How people's attention and decisions change when they know someone else can see what they are doing.

The question

People rarely use a product unobserved. Activity is visible to teammates, managers, whoever shares the document - and the moment someone knows they can be seen, their behaviour changes. That makes attention serve two functions: the eyes gather information, and they also give it away.

No established method existed to measure how behaviour changes as we balance these functions, so I built one. I paired users. The first navigated a display and selected a target while their gaze was recorded, sometimes trying to help their partner find it, sometimes trying to hide it. The second then watched that gaze, navigated the same display, and guessed the choice. That gives two measures from one task: what the first user gave away while trying to do their own job, and how the second user read it.

Each study below varies one part of that exchange. All were randomized, within-user, and modelled with mixed-effects regression. I rebuilt the lab protocols as browser-based studies on an open-source stack I taught myself - no tooling cost to the lab, and roughly ten times the lab sample - 100K+ observations per study.

• Three times the weekly data collection once studies ran remotely and asynchronously.
• Sessions about 20% shorter, sized from prior effect sizes with no loss of validity - which is what made running studies in parallel affordable.

What follows is a sample of six years of self-managed research: 10+ studies, N=1,600+, and $215K in competitive funding, delivered on time and budget.

Display complexity degrades attention management

Question. Can users balance their attentional signals in visually complex computer displays?

Design. Manipulated the display's visual complexity and the task's difficulty (display navigation while self-monitoring attention); tracked search accuracy and eye movements. n=42.

Result. Task success did not decay linearly but collapsed past a threshold. This forced users to make trade-offs between the quality of their decisions and the quality of their attentional signals. Modelling identified these thresholds and their consequences. (Report under review.)

For a product team. Density has a cliff, not a slope. Locating where it sits on your own surface turns "reduce clutter" from a heuristic into a criterion a design review can hold a decision to.

People act on trust more than truth

Question. How much are user decisions biased by another person's attention, and does it depend on whether the user trusts that person?

Design. Modelled users' choices against the strength of another person's attention nudging them to a given choice. Then manipulated whether the user trusted the attentional nudges. n=157, replicated at n=360.

Result. User trust moved decisions more than the nudges themselves - about 30 percentage points from framing alone. High trust followed the cue; low trust discounted it. (Attention, Perception & Psychophysics, 2025.)

For a product team. Credibility signals, source badges, and ranking position can move a decision more than the content they label. Most products never measure that effect, and it is not uniform across user segments.

User confidence is not always a valid measure

Question. Can a user's self-reported confidence in their attention-management ability be trusted?

Design. Collected a confidence rating after every task and tested it against users' actual performance - whether they successfully communicated with their partner - comparing instances where users did and did not know their behaviour was being shared. Eye-tracking n=40 with telemetry n=326.

Result. Confidence tracked performance when users knew they could be seen by their partner, but barely tracked performance when they thought they were not being observed. When users weren't managing their signals, their confidence had nothing to track, so their self-reports were inaccurate. (Journal of Experimental Psychology: General, 2026.)

For a product team. Self-reported confidence is worth instrumenting where users can see the consequences of what they did, and misleading where they cannot. That tells you which surveys and satisfaction scores to trust before you build a metric on one.

Segment differences show up where context has to be integrated

Question. Does the trust effect hold for every user, or do some segments read the same cue differently?

Design. Split users by the AQ-10 screening threshold for autistic traits. One experiment said nothing about the other person's intent (n=110); a second told half the users they were being cooperated with and half that they were being deceived (n=242). Same task, same stimuli, one added variable.

Result. With nothing said about intent the two groups were indistinguishable - same accuracy, same weighting of the signal. They diverged only once a belief about the other person had to be folded in, and only under low trust: lower-trait users turned away from the most-looked-at element, higher-trait users went on reading it. (Report under review.)

For a product team. Benchmark the bare interaction and your segments will look the same. The difference appears where users combine the interface with something they have been told about it - a warning, a source label, a trust badge. Test the segment there, or you will report a parity your product does not have.

Research that earns adoption

Six years of experiments demonstrated that how a user reads a screen is not a property of the screen. It moves with what they trust, who can see them, how much is on it, and who they are. Adoption is decided in that gap, and the gap can be measured and reduced. The CheckingIn case study is the same problem from the applied side: HR leaders read the dashboard perfectly well and still did not know what to act on, so the interface worked and the product did not.

I apply research to this gap in two ways. First, choosing the approach fit to the question, decision, and constraints, including a willingness to rule out an unhelpful study. Second, building a measure where none exists: stating in advance what would prove the hypothesis wrong, validating methods and data before trusting a result, and building the infrastructure that keeps it standing at speed - reproducible pipelines, documentation, collaborative or version-controlled workflows I invite the team into, and AI-assisted qualitative coding with researcher oversight.

Papers and code