AI Insight
This randomized experiment with 300 adults examined how people assign responsibility to humans versus AI systems in legal decision-making scenarios with adverse outcomes. As AI involvement increased (from low to moderate to high functionality and authority), participants attributed significantly less responsibility to human reviewers and more to the AI system. The shift in perceived human responsibility was associated with decreased psychological ownership and sense of control over the decision process.
Why it matters
These findings are relevant for designing accountability frameworks in AI-assisted legal and administrative systems, where unclear responsibility attribution could undermine justice and oversight. Understanding how people naturally assign blame in human-AI collaborative decisions may inform policy on human oversight requirements and liability structures in high-stakes automated decision-making.
Understand the Science
BackgroundArtificial intelligence (AI) is increasingly used in consequential decisions, including legal and administrative judgments. When an AI-assisted decision has an adverse outcome, observers may find it difficult to determine how responsibility should be attributed to the human and AI actors.ObjectiveThis randomized vignette experiment examined perceived responsibility attribution across three bundled AI-assisted legal-decision scenarios. The scenarios varied AI functionality, human decision authority, and opportunities to intervene together. Post-vignette psychological ownership and perceived control were examined as condition-proximal perceptions associated with responsibility judgments.MethodsThree hundred adults were retained for analysis (100 per condition). Each participant read one vignette describing an already completed legal decision and assessed the perceived responsibility of a fictional human reviewer and an AI system after an adverse outcome. Participants did not make a sentencing decision or interact with a functioning AI system. Randomized group contrasts were the principal comparisons. Secondary, exploratory mediation models used two indicator variables, with the low-AI scenario as the reference; post-vignette AI trust was excluded from the primary exploratory specification. Parallel and reverse-order models were examined as sensitivity analyses.ResultsThe manipulation check differed across conditions, F(2, 297) = 1065.77, p < 0.001, η2 = 0.878. Perceived human responsibility declined from the low- to the moderate- and high-AI scenarios, whereas perceived AI responsibility increased. The randomized mean differences in perceived human responsibility were −1.12 (moderate versus low) and −1.69 (high versus low). Exploratory indirect associations were negative via psychological ownership (B = −0.385 and −0.713), perceived control (B = −0.146 and −0.168), and the proposed ownership-to-control sequence (B = −0.190 and −0.353), respectively; corresponding confidence intervals are reported in the Results. The indirect-effect intervals were calculated using delta-method approximations.ConclusionThe randomized scenario comparisons showed differences in perceived responsibility attribution. Because the scenarios jointly varied multiple features, these findings do not isolate the causal effects of AI functionality, human authority, or opportunities to intervene. Ownership and control ratings partly reflected information supplied in the vignettes and therefore cannot establish distinct causal mediating mechanisms or their temporal sequence.