PaperlayerView exact revision on arXiv

Source-linked AI summary

Advancing Human-AI Complementarity: The Impact of User Expertise and Algorithmic Tuning on Joint Decision Making

Kori Inkpen, Shreya Chappidi, Keri Mallari, Besmira Nushi, Divya Ramesh, Pietro Michelucci, Vani Mandava, Libuše Hannah Vepřek, Gabrielle Quinn

arXiv:2208.07960v1cs.HCcs.AI

TL;DR

This paper examines how to achieve Human-AI teams that outperform either humans or AI alone, focusing on model tuning, user expertise, and perceptions. In a Stall Catchers study, AI recommendations improved team performance overall, but outcomes varied with user expertise and perceptions.

  • Problem

    Human-AI teams are expected to outperform humans or AI alone, but models trained on historical data are rarely optimized for collaborative goals and may underperform AI alone.

    Source: §2 · p. 2

  • Method

    The study examines how tuning true positive versus true negative rates intersects with human expertise and perceptions of AI.

    Source: §2 · p. 2

  • Results

    AI recommendations significantly improved team performance on Stall Catchers overall, but gains varied and some users’ performance degraded, with benefits especially variable for users near AI performance.

    Source: §6 · p. 22, §7 · p. 27

  • Takeaways & Limitations

    Collaborative deployment may require choosing the model best suited to the human partner rather than simply the most accurate model.

    Source: §2 · p. 2

    The available user population was limited, producing smaller samples of high performers and unequal expertise distributions across conditions that may constrain conclusions.

    Source: §6.7 · p. 26

Abstract

from arXiv · show

Human-AI collaboration for decision-making strives to achieve team performance that exceeds the performance of humans or AI alone. However, many factors can impact success of Human-AI teams, including a user's domain expertise, mental models of an AI system, trust in recommendations, and more. This work examines users' interaction with three simulated algorithmic models, all with similar accuracy but different tuning on their true positive and true negative rates. Our study examined user performance in a non-trivial blood vessel labeling task where participants indicated whether a given blood vessel was flowing or stalled. Our results show that while recommendations from an AI-Assistant can aid user decision making, factors such as users' baseline performance relative to the AI and complementary tuning of AI error types significantly impact overall team performance. Novice users improved, but not to the accuracy level of the AI. Highly proficient users were generally able to discern when they should follow the AI recommendation and typically maintained or improved their performance. Mid-performers, who had a similar level of accuracy to the AI, were most variable in terms of whether the AI recommendations helped or hurt their performance. In addition, we found that users' perception of the AI's performance relative on their own also had a significant impact on whether their accuracy improved when given AI recommendations. This work provides insights on the complexity of factors related to Human-AI collaboration and provides recommendations on how to develop human-centered AI algorithms to complement users in decision-making tasks.

2 INTRODUCTION

The introduction frames complementary Human-AI teams as a goal that ordinary accuracy-focused model development may not achieve, and asks how algorithmic tuning, user expertise, and perceptions affect collaboration. It presents the Stall Catchers study as a way to examine these factors and guide human-centered algorithm design.

2 INTRODUCTION

  • Models trained on historical data are rarely optimized for collaboration, so team performance may fall below AI-alone performance.Source: §2 · p. 2
  • Collaborative deployment may favor a model that fits the human partner over the most accurate model, depending on AI mistakes and explanations.Source: §2 · p. 2
  • The study asks how TPR/TNR tuning affects team performance, how expertise interacts with tuning, and how user strategies and perceptions shape assistance.Source: §2 · p. 3
  • The study uses Stall Catchers, where participants label mouse-brain blood vessels as flowing or stalled, a complex task spanning varied expertise.Source: §2 · p. 3
  • Users’ baseline expertise significantly impacts team performance, and users performing near AI accuracy are especially sensitive to algorithmic tuning.Source: §2 · p. 3
  • Users’ perceptions of AI—including relative performance, trust, and usefulness—significantly impact the utility of an AI-Assistant.Source: §2 · p. 3
  • Tuning true positive and true negative rates must account for user characteristics to improve Human-AI team performance.The study treats these rates as practical controls for model bias.Source: §2 · pp. 2–3
  • The work identifies opportunities for human-centered algorithms that complement users’ limitations, expertise, and perceptions of AI.Source: §2 · p. 3

3 RELATED WORK

Prior work reports both gains and failures in human-AI collaboration, with AI properties and human perceptions shaping assistance and team performance. Stall Catchers provides the citizen-science task context for this study.

3.1 Human AI Collaboration for decision making

  • Studies report both improved team performance with AI assistance and cases where assistance did not improve performance.Pathology studies reported higher accuracy with algorithm-assisted pathologists, while a mammography study associated computer-assisted diagnosis with decreased sensitivity and no change in overall performance or specificity.Source: §3.1 · p. 4

3.1.1 Findings on the impact of AI and AI properties on team performance.

  • Prior work examined how stated versus observed accuracy affects trust, but had not tested how tuning error types affects overall team performance or complementarity.A scheduling-assistant study compared systems with equal 50% accuracy but different false-positive and false-negative emphasis.Source: §3.1.1 · p. 4

3.1.2 Findings on the impact of human perception on team performance.

  • Human trust and reliance can diverge: users may disuse automated aids despite self-reported bias toward them, and explanations may induce automation bias among familiar users.Prior work also connected team outcomes to alignment between human and AI error boundaries, which can limit the benefit of confidence information.Source: §3.1.2 · p. 5

3.1.3 Algorithmic Aversion and Appreciation.

  • Algorithmic aversion describes preferring human advice over algorithmic advice even when the algorithm is more accurate; algorithmic appreciation describes heavier reliance on equivalent algorithmic forecasts.Prior work also reports that task expertise can relate to different evaluations of AI-labeled advice.Source: §3.1.3 · pp. 5–6

3.2 Stall Catchers

  • Stall Catchers crowdsources mouse-brain vessel labels for Alzheimer’s research, and its machine-learning models showed varied true-positive and true-negative rates without expert-like performance.The experimental interface removed game features and added a robot icon indicating the AI recommendation.Source: §3.2 · pp. 6–7

4 STUDY

The study recruited experienced and novice Stall Catchers players for a four-stage vessel-labeling experiment. Participants completed the task with and without simulated AI recommendations under conditions that varied the AI’s true-positive and true-negative rates while holding accuracy constant.

4.1 Participants

  • 58 remote U.S. Stall Catchers players aged 18–82 participated, with expertise ranging from novice to expert.All were recruited from the platform’s existing player pool and completed the task once.Source: §4.1 · p. 7

4.2 Videos

  • Participants labeled 140 videos sampled from a 7,693-video validation set with expert-verified ground truth, balanced across four stages by difficulty and vessel class.Each video depicted blood flow through mouse-brain vessels and had a difficulty metric based on historical classification accuracy.Source: §4.2 · p. 8, §4.3 · p. 8

4.3 Task Stages

  • The four stages progressed from unaided practice with feedback to unaided testing, then AI-assisted practice with feedback and AI-assisted testing without feedback.Stage 2 measured unaided skill, while Stage 4 assessed human-AI team performance.Source: §4.3.1 · p. 8, §4.3.2 · p. 8, §4.3.3 · p. 9

4.4 Experimental Design

  • A mixed-factor design assigned 58 participants among three AI-tuning conditions and compared each participant’s performance with and without AI assistance.Participants were not balanced across prior expertise levels by experimental condition.Source: §4.4 · p. 9

4.4.1 Experimental Conditions.

  • The simulated assistant held accuracy at 75% while varying true-positive and true-negative rates across balanced, high-TNR, and high-TPR conditions.Simulation enabled precise control of the rates for comparison across stages.Source: §4.4.1 · pp. 9–10

4.5 Data Collection and Procedures

  • Researchers recorded decisions, AI recommendations, ground truth, and response times, and surveyed perceived performance, AI utility, perceived error types, use, perceptions, and trust.A survey modification early in the study left some survey data missing for the first 10 participants.Source: §4.5 · pp. 10–11

5 RESULTS

AI recommendations significantly raised users’ accuracy overall, while baseline skill was associated with the size of performance gains. The results also distinguish users whose unaided accuracy was below, near, or above the AI’s benchmark.

5.1 Overall Impact of the AI-Assistant on Performance

  • Users achieved higher accuracy with AI recommendations than without them, with no significant differences among the three tuning conditions.The Stage 2-to-Stage 4 difference was significant (F1,55 = 24.61, p < .001, η2 = .31); the condition effect was not (F2,55 = .61, p = .55, η2 = .02).Source: §5.1 · p. 11
  • Stage comparisons showed no significant accuracy differences between Stage 1 and Stage 2 or between Stage 3 and Stage 4, suggesting no significant change between stages except when AI recommendations were added.The reported tests were t(1,57) = .001, p = .87 and t(1,57) = 1.86, p = .068, respectively.Source: §5.1 · p. 11

5.1.1 Baseline Performance Clusters.

  • Baseline accuracy negatively correlated with gains alongside AI (r(57) = −.59, p < .001), so lower-baseline users made larger gains.The figure compares Stage 2 unaided accuracy with Stage 4 AI-assisted accuracy and marks the AI’s 75% accuracy benchmark.Source: §5.1.1 · pp. 11–12, Figure 4 · p. 13
  • The 16 low performers and 33 mid performers scored significantly below the AI’s 75% accuracy at baseline, while nine high performers scored significantly above it.The resulting baseline-performance clusters were used in subsequent analyses.Source: §5.1.1 · p. 12

Result 2a: Regardless of AI-Assistant tuning, the accuracy of low-performers significantly increased,

AI assistance improved low- and mid-performer accuracy, though low-performers remained below the AI and mid-performer outcomes varied. High-performers maintained accuracy above the AI regardless of assistance or tuning.

Result 2a: Regardless of AI-Assistant tuning, the accuracy of low-performers significantly increased,

  • AI recommendations improved low- and mid-performer accuracy, while high-performers maintained their accuracy; low-performers remained below the AI and mid-performer outcomes varied.The low- and mid-performer improvements were statistically significant; high-performers’ accuracy did not significantly change.Source: §5.2.1 · p. 14, §5.2.2 · pp. 14–15, §5.2.3 · p. 15

5.2.1 Low-Performers.

  • Low-performer accuracy improved significantly: 15 of 16 improved, one stayed the same, none declined, and Stage 4 accuracy remained below the AI-Assistant.AI tuning had no significant effect on low-performer accuracy.Source: §5.2.1 · p. 14

5.2.2 Mid-Performers.

  • Mid-performer outcomes varied: 20 of 33 improved, two stayed the same, and 11 fell below their Stage 2 accuracy.High TPR significantly improved mid-performer accuracy; 75% exceeded AI accuracy in that condition, compared with 38% in High TNR and 17% in Balanced.Source: §5.2.2 · p. 15

5.2.3 High-Performers.

  • High-performers remained above the AI-Assistant in Stage 4, averaging 88% accuracy, with no significant change between stages or across tuning conditions.The authors note that the small number of high-performers may have limited statistical power.Source: §5.2.3 · p. 15

5.3 Complementarity: The Impact of Algorithmic Tuning

Users benefited most when AI error tuning complemented their baseline strengths. Mid- and high-performers had higher baseline true-negative than true-positive rates, while condition-specific assistance improved different rates.

5.3 Complementarity: The Impact of Algorithmic Tuning

  • Users benefited most when AI tuning complemented their baseline performance; significant gains occurred only on measures where users’ baseline was below the AI-Assistant.The study compared users’ true-positive and true-negative rates across Balanced, High TNR, and High TPR conditions.Source: §5.3 · p. 16

5.3.1 Human Tuning.

  • Mid- and high-performers had significantly higher baseline true-negative than true-positive rates, indicating stronger detection of flowing than stalled vessels.Low-performers showed no statistically significant bias toward either metric.Source: §5.3.1 · p. 16

5.3.2 A Balanced AI-Assistant.

  • In the Balanced condition, users’ true-positive rate rose from 57% to 66%, while their true-negative rate did not change significantly.The AI-Assistant’s true-positive and true-negative rates were each 75%; users’ Stage 4 true-positive rate remained below the AI’s.Source: §5.3.2 · p. 16

5.3.3 A High TNR AI-Assistant.

  • In the High TNR condition, users’ true-negative rate rose to 84% but remained below the AI-Assistant, while their true-positive rate did not change significantly.The assistant was tuned to 95% true-negative and 55% true-positive rates.Source: §5.3.3 · pp. 16–17

5.3.4 A High TPR AI-Assistant.

  • In the High TPR condition, users’ true-positive rate rose to 72% but remained below the AI-Assistant, while their true-negative rate did not change significantly.The assistant was tuned to 95% true-positive and 55% true-negative rates.Source: §5.3.4 · p. 17

5.4 Agreement with AI-Assistant

User agreement with AI recommendations depended on whether the AI was correct, its tuning, and users’ baseline performance. High TPR users were less likely than High TNR users to agree with incorrect recommendations.

5.4 Agreement with AI-Assistant

  • The study treated agreement with correct AI recommendations and disagreement with incorrect recommendations as beneficial actions for team performance.Overall agreement rates did not significantly differ across performance clusters.Source: §5.4 · p. 18, §5.4.2 · p. 18

5.4.1 Agreement by Condition.

  • Users were less likely to agree with incorrect recommendations in the High TPR condition than in the High TNR condition, while tuning did not significantly affect agreement with correct recommendations.The AI was correct 75% of the time in each condition.Source: §5.4.1 · p. 18, Table 5 · p. 18

5.4.2 Agreement by Performance Cluster.

  • High-performers were more likely than mid- or low-performers to agree with correct AI recommendations and disagree with incorrect ones.Mid-performers were also more likely than low-performers to disagree with incorrect recommendations.Source: §5.4.2 · p. 18

5.5 Users’ Perception of the AI-Assistant

Users who perceived the AI as outperforming them were more likely to improve with its recommendations, while those rating themselves higher tended to degrade. Users who improved also rated the AI more favorably.

5.5 Users’ Perception of the AI-Assistant

  • Users who rated the AI-Assistant above their own performance were more likely to improve; those who rated themselves higher tended to degrade with AI recommendations.Among users who improved, the mean self-rating was 2.87 versus 3.43 for the AI; among those who degraded, the corresponding means were 3.55 and 3.09.Source: §5.5 · pp. 19–20
  • Users who improved rated the AI-Assistant higher for helpfulness, satisfaction, and improving their ability to catch stalls than users whose accuracy decreased.Across all respondents, 54% found the AI helpful, 47% were satisfied, and 34% felt it improved their ability to find stalls.Source: §5.5 · pp. 19–20
  • After AI feedback, users initially rated the AI below themselves; after further interaction, their ratings of themselves and the AI no longer significantly differed.AI performance ratings increased significantly from Stage 3 to Stage 4, while self-ratings did not significantly change.Source: §5.5 · pp. 19–20

5.6 User Mental Models of AI-Assistant’s Errors

After Stage 4, most users were unsure how the Stall Catchers assistant made decisions or what kinds of errors it made. Many also perceived its errors as equally frequent for stalled and flowing vessels.

5.6 User Mental Models of AI-Assistant’s Errors

  • 72% of users disagreed or were unsure that they knew what types of errors the assistant made.The passage reports this alongside users’ difficulty understanding the assistant’s decisions.Source: §5.6 · p. 20
  • 62% of users believed the assistant made equal numbers of errors on stalled and flowing vessels, regardless of condition.Table 8 reports users’ mental models of assistant error types after Stage 4.Source: §5.6 · p. 21, Table 8 · p. 21

5.7 Self-Reported Use of AI

Participants reported several ways of using the assistant, including consulting it when unsure, checking disagreements, ignoring recommendations, and revisiting videos flagged as stalled.

5.7 Self-Reported Use of AI

  • Researchers coded optional responses with bottom-up open coding, followed by axial coding, into four main use strategies and an “other” category.Two researchers coded all responses, with 85% inter-rater reliability.Source: §5.7 · p. 21

5.7.1 Follow recommendations if unsure.

  • Eleven mid-performers who consulted the assistant when unsure had mixed outcomes in Balanced and High TNR conditions, while all four in High TPR improved above the assistant’s accuracy.Four of five Balanced users performed worse; both High TNR users improved.Source: §5.7.1 · pp. 21–22

5.7.2 Take a second look / verify.

  • Ten users said they used recommendations to verify decisions or revisit a video when the assistant disagreed; this strategy could direct extra scrutiny to users’ weaker areas.The proposed benefit depends on the AI complementing, rather than sharing, users’ errors.Source: §5.7.2 · p. 22

5.7.3 Ignore.

  • Six users, all mid-performers, reported ignoring recommendations, and four of them had lower accuracy in Stage 4.The authors report that mid-performers struggled with whether to use the assistant.Source: §5.7.3 · p. 22

5.7.4 Check stalled videos.

  • Five users said they acted mainly on stall recommendations, revisiting those videos to look for stalls they might have missed.Their feedback reinforced the utility of High TPR tuning, which suggests more stalls.Source: §5.7.4 · p. 22

6 DISCUSSION

AI recommendations significantly improved team performance on the Stall Catchers task, but the extent of gains varied across users and experimental factors. The discussion identifies expertise, tuning, perceptions, and interaction strategies as factors shaping human-AI performance.

6 DISCUSSION

  • AI recommendations significantly improved team performance, though some users saw marginal gains or performance declines, and gains often fell short of the AI’s level.The authors’ goal was to understand how to maximize team performance beyond the human or AI alone.Source: §6 · p. 22

6.1 Baseline Expertise

  • Baseline expertise was the largest determinant: low performers improved but stayed below AI accuracy, high performers remained above it, and mid-performers had mixed outcomes.The authors note that studies recruiting mainly novices may face similar issues.Source: §6.1 · pp. 22–23

6.2 Tuning for Complementarity

  • High TPR tuning complemented mid-performers’ high-TNR bias and yielded more consistent gains; however, incorrect recommendations in users’ strong areas could erode trust and reliance.Across conditions, users were more likely to override incorrect recommendations in the High TPR condition.Source: §6.2 · p. 23

6.3 Perceived Performance, Trust, Helpfulness, Satisfaction

  • Users who rated the assistant above their own performance were more likely to benefit, while those rating them similarly tended to lose performance; helpfulness and satisfaction also predicted success.The authors examined self-ratings of users’ own and AI performance as indicators of perceived performance and possibly trust.Source: §6.3 · p. 24

6.4 Mental Model Formation

  • Most users lacked an accurate mental model of the assistant’s errors, and complementary tuning may therefore have been less effective in supporting mental-model formation.The study withheld information about the algorithm’s tuning and overall accuracy; the authors suggest presenting reliable tuning information may help users form a mental model.Source: §6.4 · pp. 24–25

6.5 Role of the AI-Assistant

  • Users treated the assistant variously as an arbitrator or second reader, and some blindly followed or ignored it; identical expertise and mental models could still yield different outcomes.The study did not instruct users to use the AI in a particular way.Source: §6.5 · p. 25

6.6 Tuning for Tasks/Scenarios or Users

  • Complementary tuning produced the largest gains in this study, but models trained on human annotations may lack enough complementarity, and more complex tasks may require other forms of assistance.The authors propose personalization by user preference and expertise, while noting that efficiency, productivity, and cognitive effort may matter for complex tasks.Source: §6.6 · p. 26

6.7 Limitations

  • The available user population was limited, recommendations appeared before users decided, and Stall Catchers’ stalled-vessel framing may have biased attention toward stalls.The small, uneven expertise groups constrain some conclusions; the study also lacked reciprocal comments about checking flowing recommendations.Source: §6.7 · pp. 26–27

7 CONCLUSION AND FUTURE WORK

This study examines how algorithmic tuning, human expertise, and user perceptions relate to Human-AI team performance in the Stall Catchers citizen-science task. Its findings point to complementary models and suitable interactions as opportunities to boost performance, and motivate future work on personalized AI partners.

7 CONCLUSION AND FUTURE WORK

  • The study found that AI benefits varied with expertise, with outcomes especially variable for users whose performance resembled the AI’s.The work examined tuning, expertise, and user perceptions in a complex decision-making task on Stall Catchers.Source: §7 · p. 27
  • The authors identify opportunities to boost team performance through models complementary to human capabilities and interactions that support appropriate mental models, trust, and confidence.Source: §7 · p. 27
  • Future work could personalize AI partners by learning user models over time or grouping users by expertise, and could target examples difficult for experts while improving productivity.The authors note that historical data may initially be sparse and that assistance can be beneficial when it helps accurate decisions happen faster.Source: §7 · p. 27

Related papers

Loading 2208.07960v1…