What is the vigilance decrement?

The vigilance decrement is the fall in detection accuracy that happens when somebody watches for rare signals over a long period. Norman Mackworth measured it in 1948 with a clock whose hand occasionally jumped twice, and it has held up across seventy-five years of replication. The number attached to it has not held up nearly so well, and the most repeated version of it, that attention collapses after about half an hour, has travelled further than the measurement behind it.

What Mackworth actually did#

The Clock Test was built to stand in for radar and sonar watchkeeping. An unmarked clock face carries a pointer that jumps in equal steps about once a second, and at irregular intervals it makes a double jump, which the observer reports by pressing a button. Watches ran two hours. In the operational task the signal probability was a little over half a per cent, so the observer spends almost all of the watch seeing nothing worth reporting.

Detection fell between the first and second half-hour block and kept falling across the watch. That is the founding result of an entire field, and the estate grades it as peer-reviewed on the 1948 paper in the Quarterly Journal of Experimental Psychology. Much of the same material appeared at greater length in MRC Special Report 268 in 1950, which is where the 1950 date in circulation comes from.

Why the design cannot support the half-hour claim#

Mackworth analysed in half-hour blocks. A decline that shows up between block one and block two tells you the loss had occurred by the end of thirty minutes and tells you nothing about where inside that window it began, how quickly it arrived, or whether it was gradual. The half hour is the resolution of the instrument.

Follow the sentence forward and it changes category twice. In Ironies of Automation in 1983, Lisanne Bainbridge wrote: "It is impossible for even a highly motivated human being to maintain effective visual attention towards a source of information on which very little happens, for more than about half an hour. This means that it is humanly impossible to carry out the basic function of monitoring for unlikely abnormalities." A measurement interval has become a limit of the person.

Then in 2018, reviewing Bainbridge for IEEE Transactions on Human-Machine Systems, Barry Strauch writes in his own voice: "Because of research demonstrating that visual monitoring quality deteriorates after about 30 min, extended operator monitoring of automation is limited and is expected to be increasingly ineffective." No reference follows "research demonstrating". Mackworth's name does not appear anywhere in Strauch's paper, in the body, the references or the notes. The claim now stands as established background with no source attached, in a peer-reviewed journal, inside a retrospective on a paper that Google Scholar had by then recorded 1800 works citing.

Each step is defensible on its own terms. A paraphrase in a five-page conference-derived paper is not a fraud, and a reviewer restating a commonplace is doing what reviewers do. The sequence is the problem, and grading a source by what it measured instead of by how often it is quoted is this estate's only defence against that kind of drift. Whether Bainbridge attached a citation to that sentence could not be confirmed here, because the Automatica original is paywalled and was not read at source for this page.

Seventy-five years on, the mechanism is still open#

Raymond Klein and Brett Feltmate reviewed the whole literature for Frontiers in Cognition in 2025, and their reading is that the decrement itself is solid and its cause is unsettled. Read through signal detection theory, fewer detections can come from a drop in the observer's sensitivity to the difference between a target and a non-target, from a conservative shift in the observer's response criterion, or from both. Early work suggested the event rate decides which dominates: criterion shifts when events arrive slowly, sensitivity declines as the rate rises. The sensitivity account has been challenged recently, partly because false-alarm rates in these studies are so low, and partly because an observer on a long watch may simply stop attending to the source at all.

That distinction matters for anybody designing AI oversight, and it matters as a reason for caution. A person checking model output is usually in a slow-event regime, where the literature's own suggestion points at a criterion shift: the reviewer becomes more reluctant to call something wrong, while remaining perfectly able to see that it is. An intervention aimed at attention would miss that entirely. The review is explicit that no particular mechanism is established, so this is a question to hold open in a design and not an answer to build on.

What this licenses somebody to say, and what it does not#

It supports the claim that sustained attention to rare events degrades, that the degradation is a property of the task and not a failing of the individual, and that an oversight arrangement assuming a person can watch indefinitely has assumed something the evidence contradicts. That is enough to make it the load-bearing objection to most human-in-the-loop designs.

It does not support a number. Nobody should write that accuracy falls by a given percentage after thirty minutes, or that attention fails at the thirty-minute mark, on the strength of Mackworth. The specific percentages in circulation come from secondary literature and the original was not read at source for the estate's entry. It also does not transfer automatically from a clock face to a document: Mackworth's observer was watching for a visual event of known shape, and a person checking AI output is looking for an error whose form is unknown in advance, which is a harder problem and a different one.

Key sources

• Mackworth, N. H. (1948). The Breakdown of Vigilance during Prolonged Visual Search. Quarterly Journal of Experimental Psychology, 1(1), 6 to 21. DOI 10.1080/17470214808416738. Paywalled; not read at source for the estate's entry. Graded entry.
• Klein, R. M. and Feltmate, B. B. T. (2025). The vigilance decrement: its first 75 years. Frontiers in Cognition, 4. DOI 10.3389/fcogn.2025.1632885. Open access, read at source 16 September 2026. Graded entry.
• Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775 to 779. Quoted here at p. 776 as reproduced in Strauch (2018). Graded entry.
• Strauch, B. (2018). Ironies of Automation: Still Unresolved After All These Years. IEEE Transactions on Human-Machine Systems, 48(5), 419 to 433. The uncited restatement quoted above is in his section on monitoring. Graded entry.

Related SuperSkills research#

The argument this sits underneath is the ironies of automation, and the failure of the usual remedy is alarm fatigue. On what the watching does to the watcher, automation complacency and shallow jobs. On whether the arrangement counts as oversight at all, meaningful human oversight and human in the loop is not a safeguard. On other numbers this research has traced back to their source and declined to repeat, the most quoted AI statistics, checked.