AI Errors Rise When Workers Stop Feeling Responsible

An August 2026 Decision Support Systems study by Richard Henkenjohann and Manuel Trenz reports experimental evidence that workers made fewer errors in generative-AI-assisted work when they experienced stronger responsibility for the outcome. Its online laboratory experiment also found that human-first workflows increased felt responsibility but reduced subsequent co-creation participation, while higher-performing AI encouraged more interaction but diffused responsibility.

The direct answer for teams is that assigning a human reviewer is not enough. Under the study’s conditions, errors increased as felt responsibility weakened, but neither starting with human input nor increasing interaction alone preserved both ownership and active engagement. Workflows need to give people meaningful influence over the output and an explicit final decision.

Responsibility follows perceptual and behavioral routes

The researchers distinguish experienced responsibility from co-creation participation. Experienced responsibility describes whether workers feel answerable for the result. Participation describes how intensively they interact with the generative system while producing or revising that result.

These concepts represent two routes through which responsibility can develop. A workflow can create a perceptual sense of ownership by making the worker the originator of the task or its initial content. It can also create a behavioral connection through substantive revisions and repeated interaction with the shared work product.

The distinction explains why interface activity is an incomplete measure of human control. Someone can make many edits while regarding the AI proposal as the default decision. Conversely, a worker can feel ownership because the process began with their contribution, yet become relatively inactive during later stages when errors could still enter the work.

Human-first design creates a measurable trade-off

A human-first AI workflow establishes early ownership but shows reduced participation during later co-creation.

Starting with a human contribution strengthened experienced responsibility in the experiment. The worker was not simply handed a completed machine output for approval, so the result remained more closely connected to their initial assumptions and choices.

However, human-first configuration also reduced the intensity of later co-creation participation. Early authorship therefore acted as an ownership intervention, not proof that the person remained engaged throughout the workflow. A process can begin with meaningful human input and still become passive during consequential revisions.

AI-first work creates the complementary risk. An initial machine-generated answer may define the frame within which the worker reacts, particularly when the output appears capable. A May 2026 report on the research likewise describes the study’s competing perceptual and behavioral pathways and identifies 222 participants in its experiment.

The error finding supports the headline—with limits

The study’s central downstream result was that stronger experienced responsibility mitigated errors in work outcomes. Expressed inversely, errors rose when workers felt less responsible. That relationship supports the headline within the experiment, but it is not a universal error-rate estimate for every model, occupation or organization.

The evidence comes from an online laboratory experiment rather than a long-running deployment inside operating companies. Controlled experiments can isolate workflow configuration, AI performance, participation and felt responsibility. Live workplaces add deadlines, incentives, hierarchy, uneven expertise and automation bias, any of which could change the size of the observed effects.

The paper also does not establish that workers should carry unlimited personal blame for system failures. Experienced ownership may help someone catch errors, but responsibility must be matched with access to relevant evidence, authority to alter the output and permission to reject or escalate it.

Outside observers may assign even more responsibility

A separate 2026 responsibility-attribution preprint reports four experiments with 1,801 participants in AI-assisted lending scenarios. Across the studies, participants assigned an average of 10 additional responsibility points on a 0–100 scale to a human paired with AI compared with a human paired with another person; the authors attribute the effect to perceptions of autonomy, with AI viewed as a constrained implementer and the human as the actor exercising discretion.

That project examined blame allocation after mistakes, not how co-creation design changes errors, so its four experiments should not be treated as part of the Henkenjohann–Trenz study. Together, however, the projects reveal a possible organizational mismatch: outsiders may hold an employee especially responsible for an AI-assisted failure even when the internal process has reduced that employee’s felt ownership or practical ability to redirect the work.

Formal accountability can therefore exceed actual decision power. This matters when organizations name a person as the final reviewer but give that person limited time, incomplete evidence or no authority to stop the output from moving forward.

Auditable handoffs need real decision rights

An auditable human–AI handoff records review, correction and the worker’s final decision.

The practical unit for applying the evidence is the handoff, not a generic “human in the loop” label. A consequential workflow should make the person’s contribution, verification work and final authority observable instead of treating the presence of a reviewer as proof of control.

An auditable handoff can record:

the human owner and the outcome for which that person is responsible;
the assumptions, constraints and source material supplied before generation;
the material AI contributions and the worker’s substantive revisions;
the checks performed on claims, calculations or other foreseeable failure points;
the worker’s authority to accept, revise, reject or escalate the result;
the final decision and its rationale.

This is an editorial application of the findings, not a checklist validated by the experiment. Its purpose is to make both routes to responsibility observable: the worker’s ownership of the outcome and their behavioral participation in producing it. A timestamp showing that someone clicked “approve” would not demonstrate either one.

The pattern also prevents responsibility from becoming ceremonial. If the assigned owner cannot inspect inputs, challenge the system’s frame or stop the handoff, the record should expose that limitation instead of implying meaningful oversight.

Field evidence is still needed

The journal study establishes experimental evidence that felt responsibility can reduce errors and that workflow choices affect responsibility and participation differently. It does not determine which interface features, checkpoints or permission structures preserve both mechanisms during sustained workplace use.

The next evidence needs to connect actual error outcomes with interaction histories, decision authority and workers’ reported responsibility over time. Until field studies provide that comparison, the defensible conclusion is narrower: a named reviewer is not equivalent to a worker who actively shapes the output, controls the handoff and experiences the final result as their responsibility.