Steven Watson | EdgeLab | Research and Scholarship | 19 September 2026

Can a recommendation be good at finding what interests us and still make it harder to live as we intend? That question matters whenever we judge a digital service by whether people keep using it. A long session could be welcome absorption, unwanted delay or something in between. Duration alone cannot decide which.

A newly available accessibility study offers an occasion to separate relevance from control. Its importance, in my reading, is not that it settles what short-video platforms do to everyone with ADHD. It asks us to reconsider what successful recommendation should mean.

What the evidence shows

The survey of 302 adults compared 65 reporting a medical ADHD diagnosis, 85 reporting self-identified ADHD and 152 reporting neither. Participants described their previous short-video use; this was not an experiment. Recommendations matched viewing intentions similarly across groups, while ADHD groups reported greater stopping difficulties and adverse experiences. Reported effects were generally small to medium.

Diagnosis was not clinically verified, medication was unmeasured and proposed interface features were rated from descriptions, not tested in use. The results and appendices therefore support a concern about unequal experiences, not a demonstrated causal mechanism or an effective remedy.

A useful distinction, with limits

The distinction between finding something interesting and choosing how long to continue is the strongest contribution here. We should not require a person to dislike a service before taking their difficulty leaving it seriously. Equally, an account of regret should not give a designer permission to decide that all pleasure was mistaken.

My principal methodological concern is the gap between a reported pattern and an explanation of how it arises. To identify a design effect, I would want comparisons that deliberately change the interface while accounting for prior usage and circumstances. To understand an individual experience, I would want more than a group label. Neither requirement makes first-person accounts disposable; each clarifies what another kind of evidence could add.

A plausible alternative reading is that the pattern partly reflects differences in retrospective judgement or in the circumstances in which people use the platforms. A prospective diary could help distinguish these possibilities. Independently recorded stopping attempts might help too, provided participants control what is collected. Monitoring someone more intensively is not automatically a respectful way to improve their autonomy.

Another caution concerns the word ‘trap’. It is a compelling organising metaphor, but should remain a question about an interaction rather than a fixed description of a person. A useful research programme would investigate when the difficulty appears, when it does not, and which arrangements make a difference.

An AE interpretation: whose continuation counts?

The following is my interpretation through autopoietic ecology, not a finding or theoretical commitment of the authors. AE distinguishes an activity being possible now from what it does to the conditions of later activity. A feed can continue functioning while the person using it finds another intended activity harder to begin. Whether that happens, and over what interval, needs evidence.

Consider a hypothetical person who opens an app for a ten-minute break before joining a friend. The feed supplies enjoyable clips. The person keeps watching but misses the meeting. A recommendation measure could record success while the friendship bears a cost. This invented example is not evidence about the study participants. It makes the boundary of evaluation visible: which consequences are counted, and which are left elsewhere?

EdgeLab’s Concept Guide calls attention to rough closure: an operation can make a definite selection without accounting for all its dependencies. Here the useful question is how a measurable event—continued viewing—comes to stand for a much larger judgement about a worthwhile experience. The error would be to treat that translation as self-evident.

This also connects to the Research Programme’s work on burdens and repair capacity. If maintaining control requires repeated effort, who supplies that effort and who has authority to change the arrangement? An ecological account earns its place only if following those relationships reveals something that an individual-level explanation misses. The study does not prove AE.

What would a better evaluation ask?

I would begin with participant-defined intentions, including the legitimate intention to relax without being productive. Success might mean stopping when wanted, continuing without regret, or being able to revise a plan deliberately. It should not simply mean less screen time. A person who freely chooses a longer session has not necessarily failed.

A practical trial could compare optional stopping cues with a participant’s usual interface, then ask about usefulness, irritation and effects on activities outside the app. It would need sufficient duration to see whether initial benefits persist. Researchers should also examine who cannot use the options comfortably and whether setting them up creates another burden.

There is a real tension between assistance and paternalism. A control that is difficult to find leaves the user doing the work; one that cannot be questioned may take over a decision they wanted to retain. Co-design should therefore include disagreements about what counts as help, not merely preferences between finished buttons.

For designers and institutions, the constructive takeaway is to evaluate the ability to leave as seriously as the ability to enter and participate. For researchers, it is to keep experience, mechanism and remedy distinct. A system’s continued operation is an incomplete answer to the question of whether it serves the people whose attention sustains it.

Publication and review status

Misirlic, V., Mayr, G., and Lex, E. (2026). Quantifying the Engagement Trap: Impact of Short-form Video Recommender Systems on Users with ADHD. Accepted ASSETS 2026 technical paper; full manuscript released 7 September 2026, arXiv:2609.07795v1.

The title and authors match the venue’s accepted-paper list; its technical-paper policy specifies peer review. This review uses the accepted manuscript, not an unreviewed submission. The October conference and formal proceedings publication are still forthcoming; 7 September is the repository-release date, not an asserted ACM publication date.