Loading page
AlgoArena can take periodic still images during an assessment, but only if the team running it turns it on, and only after the participant grants camera access. Nothing about a participant is ever inferred from their face.
The industry treats the camera as the integrity story. We do not, for a practical reason before an ethical one. Attention monitoring produces output that looks confident but does not survive contact with real participants. It flags nervous people, people with unusual setups, people who look away to think. It misses prepared cheating entirely.
So a flag from the camera lands in front of an evaluator as a number they cannot check and cannot explain to the person it is about. That is the single most common complaint participants have about assessment platforms.
We would rather give you four signals that are boring, countable, and arguable than one signal that sounds authoritative and is not.
Not configurable, not on a roadmap, not available on request.
Where a team does enable snapshots, the images go to that team and nowhere else. No model scores them, and they never move a participant's result on their own.
Four signals, all captured in the browser, none of them biometric. The score is arithmetic, not a model judging a person, so the same session always produces the same number.
How much arrived by paste rather than typing, and whether pasted code carries the shape of model output. Only counted where the question did not permit AI.
How often focus left the assessment, and whether it left in bursts. A burst reads differently from steady checking.
How often the window lost focus entirely. Weighted lightly, because people get interrupted.
Whether a second display was present during the session.
Every deduction comes with the reason that caused it, in plain language, on the report. An evaluator can see "14 tab switches (burst of 6)" rather than a bare score, open the replay at that moment, and decide for themselves. The score is advisory and never rejects anyone.
On Agentic questions the score is not shown at all. Pasting model output is the point of that exercise, so measuring it as suspicion would be meaningless.
Before the timer starts, the participant sees what will be recorded, in the same words on every assessment: periodic still images, not continuous video. The browser asks for camera permission separately.
Snapshots are capped per session and are only ever viewed by the team running the assessment. There is no automated pass over them.
An automated job destroys snapshots once a decision has been recorded and the team's retention window has passed, and in any case within three years.
Images of a person are sensitive data under biometric privacy laws in several places, including Illinois, whether or not anything is inferred from them. That is part of why the setting is off unless a team deliberately turns it on, and why we keep the retention window short. Teams operating where consent or disclosure is required are responsible for meeting those obligations for their own participants.
Integrity context sits beside the work and never overrides it.