What the Experiments Establish—and What They Do Not

This chapter evaluates the evidence behind the responsibility model in AI–human teams and AI-supported oversight. The question is not whether the results are interesting, but how far they justify a claim: do they show a robust response to the structure of human–AI control, or only a reaction to particular stories and interfaces? Reading the experiments carefully allows a useful conclusion without turning stated responsibility into proof of actual causation, legal liability, or universal behavior.

What the two research programs actually measure

The first paper reports four experiments about AI-Induced Human Responsibility. Participants read scenarios involving bank loan processing and then allocated responsibility between a human and either an AI system or another human. The studies vary the harm level, the participant’s perspective, and the alternative explanations measured, but they share a common method: written scenarios followed by judgments about responsibility.

That method is well suited to the paper’s central question. A randomized comparison can hold the described task and harmful outcome constant while changing whether the teammate is human or AI. If responsibility judgments differ, the difference is evidence that team composition changes attribution under the scenario’s conditions. The first study’s result—47.81 versus 40.81 on the self-responsibility scale, with d = .24—supports that narrower claim. It does not show that a real employee caused discriminatory lending or that a real bank should assign more blame to an employee partnered with AI.

The distinction matters because a scenario experiment controls ambiguity by withholding information. In the loan vignette, participants are told that the team is responsible, but they are not given a detailed causal reconstruction showing which teammate made which error. That makes the setting useful for studying default attribution. It also means that the result should not be read as a finding about actual fault. The study shows how people fill an evidentiary gap, not what a complete investigation would conclude after inspecting the model, data, interface, policies, and decisions.

The four studies provide more than one observation of the same comparison. Study 1 establishes the initial effect. Study 2 changes the level of harm and also measures perceived autonomy. Study 3 tests whether threat to self-worth explains the result. Study 4 examines self versus other perspective and tests mind perception as an alternative. Across the paper, the reported sample totals 1,801 participants. This convergence makes the proposed pattern more credible within the scenario family than a single vignette would be.

The second paper asks a different question through a different method. In a simulated drone-oversight task, 274 participants monitored critical situations and selected from an interface that allowed six, four, two, or one action. The study measures perceived responsibility of the overseer, the AI system, and the developer, along with perceived causality and knowledge. It therefore provides a controlled test of how changing the human’s available action space affects responsibility judgments and task performance.

The drone design has a useful advantage over a purely written vignette: participants make repeated decisions under time pressure and encounter an outcome that the study ensures they experience. It also has clear limitations.