 instances from the whole dataset with an AI accuracy of 95%, and 5 instances each from two subsets of the dataset with an AI accuracy of 40% and 20% respectively (simulating subgroups behavior descriptions would be useful for). In the AI + Behavior Description condition, participants were shown behavior descriptions for instances in group 1 and 2. While the instances shown to participants were randomly chosen from a larger subset of data, each participant saw the same number of AI errors to ensure they observed the same AI accuracies.
Fig. 2 Experimental setup . Each participant was randomly assigned to a condition and dataset (3x3 between-subjects study, 25 participants per condition, 225 participants total). For each dataset, participants saw 30 instances, 20 instances from the whole dataset with an AI accuracy of 95%, and 5 instances each from two subsets of the dataset with an AI accuracy of 40% and 20% respectively (simulating subgroups behavior descriptions would be useful for). In the AI + Behavior Description condition, participants were shown behavior descriptions for instances in group 1 and 2. While the instances shown to participants were randomly chosen from a larger subset of data, each participant saw the same number of AI errors to ensure they observed the same AI accuracies.From the PDF
Fig. 3 UI screenshots for the fake reviews (left) and satellite image classification (right) tasks. Each participant labeled 30 instances, distributed according to the instance groups described in Figure 2 and Section 4.1.2. Both screenshots are shown on a labeling step for the AI + Behavior Description (BD) condition and on instances that are part of a behavior description group. In the AI condition participants are not shown the additional text for instances in a BD group, and in the No AI condition participants are not shown the AI output. The bird classification task used the same format as the satellite classification task shown.
Fig. 3 UI screenshots for the fake reviews (left) and satellite image classification (right) tasks. Each participant labeled 30 instances, distributed according to the instance groups described in Figure 2 and Section 4.1.2. Both screenshots are shown on a labeling step for the AI + Behavior Description (BD) condition and on instances that are part of a behavior description group. In the AI condition participants are not shown the additional text for instances in a BD group, and in the No AI condition participants are not shown the AI output. The bird classification task used the same format as the satellite classification task shown.From the PDF
Image 4
From the PDF
Fig. 4 Average participant accuracy by task and condition. The vertical orange bar indicates the AI accuracy, what would be the participant's accuracy if they picked the AI response every time. The blue shaded area indicates complementarity , the region where the human+AI accuracy is higher than either the human or AI alone. We find that behavior descriptions led to higher accuracy in the reviews and birds tasks, with complementarity in the birds task (red point in rightmost chart). The error bars represent standard error.
Fig. 4 Average participant accuracy by task and condition. The vertical orange bar indicates the AI accuracy, what would be the participant's accuracy if they picked the AI response every time. The blue shaded area indicates complementarity , the region where the human+AI accuracy is higher than either the human or AI alone. We find that behavior descriptions led to higher accuracy in the reviews and birds tasks, with complementarity in the birds task (red point in rightmost chart). The error bars represent standard error.From the PDF
Fig. 5 Average team accuracy by task, condition and instance group. We further break down accuracy by instance type (see Section 4.1.1): the main group (20 instances), and two behavior description groups (5 instances each). The average human-AI accuracy across the three groups of instances gives us an idea of how behavior descriptions improve the performance of human-AI teams. We find that participants relied more on the AI when shown BDs in every task. Participant performance on the different behavior description groups was mixed, from no effect to significant improvement in group 2 birds (bottom right). These results highlight the two effects of behavior descriptions, increasing human reliance on a more accurate AI and overriding systematic AI errors. The error bars represent standard error.
Fig. 5 Average team accuracy by task, condition and instance group. We further break down accuracy by instance type (see Section 4.1.1): the main group (20 instances), and two behavior description groups (5 instances each). The average human-AI accuracy across the three groups of instances gives us an idea of how behavior descriptions improve the performance of human-AI teams. We find that participants relied more on the AI when shown BDs in every task. Participant performance on the different behavior description groups was mixed, from no effect to significant improvement in group 2 birds (bottom right). These results highlight the two effects of behavior descriptions, increasing human reliance on a more accurate AI and overriding systematic AI errors. The error bars represent standard error.From the PDF
Fig. 6 Likert-scale responses on perception of AI. The diverging stacked bar chart centered around the neutral response shows that participants across all conditions and subjective measures overwhelmingly viewed the AI favorably. There were no significant differences in user's perception of the AI when they were give behavior descriptions.
Fig. 6 Likert-scale responses on perception of AI. The diverging stacked bar chart centered around the neutral response shows that participants across all conditions and subjective measures overwhelmingly viewed the AI favorably. There were no significant differences in user's perception of the AI when they were give behavior descriptions.From the PDF
Fig. 7 Learning curves by task and condition. We fit linear models of accuracy on round number to measure learning effects. We found that the two conditions in which BDs significantly improved performance also had significant learning effects, the AI + BD conditions in the reviews and birds tasks (denoted by *).
Fig. 7 Learning curves by task and condition. We fit linear models of accuracy on round number to measure learning effects. We found that the two conditions in which BDs significantly improved performance also had significant learning effects, the AI + BD conditions in the reviews and birds tasks (denoted by *).From the PDF
Image 9
From the PDF
Fig. 8 The instructions shown to participants for each task before labeling.
Fig. 8 The instructions shown to participants for each task before labeling.From the PDF

Verification

Claims the paper makes

14 Valid · 3 Not valid · 12 Not checked · 84 Own results and interpretations

54 more claims · 3 with their cited sources checked

Create a free account to see all 113 claims and to verify the 3 references still open.

See all claims — free account
  • It may state two separate claims as one. 34
  • Probably not a substantial claim. 24

Create a free account to see these claims.

Citations

Cites

Show all 76 references

Research works printed but not linked: 5

  1. Ruogu Kang, Laura Dabbish, Nathaniel Fruchter, and Sara Kiesler. 2015. My Data Just Goes Everywhere: User Mental Models of the Internet and Implications for Privacy and Security. In Eleventh Symposium On Usable Privacy and Security (SOUPS 2015). USENIX Association, Ottawa, 39--52. https://www.usenix.org/conference/soups2015/proceedings/presentation/kang
  2. 10.1038/s41586-019--1138-y · doi:10.1038/s41586-019--1138-y
  3. 10.18653/v1 · doi:10.18653/v1
  4. 10.1007/s10648-010--9128--5 · doi:10.1007/s10648-010--9128--5
  5. P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. 2010. Caltech-UCSD Birds 200. Technical Report CNS-TR-2010-001. California Institute of Technology.

Cited by

See all 87

Metadata

Authors
Cabrera, Á. A., Perer, A., & Hong, J. I.
Affiliations
Carnegie Mellon University (US)
Venue
Proceedings of the ACM on Human-Computer Interaction
Published
2023-04-14
Volume
7
Issue
CSCW1
Pages
1-21
ISSN
2573-0142 (ISSN-L)
OpenAlex
W4317464833
Cited by
87
References
76
Citations by year
2026: 37 · 2025: 30 · 2024: 18 · 2023: 2
FWCI
25.67
Type
journal article
Language
en
Open access
Locations
  • Proceedings of the ACM on Human-Computer Interaction (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
National Science Foundation (IIS-2040942)
Topics
  • Social Sciences › Human Factors and Ergonomics › Human-Automation Interaction and Safety
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Physical Sciences › Computer Science › Human-Computer Interaction › Usability and User Interface Design
Keywords
human-AI collaboration, behavioral description, AI-assisted decision-making, trust in AI, mental models, fake review detection, satellite image classification, bird classification, explainable AI

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Improving Human-AI Collaboration With Descriptions of AI Behavior. https://app.cochat.ai/papers/10.1145/3579612
The paper
Cabrera, Á. A., Perer, A., & Hong, J. I. (2023). Improving Human-AI Collaboration With Descriptions of AI Behavior. Proceedings of the ACM on Human-Computer Interaction. https://doi.org/10.1145/3579612

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.