A recent Management Science study shows that the biggest gains from AI come not simply from model accuracy but from workers who are well calibrated about their own abilities. Organizations that treat AI adoption as a mainly technical problem—benchmarking models, buying licenses and redesigning workflows—risk modest returns if they ignore whether employees know when to defer to the tool and when to trust themselves.
The Management Science experiment tested 732 people on a task: deciding whether faces in photos were over age 21, sometimes with an AI confidence score and sometimes without. Average performance rose with AI assistance, and lower-ability participants improved more than stronger ones, consistent with evidence that AI can act as a leveling technology. The paper separates raw ability from self-knowledge: two workers with the same baseline skill can gain very different value from the same system because one is better at recognizing when the system is likely to outperform them.
AI adoption and worker calibration
Miscalibration hurts in two directions. Overconfident workers ignore useful AI advice; underconfident workers defer when their own judgment would have been superior. In both cases, the potential value of augmentation leaks away: companies can install a capable system and still see mediocre gains because the bottleneck is users’ sense of their own competence, not the model itself.
The Management Science findings complement other recent research. A widely discussed Science paper on generative AI and professional writing found that ChatGPT meaningfully increased productivity and quality, with the largest gains for weaker initial performers. A large field experiment published in the Quarterly Journal of Economics showed AI assistance raised customer-support productivity by 15% on average and helped novice and low-skilled workers far more than top performers. An NBER paper argues that AI could broaden access to expertise and help restore middle-skill work; the Management Science study supports that potential but adds a critical condition: calibration matters.
When people use AI with their actual, imperfect self-beliefs, inequality in performance falls. With perfect calibration, the study notes, inequality would fall even more. In short, AI can narrow performance gaps, but it does not do so automatically—human judgment remains a gatekeeper.
Train people to work with AI, not just to use it
That insight suggests a shift in training priorities. Rather than focusing solely on technical instruction, certification and prompt training, organizations should teach workers to estimate uncertainty, read model confidence signals, recognize edge cases and adjust behavior based on feedback. The goal is disciplined partnership with AI, not blind trust.
There is reason to believe calibration can improve. A study in Futures & Foresight Science found that an interactive training app reduced overconfidence and improved calibration in under 30 minutes. Decades earlier, Sarah Lichtenstein and Baruch Fischhoff showed that feedback-based calibration training can make people better judges of their own probabilities. More recent work on automated calibration training for forecasters reaches a similar conclusion: calibration is not immutable.
Practical implications are straightforward. A call center, insurer, hospital or legal team may not be able to turn average employees into experts overnight, but it can make them much better at recognizing when the machine is likely right and when skepticism is warranted. That focused training amplifies the real mechanism by which AI acts as a force multiplier: rewarding workers who can distinguish confidence from competence.
Adapted from The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). Dr. Gleb Tsipursky, described by The New York Times as the “Office Whisperer,” is CEO of the AI consultancy Disaster Avoidance Experts and the author of eight books, including The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).
