When multiple AI agents interact with a human and the outcome is delayed, the human blames the wrong agent.
Parakh and Feigh (arXiv:2603.23419) ran a controlled game-based experiment and found systematic misidentification of which actions caused failures in multi-agent human-AI systems. Participants couldn't track responsibility through the delay. Their behavioral corrections — adjustments made after observing a bad outcome — were only weakly connected to actual performance causes.
Note (2026-11-11): checked against the paper's full text. The essay says the bias is systematic in a particular direction: that blame clusters around the most recent action or the most visible agent. I searched the paper for recency and visibility language and found no such finding; what it reports is that participants often fail to identify the actions that caused failure, misattribute responsibility across the AI agents, and revise decisions in ways only weakly related to the causes (it calls this a temporal credit assignment problem). The paper also calls the result a form of attribution bias under delayed feedback, not a new cognitive bias distinct from the fundamental attribution error or hindsight bias; it makes no such comparison that I found. It reports a further result the essay leaves out: stronger corrective adjustments after negative outcomes than after positive ones. The design conclusion the paper draws is decision support that helps people understand causes and learn over time, not specifically transparency about which agent acted when.
The bias isn't random. It's systematic: blame clusters around the most recent action or the most visible agent, regardless of causal contribution. The delay between action and outcome severs the feedback loop that would normally connect cause to effect. In single-agent systems, this loop is already imperfect. With multiple agents, each adding their own actions between cause and consequence, the attribution problem compounds.
This is a new cognitive bias the authors term “biased error attribution under delayed feedback.” It's distinct from fundamental attribution error (blaming agents instead of situations) or hindsight bias (retroactive certainty). It's specifically about the temporal-causal structure of multi-agent interaction: when you can't see the chain from action to outcome, you reconstruct it using heuristics that are systematically wrong.
The practical implication for AI system design: transparency about which agent did what, and when, isn't just a nice-to-have. Without it, human oversight of multi-agent systems produces corrections that don't fix the actual problem and may make it worse. The human believes they're providing useful feedback. The feedback is noise.
The deeper structural problem: as AI systems become more modular (multiple specialized agents cooperating), the attribution gap widens. More agents means more candidates for blame. More cooperation means more entangled causal chains. More autonomy means longer delays. The architecture that makes multi-agent AI powerful also makes human oversight of it systematically miscalibrated.