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

The DOI matches its Crossref record; 98 of the 101 references it cites were found in a registry or an index; 14 of the 28 sources its claims cite were checked for support.

Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance

Open access
Bansal, Gagan, Wu, Tongshuang, Zhou, Joyce, Fok, Raymond, Nushi, Besmira, Kamar, Ece, Ribeiro, Marco Tulio, Weld, Daniel · 2021 · Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems

Abstract

Many researchers motivate explainable AI with studies showing that human-AI team performance on decision-making tasks improves when the AI explains its recommendations. However, prior studies observed improvements from explanations only when the AI, alone, outperformed both the human and the best team. Can explanations help lead to complementary performance, where team accuracy is higher than either the human or the AI working solo? We conduct mixed-method user studies on three datasets, where an AI with accuracy comparable to humans helps participants solve a task (explaining itself in some conditions). While we observed complementary improvements from AI augmentation, they were not increased by explanations. Rather, explanations increased the chance that humans will accept the AI’s recommendation, regardless of its correctness. Our result poses new challenges for human-centered AI: Can we develop explanatory approaches that encourage appropriate trust in AI, and therefore help generate (or improve) complementary performance?

Overview

CoChat Summary

TL;DR

This paper tests whether AI explanations produce complementary human-AI team performance when AI accuracy is comparable to human accuracy. Across three crowdsourced tasks, teams outperformed solo humans and solo AI, but explanations did not significantly improve team accuracy over simply showing AI confidence, and they increased acceptance of AI recommendations regardless of correctness.

Key findings

  • Complementary performance occurred in every human-AI teaming condition. For Beer, Team (Conf) reached μ = 0.89 ± 0.05, surpassing AI alone (0.84) and unassisted humans (0.82 ± 0.09); Amzbook and LSAT showed relative gains of 2.2% and 20.1% over unassisted workers.
  • Explanations did not significantly improve team accuracy over the confidence-only baseline. For Beer, Team (Conf) was 0.89 ± 0.05 vs. Team (Explain-Top-1, AI) 0.88 ± 0.06 (z = -1.18, p = .24); similar non-significant patterns appeared for Amzbook (p = .22) and LSAT (p = .64).
  • Explain-Top-1 and Explain-Top-2 did not differ significantly: Beer z = 0.85, p = .40; Amzbook z = 0.81, p = .42; LSAT z = 0.42, p = .68.
  • Explanations increased reliance on AI recommendations even when incorrect. Splitting by AI correctness, Explain-Top-1 improved accuracy when AI was correct but worsened it when AI was incorrect; Adaptive explanations reduced blind trust in low-confidence cases but did not significantly improve final team accuracy.

Methods

  • Mixed-method user studies on three tasks: sentiment analysis of Beer and Amzbook reviews and LSAT logical-reasoning questions. Crowdworkers were recruited from Amazon Mechanical Turk; 566 (Beer), 552 (Amzbook), and 508 (LSAT) participated, with about 100 participants per condition and 1,626 users total across tasks.
  • AI systems were RoBERTa-based classifiers. Sentiment explanations used LIME saliency highlights; LSAT used expert-generated narrative explanations. Confidence was calibrated with isotonic regression; Adaptive explanations switched between top-1 and top-2 based on median confidence thresholds of 89.2% (Beer) and 88.9% (Amzbook).
  • Main analysis compared team decision accuracy across conditions using Student’s T-tests with Bonferroni correction. Qualitative analysis coded participants’ self-reported collaboration strategies, with average Cohen’s κ = 0.71 and average agreement 93% ± 6.5%.

Limitations

  • The paper states that it evaluated only decision accuracy, not other possible benefits such as speed or multi-dimensional complementarity; it used crowdsourceable tasks and may not generalize to high-stakes expert domains; it explored only two explanation presentation formats; and it did not investigate user expertise.
  • Evident from the text: LSAT used only 20 examples and exclusively expert explanations; the Adaptive threshold was set by median confidence rather than learned; AI recommendations were shown before participants solved the task, which the authors note may invoke anchoring; and subjective survey ratings were analyzed post-hoc without claims of statistical significance.

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

Integrity signals

Registry

Crossref · 10.1145/3411764.3445717

Cites

98 of 98 references verified0 citation errors · 0 not checkable

Statistics

11 tests recomputed · 0 with the reported and recomputed p on opposite sides of .05

Verification

Claims the paper makes

11 Valid · 3 Not valid · 11 Not checked · 82 Own results and interpretations

44 more claims · 1 with their cited sources checked

Create a free account to see all 107 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 37
  • It may state two separate claims as one. 33
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

60 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Statistics that don’t add up

Each reported p-value recomputed from its test statistic and degrees of freedom.

1 of 1

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Doesn’t add upz = 0.427reported p = .64 · recomputed p = .669

    “z =0.427, p = .64”

  • Consistentz = 0.85reported p = .4 · recomputed p = .395

    “z =0.85, p = .40”

  • Consistentz = 0.81reported p = .42 · recomputed p = .418

    “z =0.81, p = .42”

  • Consistentz = 0.42reported p = .68 · recomputed p = .674

    “z =0.42, p = .68”

  • Consistentz = -1.18reported p = .24 · recomputed p = .238

    “z =-1.18, p = .24”

5 of 11 · the rest aren’t loaded

6 more recomputed tests

Create a free account to see every test the Stats check recomputed.

See all tests — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 101 references

Research works printed but not linked: 5

  1. 10.5555/647985.743843 · doi:10.5555/647985.743843
  2. 10.18653/v1 · doi:10.18653/v1
  3. 10.5555/3524938.3525594 · doi:10.5555/3524938.3525594
  4. 10.18653/v1 · doi:10.18653/v1
  5. LSAT Prep Books Team. 2017. LSAT prep book study guide: quick study & practice test questions for the Law School Admissions council’s (LSAC) Law school admission test. Mometrix Test Preparation, Beaumont, TX.

Cited by

See all 757

Metadata

Authors
Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D.
Affiliations
  • University of Washington (US)
  • Microsoft (United States) (US)
  • Microsoft Research, United States
Venue
Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems
Publisher
ACM
Published
2021-05-06
Pages
1-16
OpenAlex
W3037634279
Cited by
757
References
101
Citations by year
2026: 238 · 2025: 148 · 2024: 128 · 2023: 125 · 2022: 88 · 2021: 26
FWCI
42.30
Type
conference paper
Language
en
Open access
Locations
  • CHI Conference on Human Factors in Computing Systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Microsoft Research
  • the University of Washington WRF/Cable Professorship
  • Office of Naval Research (N00014-18-1-2193)
  • NSF RAPID grant (2040196)
  • the Allen Institute for Artificial Intelligence (AI2)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Human Factors and Ergonomics › Human-Automation Interaction and Safety
  • Physical Sciences › Computer Science › Human-Computer Interaction › Usability and User Interface Design
Keywords
explainable AI, AI explanations, human-AI teaming, AI augmentation, trust in AI, human-AI decision-making, recommendation acceptance, human-centered 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).

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: Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. https://app.cochat.ai/papers/10.1145/3411764.3445717
The paper
Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. (2021). Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3411764.3445717

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.