New research explores how deliberately imperfect outputs may help keep human reviewers vigilant.
Listen to this article:
Modern AI systems increasingly present us with products that appear finished: complete slide decks with images, polished reports and emails, and even full web apps coded entirely by the AI. We therefore move into the role of the reviewer, but as AI outputs have gotten better and better over the years, checking can start to feel ceremonial. Earlier HBS AI Institute research identified the risk of “falling asleep at the wheel” when we approve AI outputs sight unseen. But if better AI can reduce the effort humans put into oversight, can organizations design workflows to keep people alert? The new working paper “The Right Amount of Wrong: Error Injection for AI Oversight,” co-written by HBS AI Associate Michael Lingzhi Li, tackles this question with a counterintuitive idea: intentionally put realistic mistakes into some AI outputs before humans review them.
Key Insight: The Error Injection Theory
“Organizations increasingly rely on human review as a safeguard against AI errors, but reviewer performance is difficult to observe.” [1]
Li and his co-authors identify two management problems. The first is vigilance: when an AI system is right almost all the time, reviewers have fewer reasons to question each new output, making sustained attention harder to maintain. The second is measurement: completing reviews is easy to track, but the accuracy of attention behind those reviews is much harder for managers to see. Controlled error injection is designed to address both problems. An organization can insert realistic mistakes that it knows are wrong, giving reviewers reasons to stay attentive while creating measurable tests of oversight.
To actually put these ideas to the test, the researchers ran a preregistered, double-blind, 12-week experiment with 17 PharmD students performing a simulated pharmacovigilance review task. Each participant reviewed 50 AI-generated summaries per week, compared them to their underlying source reports, and flagged any inaccuracies. Participants rotated every three weeks through four conditions: no injected errors, or low (1 error every 3 summaries), medium (1 per summary), and high (3 per summary). Because every reviewer experienced all four conditions, the researchers could compare how the same individuals behaved under different error environments while controlling for treatment order and carryover effects.
Key Insight: There Really is a “Right Amount of Wrong”
“A natural intuition is that any mechanism that reminds reviewers that AI systems can make mistakes should improve oversight. Our results suggest that this intuition is incomplete.” [2]
The experiment’s centerpiece is whether error injection changed reviewers’ ability to catch naturally occurring AI errors. With no injection, they detected about 24% of natural errors. Under low injection, detection fell to about 20%. Medium injection produced roughly 25%, statistically similar to the control. High injection increased detection to about 31%. Why would a small dose of mistakes make oversight worse? At low levels, the planted mistakes may have imposed an additional burden without fundamentally changing how risky reviewers believed the AI output to be. High injection was the only condition where planted-error prevalence exceeded the natural-error rate, which may have made mistakes salient enough to recalibrate expectations. Reviewers in the medium- and high-injection conditions spent about 16% and 18% longer on each review than the control group. High-frequency injection also significantly reduced trust in the AI system. Yet false-positive (correct pieces of information mistakenly flagged as errors) rates did not significantly rise. Under high injection, reviewers caught more real mistakes, while their tendency to incorrectly flag accurate information remained statistically unchanged. The authors describe this as “calibrated distrust”: in a high-stakes workflow, maximum trust may be less useful than enough skepticism to keep verification active.
Key Insight: Designing Workflows for Real Accountability
“Error injection can serve both as an intervention and as a measurement system.” [3]
Because organizations know which errors they deliberately planted, those errors create ground-truth tests of the reviewer. Across the 17 participants, reviewers who caught more injected errors also tended to catch more naturally occurring ones. Controlled errors could therefore potentially do double duty: changing review behavior while giving managers a signal of whether their oversight is vigilant. However, even under high injection, reviewers caught only about one-third of visible natural AI errors. The authors’ broader implication is that “human in the loop” does not guarantee meaningful supervision. Oversight quality depends on workflow design: how often reviewers encounter mistakes, how costly verification is, and whether performance can be observed. Error injection is just one piece of this puzzle.
Why This Matters
Organizations may need to evaluate a new capability: not just whether people can perform a task (with or without AI), but whether they can responsibly supervise AI performing it. For example, a technical coding interview might give a candidate AI-generated code containing realistic seeded errors and ask them to diagnose, correct, and explain what went wrong. Similar exercises could apply to financial analysts, writing, research, or operations. If oversight quality is invisible, it will drift. Building deliberate, measurable friction into AI-assisted work might be the next step in responsible AI management.
References & Notes
[1] Fang, Kacey, Greg Powell, Jeffery Painter, Andrew Bate, and Michael Lingzhi Li, “The Right Amount of Wrong: Error Injection for AI Oversight” (2026): 3. https://dx.doi.org/10.2139/ssrn.7108438
[2] Fang et al., “The Right Amount of Wrong,” 23.
[3] Fang et al., “The Right Amount of Wrong,” 24.
This article was created through a human-led, AI-assisted process.
Meet the Authors
Kacey Fang is a PhD candidate in health policy at Harvard Medical School.
Greg Powell is Director, Safety Evaluation and Risk Management, at GSK.
Jeffery Painter is Senior Director, Safety & Quantitative Innovation, at GSK.
Andrew Bate is VP, Head of Safety Innovation & Analytics, at GSK.
Michael Lingzhi Li is Assistant Professor of Business Administration and Mary Ellen Jay and Jeffrey Jay Fellow at Harvard Business School and an Associate at the HBS AI Institute.