cs.HCMar 24, 2026

Biased Error Attribution in Multi-Agent Human-AI Systems Under Delayed Feedback

Authors: Teerthaa Parakh, Karen M. Feigh

Organizations: Georgia Institute of Technology, Atlanta, GA, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Human decision-making is strongly influenced by cognitive biases, particularly under conditions of uncertainty and risk. While prior work has examined bias in single-step decisions with immediate outcomes and in human interaction with a single autonomous agent, comparatively little attention has been paid to decision-making under delayed outcomes involving multiple AI agents, where decisions at each step affect subsequent states. In this work, we study how delayed outcomes shape decision-making and responsibility attribution in a multi-agent human-AI task. Using a controlled game-based experiment, we analyze how participants adjust their behavior following positive and negative outcomes. We observe asymmetric responses to gains and losses, with stronger corrective adjustments after negative outcomes. Importantly, participants often fail to correctly identify the actions that caused failure and misattribute responsibility across AI agents, leading to systematic revisions of decisions that are weakly related to the underlying causes of poor performance. We refer to this phenomenon as a form of attribution bias, manifested as biased error attribution under delayed feedback. Our findings highlight how cognitive biases can be amplified in human-AI systems with delayed outcomes and multiple autonomous agents, underscoring the need for decision-support systems that better support causal understanding and learning over time.

Figures & tables

Figure 1. The game environment consists of two regions: the blue ally region and red adversary region. Each region contains a carrier at its center that forces must protect while attempting to destroy the opponent’s carrier by launching aircraft. Enemies can enter the ally region at any time, requiring strategic resource allocation for both offensive and defensive operations. The ally side has a total of 5 aircraft available and must decide how many to deploy in each region across 7 time steps. The interface includes a clock and chat box in the bottom right where AI suggestions appear and participants enter their decisions, while scores for both forces are displayed in the top right corner. game window screenshot
Figure 2. Chat interface displaying AI agent suggestions, remaining aircraft count, and current decision step in the game. Chat interface displaying AI agent suggestions, remaining aircraft count, and current decision step in the game.
ConditionOrder AOrder B
Confidence85
No-Confidence85
Table 1. Number of participants per experimental condition. Order A corresponds to the medium-low-high-medium-low-high sequence, and Order B corresponds to the medium-high-low-medium-low-high sequence.
Figure 3. Overview of the experimental procedure. A flowchart illustrating the experimental procedure. The chart shows sequence of study phases, including consent, instructions, questionnaires, practice games, six repeated test game–questionnaire cycles, a final questionnaire, and compensation, connected by arrows indicating the order of occurrence.
Figure 4. Strategy labeling scheme showing overall strategy labels ( SS -0 to SS -5). Strategies are arranged from risk-seeking to increasingly loss-averse along the horizontal axis, with an additional erratic category ( SS -5) representing qualitatively distinct defensive behavior. Within the balanced strategies, offense deployment timing labels ( OO -1 and OO -2) are shown along the vertical axis. Colored regions indicate risk-seeking (red), balanced (blue), loss-averse (green), and erratic (brown) strategy categories. Diagram showing the strategy labeling scheme used in the study. The top section presents overall strategy categories arranged from left to right according to increasing defensive allocation (loss aversion): S-0 (risk-seeking), S-1 to S-3 (balanced), S-4 (loss-averse), and S-5 (erratic). The balanced category is expanded in the lower section to illustrate timing-based subcategories. Across columns (S-1 to S-3), defensive deployment timing ranges from delayed–gradual to immediate–batch. Along the vertical axis, offense deployment timing varies from early (O-1) to late (O-2). Each cell represents a specific combination of defensive and offensive deployment timing.
Figure 5. Histogram of game scores by difficulty level across all games (26 participants, 6 games each). The score distribution for low difficulty is shifted toward positive values, whereas high difficulty is shifted toward negative values. Histogram of game scores by difficulty level across all games (26 participants, 6 games each). The score distribution for low difficulty is shifted toward positive values, whereas high difficulty is shifted toward negative values.
PredictorEst.SE95% CIpp
Intercept (S-4 loss-averse, bad outcome)0.28770.500[-0.69, 1.27].565
S-0 (risk-seeking)1.09860.548[0.03, 2.17].045*
S-1 (balanced allocation & delayed-gradual timing)-0.69310.339[-1.36, -0.03].041*
S-2 (balanced allocation & mixed timing)0.11780.645[-1.15, 1.38].855
S-3 (balanced allocation & immediate-batch timing)1.36100.537[0.31, 2.41].011*
S-5 (erratic)0.27190.627[-0.96, 1.50].664
Table 2. Poisson regression of strategy transition counts by prior strategy and outcome.
Figure 6. Strategy transition matrices following good and bad games. Each matrix shows how participants transitioned between strategy labels (S-0 to S-5) from one game to the next, conditioned on whether the previous game outcome was good or bad. Cell values indicate the number of observed transitions. Two heatmaps showing transitions between defense strategy labels across consecutive games. The left heatmap corresponds to transitions following good games, and the right heatmap corresponds to transitions following bad games. Rows represent the strategy used in the previous game, columns represent the strategy used in the subsequent game, and numbers within cells indicate transition counts.
Figure 7. Density distributions of signed changes in strategy labels between consecutive games. Negative values indicate shifts toward more risk-seeking strategies, whereas positive values indicate shifts toward greater loss aversion. (a) Moderate outcomes (good: score ≥1\geq 1 , bad: score ≤−1\leq-1 ). (b) Extreme outcomes (good: score ≥4\geq 4 , bad: score ≤−4\leq-4 ). Two side-by-side density histograms showing changes in strategy between consecutive games. The horizontal axis shows signed changes in strategy labels, and the vertical axis shows density. Each histogram overlays distributions for good outcomes and bad outcomes. The left histogram corresponds to moderate outcomes, and the right histogram corresponds to extreme outcomes.
Figure 8. Offense strategy transitions when overall strategy remains unchanged, showing frequent shifts from early ( OO -1) to late ( OO -2) offense deployment following bad outcomes. Heatmaps showing how participants transition between offense strategies from one game to the next for good and bad outcomes, when their overall strategy does not change.
PredictorEst.SE95% CIpp
Intercept (O-1 early offense, bad outcome)1.70470.302[1.11, 2.30]<.001
O-2 (late offense)-1.70470.769[-3.21, -0.20].027*
Outcome (Good)-1.01160.584[-2.16, 0.13].083
Strategy ×\times Outcome (Good)    
O-2 (late offense) ×\times Good Outcome2.62100.970[0.72, 4.52].007*
Table 3. Poisson regression of offense timing transitions by prior offense strategy and outcome.
Figure 9. Offense strategy transitions when defensive strategies vary within the balanced range ( SS -1 to SS -3), showing increased shifts from early ( OO -1) to late ( OO -2) offense deployment following bad outcomes. Heatmaps showing transitions between offense strategies across consecutive games for good and bad outcomes, including cases where defensive strategies change within the balanced range (SS-1 to SS-3).
PredictorEst.SE95% CIpp
Intercept (O-1 Early offense, bad)2.07940.250[1.59, 2.57]<.001
O-2 (late offense)-0.98080.479[-1.92, -0.04].040*
Outcome (Good)-0.69310.433[-1.54, 0.16].109
Strategy ×\times Outcome (Good)    
O-2 (late offense) ×\times Good1.73460.643[0.48, 2.99].007*
Table 4. Poisson regression of offense timing transitions when defensive strategies vary within the balanced range ( SS -1 to SS -3).

Explore similar work

CardsList

  1. Responsibility in Multi-Agent Sequential Decision-Making: Comparing Human Judgments to Formal Models of Causal Attribution

    Aug 5, 2026Nripsuta Ani Saxena, Stelios Triantafyllou, Goran RadanovićActual CausalityMulti-Agent Systems

    University of Southern California Los Angeles, California, USA · Max Planck Institute for Software Systems Saarbrücken, Germany

  2. AI-Induced Human Responsibility (AIHR) in AI-Human teams

    Apr 10, 2026Greg Nyilasy, Brock Bastian, Jennifer Overbeck +1AI AccountabilityHuman-AI Decision Making

    Department of Management and Marketing, Faculty of Business and Economics, the University of Melbourne, 198 Berkeley St, Carlton, VIC 3053, Australia · School of Psychological Sciences, Faculty of Medicine, Dentistry, and Health Sciences, the University of Melbourne, Parkville, VIC 3010, Australia · Melbourne Business School, the University of Melbourne, 200 Leicester St, Carlton, VIC 3053, Australia

  3. Human Attribution of Causality to AI Across Agency, Misuse, and Misalignment

    Feb 17, 2026Maria Victoria Carro, David LagnadoCausal AttributionAI Accountability

    Università degli Studi di Genova, IT · FAIR IALAB, University of Buenos Aires, AR · University College London, UK