Eyes Reveal When Supervisors Drift Out of the Loop

Modern automation was supposed to make human operators obsolete. Instead, it has quietly transformed them into passive supervisors of systems so reliable that they almost never need a human touch. That shift has a hidden cost. When operators spend long stretches watching an automated system that rarely fails, their performance degrades precisely when it matters most: the moment the automation surprises them with an unexpected failure. Researchers call this the out-of-the-loop performance problem, and despite decades of investigation, it has remained stubbornly difficult to characterize, quantify, and predict. A new study published in Frontiers in Neuroergonomics now offers something the field has lacked: objective, eye-based markers that can separate two distinct mental states—task disengagement and declining vigilance—both of which push operators out of the loop.

The research, conducted by Stefania C. Ficarella of ONERA and Aix-Marseille Université, Andrea Desantis of ONERA, and colleagues in collaboration with Airbus Operations, tackled a question that has long confounded human-factors scientists. Vigilance, the ability to stay alert and sustain attention over time, is known to decline as time-on-task increases. Task disengagement, the reduced allocation of processing resources to the task at hand, emerges when operators are not frequently solicited. Both contribute to out-of-the-loop effects, but in most real-world settings the two are confounded. A pilot who has been flying for six hours is simultaneously fatigued and under-stimulated. Which factor is driving the degradation? Until now, no objective metric could tell them apart.

To disentangle the two, the team designed an elegant supervisory paradigm. Twelve healthy volunteers with normal or corrected-to-normal vision watched a simplified aircraft interface displaying four flight parameters: altitude, speed, flap position, and heading. Their job was to monitor an autopilot and validate or reject its decisions. The rule was deceptively simple. Whenever a parameter deviated by more than 20 units (or 1 unit for flaps), participants compared the direction of that change to the most recent significant change they had observed. Same direction, regardless of parameter, meant the autopilot was correct; opposite direction meant it was wrong. Participants pressed designated arrow keys to report their judgment. Background noise—small sinusoidal fluctuations in all four parameters—made detection genuinely challenging, and the central display remained fixed to encourage eye movements.

The critical manipulation was temporal. The researchers varied the time delay between significant parameter changes across fifteen intervals ranging from 8 to 92 seconds, with the frequency of each interval following an exponential function so that short delays appeared most often. The logic was that the longer an operator waits before taking a relevant action, the more likely disengagement becomes—through task withdrawal or mind-wandering. Vigilance, by contrast, was manipulated through block repetition: the same task was divided into ten blocks, with the first block corresponding to the first 15 responses and the last to the final 15. The full session lasted roughly 90 minutes. Throughout, a Smart Eye Pro 3.0 system with two infrared illuminators and three cameras sampled gaze at 120 Hz, recording fixations, saccades, blinks, and eyelid opening.

The behavioral results confirmed the performance cost of disengagement. Using hierarchical Bayesian mixed-effects models, the team found strong evidence that reaction times increased with longer delays: the longer participants had to wait before intervening, the slower they responded when the moment finally arrived. Accuracy, by contrast, remained high overall—around 90 percent—and showed no credible effect of delay or block. Correct responses were slightly slower than incorrect ones, consistent with a small speed–accuracy trade-off in which more careful judgments took a bit longer. The performance problem, in other words, was not about making mistakes. It was about losing the readiness to act quickly.

The eye-tracking data revealed something far more interesting: two dissociable oculomotor signatures. Under disengagement—longer delays—the number of saccades decreased, the number of fixations decreased, and fixation durations increased. Visual exploration of the interface shrank; participants scanned less and dwelled longer at each location, a pattern the authors suggest is consistent with episodes of mind-wandering or attentional drift during extended intervals without stimulation. Under declining vigilance—as blocks progressed—the picture changed. Blink frequency increased, blink durations increased, and eyelid opening decreased, all classic markers of reduced arousal and fatigue. Fixation durations, remarkably, moved in opposite directions for the two states: they lengthened with increasing delay but shortened across blocks. The number of saccades and fixations remained stable across blocks, indicating that vigilance decline did not reduce visual exploration per se but altered how sustained visual processing was maintained.

This dissociation matters because it suggests different underlying mechanisms. Longer delays between task-relevant events may encourage transient disengagement from active monitoring, reducing motor and cognitive readiness so that participants are less prepared to detect and evaluate system changes when they occur. The shortening of fixation durations across blocks, occurring alongside increased blinking and narrowed eyelids, reflects less sustained visual processing rather than reduced scanning. The authors note that the mid-experiment break may have partially restored vigilance, attenuating time-on-task effects on behavior—yet clear oculomotor patterns associated with time spent on task still emerged. They also acknowledge that fatigue or boredom cannot be fully excluded over a 90-minute session, though a slight decrease in reaction times across blocks argues against boredom or fatigue being the dominant factor, since those would be expected to slow responses as time-on-task increases.

The most forward-looking result came from the classification analyses. The team trained logistic regression classifiers to predict, from oculomotor features alone, whether a participant would respond slowly or quickly on each upcoming trial. Features included the number and duration of saccades, fixations, and blinks, plus pupil size and eyelid opening, computed across three time windows preceding each parameter change. Using Monte Carlo cross-validation repeated 1,000 times, with training on 80 percent of trials and matched numbers of fast and slow responses, the classifiers performed reliably above chance. The highest classification accuracy reached 56 percent, with area-under-the-curve values between 0.54 and 0.58 across time windows, and the entire time cluster was statistically significant. Neighborhood component analysis identified the three most informative features: the number of saccades, the duration of fixations, and the duration of blinks.

Those numbers are modest, but their implications are substantial. They demonstrate that an operator’s behavioral state—specifically, whether they are about to react sluggishly—can be anticipated from eye movements recorded seconds before the critical event. In a cockpit, that means an eye-tracking system could in principle flag a pilot who is drifting out of the loop before the autopilot fails, opening the door to adaptive interfaces that re-engage attention proactively rather than reactively. The authors emphasize that the findings are relevant not only for monitoring piloting activities but also for the broader science of how task disengagement shapes human cognition.

Limitations remain. The sample of twelve participants is small, which may constrain generalizability, and the laboratory paradigm cannot fully capture the vigilance and engagement decrements pilots experience during long flights. The authors call for future studies to replicate these effects in larger and more diverse samples, including pilots during flight simulations, and note that the limited ecological validity of the cognitive task requires validation using cockpit simulators. Still, the study delivers what the field has needed: proof that disengagement and vigilance decline, long conflated in the out-of-the-loop literature, leave distinct fingerprints in the eyes—and that those fingerprints can predict performance before it falters.

Subject of Research: Distinct oculomotor signatures for task disengagement and reduction in vigilance during a supervisory task

Article Title: Distinct oculomotor signatures for task disengagement and reduction in vigilance during a supervisory task

Article References: Ficarella, S. C., Maille, N., Lantos, N., Le Goff, K., Sciabica, J.-F., Sarrazin, J.-C., & Desantis, A. (2026). Distinct oculomotor signatures for task disengagement and reduction in vigilance during a supervisory task. Frontiers in Neuroergonomics, 7, Article 1706491. https://doi.org/10.3389/fnrgo.2026.1706491

Image Credits: AI Generated

DOI: 10.3389/fnrgo.2026.1706491

Keywords: Distinct, oculomotor, signatures, task, disengagement, reduction, vigilance, during, supervisory, scientific research