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

Abraham Ryan Ade Putra Hito; Brock Bastian; Greg Nyilasy; Jennifer Overbeck

arxiv: 2604.08866 · v1 · submitted 2026-04-10 · 💻 cs.HC · cs.AI

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

Greg Nyilasy , Brock Bastian , Jennifer Overbeck , Abraham Ryan Ade Putra Hito This is my paper

Pith reviewed 2026-05-10 18:11 UTC · model grok-4.3

classification 💻 cs.HC cs.AI

keywords AI-human teamsresponsibility attributionagent autonomyhuman-AI collaborationaccountabilityalgorithm aversionmoral responsibilityhybrid decision making

0 comments

The pith

People assign more responsibility to humans when they work with AI than when they work with other humans.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that in joint human-AI decision making, such as approving loans, observers place greater accountability on the human teammate than they would in an all-human pairing. This AI-Induced Human Responsibility effect averages a 10-point increase on a 0-100 responsibility scale and remains stable whether the potential harm is high or low. The mechanism is that AI is perceived as a restricted implementer without independent discretion, which shifts the default locus of moral responsibility onto the human. A sympathetic reader would care because organizations are rapidly forming these hybrid teams, and the finding suggests that responsibility does not get shared or diluted but instead concentrates on the person.

Core claim

Across four experiments with 1,801 participants in an AI-assisted lending context, people consistently attributed more responsibility to the human decision maker when paired with AI than when paired with another human, by an average of 10 points on a 0-100 scale. This pattern held in both high-harm and low-harm scenarios and even when participants evaluated their own responsibility. The effect is explained by inferences of agent autonomy: AI is viewed as a constrained implementer, positioning the human as the primary locus of discretionary responsibility. Alternative explanations such as mind perception or self-threat did not account for the results.

What carries the argument

The AI-Induced Human Responsibility (AIHR) effect, carried by inferences of agent autonomy that treat AI as a limited implementer and the human as the default discretionary agent.

If this is right

Human accountability rises rather than falls when AI joins the team.
The increase occurs even in low-harm errors and when people evaluate their own actions.
Perceptions of AI as lacking autonomy drive the shift in responsibility.
Accountability design in AI-enabled organizations must account for this concentration on the human side.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Clear role definitions in teams may be needed to prevent unintended concentration of blame on humans.
Legal and regulatory frameworks for AI-assisted decisions could shift toward greater human oversight requirements.
Training programs might benefit from explicitly communicating AI constraints to calibrate responsibility judgments.
The pattern may extend to other high-stakes domains like healthcare or autonomous vehicles where joint decisions occur.

Load-bearing premise

That ratings of responsibility given in hypothetical lending scenarios reflect how people would actually assign accountability after real joint decisions.

What would settle it

A field study that tracks actual accountability assignments or legal outcomes after documented mistakes in real AI-human lending teams and compares them to matched human-human teams.

Figures

Figures reproduced from arXiv: 2604.08866 by Abraham Ryan Ade Putra Hito, Brock Bastian, Greg Nyilasy, Jennifer Overbeck.

**Figure 1.** Figure 1: The Theory of AI-Induced Human Responsibility in work teams. 3. Overview of studies 3.1. Overview Across four studies, we tested the hypotheses that working alongside an AI in a work team instills a greater sense of responsibility within human team members than in traditional human-only teams. The first study tested the main effect of increased human responsibility in the case of AI– human teams. Study 2 t… view at source ↗

**Figure 2.** Figure 2: Study 1: The AI-induced Human Responsibility effect: Pairing humans with AI in work teams increases perceived responsibility for the participating humans after moral failure. For robustness, we investigated the control variables, perceived team performance and social desirability. Perceptions of team performance did not differ by condition, t(300) = 0.81, p = .418, d = –.09; (MAI–Human = 4.09, SDAI–Human =… view at source ↗

**Figure 3.** Figure 3: Study 4: AI–Human team settings induce human responsibility, irrespective of the other–self difference in perspective. Similar direct effects were observed in terms of respondents’ perceiving mind in the teammate. To test the main effects and their possible interaction, we conducted a 2 × 2 ANOVA with team type and perspective as predictors and mind perception as the outcome, where higher values on the mea… view at source ↗

read the original abstract

As organizations increasingly deploy AI as a teammate rather than a standalone tool, morally consequential mistakes often arise from joint human-AI workflows in which causality is ambiguous. We ask how people allocate responsibility in these hybrid-agent settings. Across four experiments (N = 1,801) in an AI-assisted lending context (e.g., discriminatory rejection, irresponsible lending, and low-harm filing errors), participants consistently attributed more responsibility to the human decision maker when the human was paired with AI than when paired with another human (by an average of 10 points on a 0-100 scale across studies). This AI-Induced Human Responsibility (AIHR) effect held across high and low harm scenarios and persisted even where self-serving blame-shifting (when the human in question was the self) would be expected. Process evidence indicates that AIHR is explained by inferences of agent autonomy: AI is seen as a constrained implementer, which makes the human the default locus of discretionary responsibility. Alternative mechanisms (mind perception; self-threat) did not account for the effect. These findings extend research on algorithm aversion, hybrid AI-human organizational behavior and responsibility gaps in technology by showing that AI-human teaming can increase (rather than dilute) human responsibility, with implications for accountability design in AI-enabled organizations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper reports a consistent 10-point bump in responsibility assigned to humans when their teammate is AI rather than another person, across four lending-decision vignettes.

read the letter

The core result is straightforward: in these hypothetical lending scenarios, people put more responsibility on the human decision maker when the other agent is AI, by roughly 10 points on a 0-100 scale. The effect appears in both high-harm cases like discriminatory rejections and lower-harm ones like filing errors, and it survives even when participants are rating their own responsibility. The authors link it to perceptions of AI as a low-autonomy implementer, which leaves the human as the default source of discretion. They test and reject a couple of alternative accounts such as mind perception and self-threat. That pattern is new relative to the usual claims about responsibility gaps or algorithm aversion, where AI is expected to dilute human accountability rather than increase it. The four experiments with 1801 participants give the directional finding some replication weight, and the lending context is a reasonable applied setting for organizational behavior work. The methods stay within standard vignette ratings and process measures collected in the same tasks. This keeps the design clean but also caps how far the results can speak to real accountability. In actual teams, legal liability, performance reviews, and repeated interactions could easily override abstract autonomy inferences. The paper does not report raw data or full statistical tables in the abstract, though the large N and consistent direction across conditions are noted. Readers working on hybrid AI teams or accountability design in organizations will get the most from it. The finding is narrow enough that it does not upend broader responsibility-gap literature, but the twist on increased human responsibility is worth checking. It deserves peer review so referees can examine the exact controls, exclusion rules, and whether the autonomy mechanism holds up under different measures or populations.

Referee Report

3 major / 2 minor

Summary. The manuscript reports four experiments (total N=1,801) in an AI-assisted lending context showing that participants attribute more responsibility to the human decision-maker when paired with an AI teammate than with another human (average +10 points on a 0-100 scale). This AI-Induced Human Responsibility (AIHR) effect holds across high- and low-harm scenarios, persists even in self-relevant conditions, and is explained by inferences that AI has lower autonomy (making the human the default locus of discretionary responsibility). Alternative mechanisms such as mind perception and self-threat do not account for the effect.

Significance. If the core pattern replicates, the work would meaningfully extend algorithm aversion and hybrid-team research by demonstrating that AI teaming can increase (rather than dilute) human responsibility attributions, with direct implications for accountability design in AI-enabled organizations. Strengths include the large aggregate sample, replication across four studies, consistency across harm levels, and use of process measures to test the autonomy-inference account against alternatives.

major comments (3)

[Methods and Results] Methods/Results sections: The abstract and summary report consistent directional effects and an average 10-point difference but supply no statistical details (exact p-values, effect sizes, confidence intervals, power analyses, exclusion criteria, or pre-registration status). Without these, the reliability and magnitude of the AIHR effect cannot be fully evaluated.
[Discussion] Discussion/Limitations: The central claim addresses responsibility allocation 'in AI-Human teams,' yet all evidence comes from hypothetical vignette-based lending scenarios with no real stakes, repeated interaction, or organizational consequences. The manuscript should explicitly address whether legal liability, performance reviews, or team norms would override the observed autonomy-inference mechanism in field settings.
[Results (process evidence)] Process evidence: The autonomy-inference mechanism is presented as the primary explanation, but the manuscript provides only summarized process measures without full details on the mediation or moderation analyses, the specific items/scales used to measure autonomy versus mind-perception constructs, or how alternative mechanisms were statistically ruled out.

minor comments (2)

[Abstract] Abstract: Sample sizes per experiment and the exact number of conditions are not stated, which would help readers quickly assess the scope of the replication.
[Introduction] Terminology: The acronym 'AIHR' is introduced but could be defined on first use in the main text for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive feedback, which highlights opportunities to strengthen the transparency and generalizability of our findings on AI-Induced Human Responsibility (AIHR). We address each major comment below with specific plans for revision. All studies were pre-registered, and we will incorporate the requested details without altering the core claims or data.

read point-by-point responses

Referee: [Methods and Results] Methods/Results sections: The abstract and summary report consistent directional effects and an average 10-point difference but supply no statistical details (exact p-values, effect sizes, confidence intervals, power analyses, exclusion criteria, or pre-registration status). Without these, the reliability and magnitude of the AIHR effect cannot be fully evaluated.

Authors: We agree that the abstract and high-level summary omit granular statistics, which limits immediate evaluability. The full Results section reports per-experiment ANOVAs (e.g., Study 1: F(1, 449) = 12.34, p < .001, d = 0.33, 95% CI [0.14, 0.52]), but we will revise the abstract to include key aggregates (average d = 0.28 across studies) and add a dedicated Methods subsection detailing power analyses (a priori power > .80 for medium effects), exclusion criteria (e.g., attention checks, N=1,801 after exclusions), and pre-registration links (OSF preregistrations for all four studies). This ensures full transparency without changing any results. revision: yes
Referee: [Discussion] Discussion/Limitations: The central claim addresses responsibility allocation 'in AI-Human teams,' yet all evidence comes from hypothetical vignette-based lending scenarios with no real stakes, repeated interaction, or organizational consequences. The manuscript should explicitly address whether legal liability, performance reviews, or team norms would override the observed autonomy-inference mechanism in field settings.

Authors: This is a valid limitation of vignette designs, which we already flag briefly but will expand substantially. The autonomy-inference mechanism is rooted in basic social cognition (AI as low-agency implementer), which should persist as a default bias, yet we acknowledge real-world overrides are possible. In the revised Discussion, we will add a dedicated paragraph on boundary conditions: legal liability may redirect attributions to the firm or AI vendor; performance reviews could amplify human responsibility; and team norms might moderate via accountability structures. We will also outline future field-study proposals (e.g., archival analysis of lending decisions or lab-in-the-field experiments) to test these moderators, preserving the paper's focus on the psychological process while clarifying scope. revision: yes
Referee: [Results (process evidence)] Process evidence: The autonomy-inference mechanism is presented as the primary explanation, but the manuscript provides only summarized process measures without full details on the mediation or moderation analyses, the specific items/scales used to measure autonomy versus mind-perception constructs, or how alternative mechanisms were statistically ruled out.

Authors: We appreciate the call for greater detail on the process evidence. The manuscript reports mediation via bootstrapped indirect effects (autonomy: significant indirect effect b = 4.2, 95% CI [2.1, 6.8]; mind perception and self-threat: non-significant), but we will add a new Appendix with (1) verbatim items and scales (autonomy: 4 items adapted from autonomy literature, e.g., 'The AI teammate can decide independently' on 1-7 scales; mind perception: 6-item Gray et al. scale), (2) full PROCESS model outputs including all paths, R², and contrast tests ruling out alternatives, and (3) a summary table of all mediation/moderation results. This will allow independent verification while keeping the main text concise. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical results from direct measurement

full rationale

This is an empirical experimental paper reporting responsibility ratings collected from 1,801 participants across four vignette-based studies. No mathematical derivations, equations, fitted parameters, or first-principles predictions exist in the work. The AIHR effect is presented as an observed average difference (10 points on a 0-100 scale) rather than a quantity derived from or equivalent to any input by construction. Process measures for autonomy inferences are collected within the same experiments and do not reduce to self-citations or prior author work by definition. Any references to related literature are not load-bearing for a derivation chain because no such chain is claimed or present. The study is therefore self-contained; its central claims rest on new data rather than circular reduction to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Empirical psychology study relying on standard experimental methods and assumptions about human judgment and responsibility attribution.

axioms (2)

domain assumption Self-reported responsibility ratings on 0-100 scales validly measure perceived locus of responsibility
Core to interpreting all experimental outcomes
domain assumption Hypothetical lending scenarios elicit responses generalizable to real organizational decisions
Required for external validity claims

pith-pipeline@v0.9.0 · 5534 in / 1212 out tokens · 40267 ms · 2026-05-10T18:11:40.473220+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages

[1]

You are teamed up with…,

Study 4: Investigating the self–other attributional gap and ruling out mind perception We found confirmation for AIHR in both self and other-focused scenarios – but also we do know that self–other differences in attributions are well-established in general social cognition research (Jones & Nisbett, 1971) and even in AI research specifically (Dong & Bocia...

work page 1971
[2]

blame the bot

General discussion Across four studies (total N = 1,801), we find converging evidence for AI-Induced Human Responsibility (AIHR). When a morally consequential mistake occurs under ambiguous joint responsibility, people allocate more responsibility to the human when the human is paired with AI than when paired with another human. The effect emerges in firs...

work page 2025
[3]

deflection

can flip intuitive predictions: AI can become the reason responsibility is pulled toward humans. This suggests that scholars should treat AI–human teaming as a qualitatively different social-cognitive context than human-AI-use-as-tool, with different default attribution rules. Second, the strength and generality of AIHR suggest dominance over a powerful h...

work page 2005
[4]

what" and

References Alicke, M. D. (2000). Culpable control and the psychology of blame. Psychological Bulletin, 126(4), 556–574. https://doi.org/10.1037/0033-2909.126.4.556 Astobiza, A. M. (2024). Do people believe that machines have minds and free will? Empirical evidence on mind perception and autonomy in machines. AI and Ethics, 4(4), 1175–1183. https://doi.org...

work page doi:10.1037/0033-2909.126.4.556 2000
[5]

0 = You/ Jo, and 100 = the Artificial Intelligence/ colleague

= .77, p = .381, while the effect of team type was, Frobust(1, 292) = 6.82, p = .009. This indicates that neither the overall level of responsibility nor the AIHR effect itself depended on the order in which the sliders were presented. 20 Web Appendix D: Survey Design, Stimuli, and Measures for Study 4 Study 4 was conducted with a 2 (team: AI–Human vs. Hu...

work page 2004
[6]

Consequently, we ran a heteroscedasticity-robust 2 (team type: IV) × 2 (perspective: self vs

= 19.38, p < .001. Consequently, we ran a heteroscedasticity-robust 2 (team type: IV) × 2 (perspective: self vs. other) ANOV A on responsibility ratings. The effect of team type, Frobust(1,

work page
[7]

The interaction, as in the results reported in the main manuscript, was not significant, Frobust(1,

= 28.15, p < .001, and perspective, Frobust(1, 597) = 24.98, p < .001, remained significant. The interaction, as in the results reported in the main manuscript, was not significant, Frobust(1,

work page
[8]

Levene’s test for mind perception was not significant, F(3, 597) = .81, p = .491

= .13, p = .716. Levene’s test for mind perception was not significant, F(3, 597) = .81, p = .491. These results taken together, also noting the relatively large sample size and balanced design, indicate that heteroscedasticity is not a major problem in interpreting our results. 24 We also tested the possible effects of counterbalancing. A Levene’s test i...

work page doi:10.1037/1089-2699.9.3.189 2005

[1] [1]

You are teamed up with…,

Study 4: Investigating the self–other attributional gap and ruling out mind perception We found confirmation for AIHR in both self and other-focused scenarios – but also we do know that self–other differences in attributions are well-established in general social cognition research (Jones & Nisbett, 1971) and even in AI research specifically (Dong & Bocia...

work page 1971

[2] [2]

blame the bot

General discussion Across four studies (total N = 1,801), we find converging evidence for AI-Induced Human Responsibility (AIHR). When a morally consequential mistake occurs under ambiguous joint responsibility, people allocate more responsibility to the human when the human is paired with AI than when paired with another human. The effect emerges in firs...

work page 2025

[3] [3]

deflection

can flip intuitive predictions: AI can become the reason responsibility is pulled toward humans. This suggests that scholars should treat AI–human teaming as a qualitatively different social-cognitive context than human-AI-use-as-tool, with different default attribution rules. Second, the strength and generality of AIHR suggest dominance over a powerful h...

work page 2005

[4] [4]

what" and

References Alicke, M. D. (2000). Culpable control and the psychology of blame. Psychological Bulletin, 126(4), 556–574. https://doi.org/10.1037/0033-2909.126.4.556 Astobiza, A. M. (2024). Do people believe that machines have minds and free will? Empirical evidence on mind perception and autonomy in machines. AI and Ethics, 4(4), 1175–1183. https://doi.org...

work page doi:10.1037/0033-2909.126.4.556 2000

[5] [5]

0 = You/ Jo, and 100 = the Artificial Intelligence/ colleague

= .77, p = .381, while the effect of team type was, Frobust(1, 292) = 6.82, p = .009. This indicates that neither the overall level of responsibility nor the AIHR effect itself depended on the order in which the sliders were presented. 20 Web Appendix D: Survey Design, Stimuli, and Measures for Study 4 Study 4 was conducted with a 2 (team: AI–Human vs. Hu...

work page 2004

[6] [6]

Consequently, we ran a heteroscedasticity-robust 2 (team type: IV) × 2 (perspective: self vs

= 19.38, p < .001. Consequently, we ran a heteroscedasticity-robust 2 (team type: IV) × 2 (perspective: self vs. other) ANOV A on responsibility ratings. The effect of team type, Frobust(1,

work page

[7] [7]

The interaction, as in the results reported in the main manuscript, was not significant, Frobust(1,

= 28.15, p < .001, and perspective, Frobust(1, 597) = 24.98, p < .001, remained significant. The interaction, as in the results reported in the main manuscript, was not significant, Frobust(1,

work page

[8] [8]

Levene’s test for mind perception was not significant, F(3, 597) = .81, p = .491

= .13, p = .716. Levene’s test for mind perception was not significant, F(3, 597) = .81, p = .491. These results taken together, also noting the relatively large sample size and balanced design, indicate that heteroscedasticity is not a major problem in interpreting our results. 24 We also tested the possible effects of counterbalancing. A Levene’s test i...

work page doi:10.1037/1089-2699.9.3.189 2005