CoChat

Emerging Reliance Behaviors in Human-AI Content Grounded Data Generation: The Role of Cognitive Forcing Functions and Hallucinations

Open access
Ashktorab, Zahra, Desmond, Michael, Pan, Qian, Johnson, James, Brachman, Michelle, Dugan, Casey, Danilevsky, Marina, Geyer, Werner · 2025 · Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work

The DOI matches its Crossref record; 66 of the 66 references it cites were found in a registry or an index; 10 of the 12 sources its claims cite were checked for support.

Abstract

We investigate the impact of hallucinations and Cognitive Forcing Functions in human-AI collaborative content grounded data generation, focusing on the use of Large Language Models (LLMs) to assist in generating high-quality conversational data. Through a study with 34 users who each completed 8 tasks (n=272), we found that hallucinations significantly reduce data quality. While Cognitive Forcing Functions do not always alleviate these effects, their presence influences how users integrate AI responses. Specifically, we observed emerging reliance behaviors, with users often appending AI-generated responses to their correct answers, even when the AI’s suggestions conflicted. This points to a potential drawback of Cognitive Forcing Functions, particularly when AI suggestions are inaccurate. Users who overrelied on AI-generated text produced lower-quality data, emphasizing the nuanced dynamics of overreliance in human-LLM collaboration compared to traditional human-AI decision-making.

Overview

CoChat Summary

TL;DR

In a 34-participant study (8 tasks each, n=272) of LLM-assisted content-grounded dialogue creation, hallucinated AI responses significantly reduced data quality and significantly reduced users' incorporation of AI text, while cognitive forcing functions (CFFs) did not mitigate either effect. The authors identify new reliance behaviors in co-creation, including users appending hallucinated AI text to their own already-correct answers.

Key findings

  • Hallucinations significantly lowered data quality (F(1, 229) = 73.07, p < .001): quality was M = 4.34 (SE = 0.137) with hallucinations present vs. M = 5.77 (SE = 0.137) absent, mean difference −1.43 (t(229) = −8.548, p < .0001). H1 accepted.
  • Hallucinations significantly decreased AI use (F(1, 229) = 463.519, p < .001): AI usage M = 0.133 (SE = 0.0299) with hallucinations vs. M = 0.916 (SE = 0.0299) without (t(229) = −21.529, p < .0001). H3 accepted.
  • CFFs did not mitigate hallucination effects: no significant main effect of CFF type (F(2, 31) = 0.19, p = .83) or CFF presence (F(1, 229) = 3.12, p = .079), and no significant CFF × CFF presence × hallucination interaction on quality (F(2, 229) = 1.870, p = .15); H2A–H2C and H4A–H4C rejected. A significant CFF Presence × CFF interaction on AI use (F(2, 229) = 4.944, p = 0.008) showed more AI use in Formulate when the CFF was absent (M = 0.614, SE = 0.0523) than in Highlight (M = 0.423, SE = 0.0481), t(74.5) = 2.684, p = 0.0907.
  • Reliance classification: 13 underreliers, 9 appropriate reliers, 6 both over- and underreliers, 6 solely overreliers. Reliance type significantly affected summed data quality (F(3, 268) = 110.05, p < 0.0001): appropriate reliers (M = 43.33) and underreliers (M = 42.77) outperformed overreliers (M = 36) and overreliers-underreliers (M = 34.33) (p < 0.01 and p < 0.05 respectively). Five hallucination-reliance patterns were coded: exact copy, copy and minor edit, exact copy and append with correct response, copy and removal of information, copy and added content from the reference document.

Methods

  • Mixed between-within web experiment: 34 participants recruited via internal Slack channels and word of mouth, each completing 8 HR/customer-support tasks (n = 272). Between-subjects variable was CFF Type (Formulate, Read-First, Highlight); within-subjects variables were CFF presence (visible or not) and hallucination presence. Four tasks contained hallucinated AI responses and four did not; task order was randomized.
  • Tasks were drawn from real HR support documents (8 unique reference documents, M = 555.25 words, SD = 274.41) with prior annotator-created conversations; final turns used AI responses from Flan-T5-XXL and a fine-tuned content-grounded variant. Gold responses were produced independently by three authors and reconciled.
  • Data quality was rated with a 6-item yes/no rubric spanning faithfulness, accuracy, and completeness (adapted from Grice's maxims), plus one item on whether any AI-generated text appeared in the final response. Analysis used linear mixed-effects models with random intercepts per participant and F-tests with Satterthwaite approximations; reliance-type effects were tested with ANOVA and Tukey post-hoc comparisons.

Limitations

  • Stated: the focus on conversational customer support may limit applicability to other domains; gold-response creation may not be scalable; the participant pool was restricted to internal company members due to HR document confidentiality; the small sample size limits generalizability, and future work should test larger, more diverse pools.
  • Stated: only ~25% of participants reported being from HR, though participants on average reported high confidence in composing high-quality data.
  • Evident from the text: reliance was measured as a binary indicator of whether any AI text appeared in the final response, and the authors note that logged copy-button events were not a reliable proxy for AI text usage, requiring manual rubric-based evaluation that they describe as neither scalable nor efficient.

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

Integrity signals

Registry

Crossref · 10.1145/3729176.3729179

Cites

66 of 66 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

8 Valid · 0 Not valid · 2 Not checked · 60 Own results and interpretations

31 more claims

Create a free account to see all 70 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 26
  • It may state two separate claims as one. 12

Create a free account to see these claims.

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.

  • Not confirmedF(2, 229) = 1.87reported p = .15 · recomputed p = .156

    “F(2, 229) = 1.870, p = .15”

  • ConsistentF(1, 229) = 73.07reported p < .001 · recomputed p < .001

    “F(1, 229) = 73.07, p < .001”

  • Consistentt(229) = -8.548reported p < .0001 · recomputed p < .001

    “t(229) = -8.548, p < .0001”

  • ConsistentF(2, 31) = 0.19reported p = .83 · recomputed p = .828

    “F(2, 31) = 0.19, p = .83”

  • ConsistentF(1, 229) = 3.12reported p = .079 · recomputed p = .079

    “F(1, 229) = 3.12, p = .079”

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.

5 of 6 · the rest aren’t loaded

1 more finding

Create a free account to see every finding of this check.

See all findings — free account

Citations

Cites

Show all 66 references

Research works printed but not linked: 3

  1. Tim Fischer, Steffen Remus, and Chris Biemann. 2022. Measuring Faithfulness of Abstractive Summaries. In Proceedings of the 18th Conference on Natural Language Processing (KONVENS 2022), Robin Schaefer, Xiaoyu Bai, Manfred Stede, and Torsten Zesch (Eds.). KONVENS 2022 Organizers, Potsdam, Germany, 63–73. https://aclanthology.org/2022.konvens-1.8/
  2. Daniel Kahneman. 2011. Thinking, Fast and Slow. Macmillan, New York, NY.
  3. Chin-Yew Lin. 2004. ROUGE: A Package for Automatic Evaluation of Summaries. In Text Summarization Branches Out. Association for Computational Linguistics, Barcelona, Spain, 74–81. https://aclanthology.org/W04-1013/

Cited by

See all 10

Metadata

Authors
Ashktorab, Z., Desmond, M., Pan, Q., Johnson, J., Brachman, M., Dugan, C., Danilevsky, M., & Geyer, W.
Affiliations
  • IBM (United States) (US)
  • IBM Research - Cambridge (United States) (US)
  • RAND Corporation (US)
  • IBM Research - Almaden (US)
Venue
Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work
Publisher
ACM
Published
2025-06-22
Pages
1-17
OpenAlex
W4411505989
Cited by
10
References
57
Citations by year
2026: 10
FWCI
10.33
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • Annual Symposium on Human-Computer Interaction for Work (CHIWORK) (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
Keywords
human-AI collaboration, hallucinations, human-AI interaction

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 2026-10-27, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-27 (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: Emerging Reliance Behaviors in Human-AI Content Grounded Data Generation: The Role of Cognitive Forcing Functions and Hallucinations. https://app.cochat.ai/papers/10.1145/3729176.3729179
The paper
Ashktorab, Z., Desmond, M., Pan, Q., Johnson, J., Brachman, M., Dugan, C., Danilevsky, M., & Geyer, W. (2025). Emerging Reliance Behaviors in Human-AI Content Grounded Data Generation: The Role of Cognitive Forcing Functions and Hallucinations. Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work. https://doi.org/10.1145/3729176.3729179

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.