Can Coding Assessments Track Your Eye Movement?

Some coding assessment platforms do analyze eye and head movement from your webcam, but not in the way the phrase "eye tracking" suggests. HackerRank documents gaze analysis as a Proctor Mode feature that looks for one specific pattern — looking away from the screen and returning to type. CodeSignal's published anti-cheating page does not list a gaze signal at all. The output is a behavioral flag routed to a human reviewer, not a determination that you cheated. This article explains what that signal measures, why ordinary behavior triggers it, and what you can reasonably do about it before a proctored assessment starts.

What the platforms actually say they do

HackerRank's Proctor Mode documentation describes gaze analysis as a system that "analyzes the candidate's eye and head movements captured by the webcam to identify repeated patterns of looking away from the screen and returning to type, which may indicate the use of external resources." The same page states that only "high-confidence behavioral patterns are flagged as suspicious activity for review," and that Proctor Mode classifies results as High or Medium severity (HackerRank, Proctor Mode).

HackerRank's July 2026 release notes introduced gaze detection as a limited-availability feature and described it as a medium-severity signal specifically in order to reduce false positives (HackerRank, July 2026 release notes). That is a vendor decision worth reading closely: the vendor itself is treating gaze as weaker evidence than the signals it has shipped for years.

Adoption across the category is uneven. As of August 2026, CodeSignal's cheating and fraud page describes a Suspicion Score built from "solution similarity, telemetry, and copy-paste activity," plus recording of "video, audio, and screen activity for the entire assessment session" with "human reviewers validating flagged activity." It does not list gaze, eye movement, or attention tracking among its signals (CodeSignal, Cheating and Fraud). Two platforms with similar market positions have made different choices here, which is why the disclosure screen on your specific test matters more than any general article about proctoring — including this one. For the wider picture of what one platform captures, see our breakdown of what HackerRank proctoring detects and records.

Gaze estimation is not calibrated eye tracking

Laboratory eye tracking uses dedicated hardware and a per-person calibration step to map pupil position onto screen coordinates. Webcam gaze estimation in a proctoring product has neither. It infers a rough head-and-eye direction from a low-frame-rate video feed of an uncontrolled room.

A 2024 paper on AI-assisted gaze detection for proctoring is candid about the gap. Its authors note that "the gaze plot only shows the gaze directions of the test takers, it doesn't show where on the screen the test taker is actually looking at," and that they expect an ML-only system to perform poorly "because calibration will be needed to determine the exact relative positional relationships between the screen, the camera, and the test taker" (Shih et al., arXiv:2409.16923). Their design conclusion is to use the model as a navigation aid that helps a human proctor jump to relevant timestamps, not as a classifier.

That is the honest technical frame. A gaze signal can tell a reviewer roughly when your face turned away and came back. It cannot tell anyone what you were looking at, whether anything was there, or what you read.

Why ordinary behavior produces the same pattern

The pattern the model is built to find — look away, look back, resume typing — is not distinctive. It also describes:

• Thinking. Many people break eye contact with a screen while working through a problem.
• Reading a long problem statement, then returning to the editor to type what you worked out.
• A dual-monitor or laptop-plus-stand setup, where a permitted resource sits outside the webcam's assumed axis.
• Ambient interruption: a person walking past, a pet, a delivery, a phone lighting up on the desk.
• Physical setup, where an external keyboard positions you off-center relative to the camera.

It also describes disability. The Center for Democracy and Technology's analysis of automated proctoring warned in 2020 that such software "could flag blind or autistic students who have atypical eye movements," and made the structural point plainly: "The point is to identify and flag atypical movement, behavior, or communication; disabled people are by definition going to move, behave, and communicate in atypical ways" (CDT, November 16, 2020). That critique was written about academic exams, and it transfers directly to hiring assessments that reuse the same detection approach.

No vendor has published accuracy or false-positive rates for gaze detection in a hiring assessment, and we are not aware of independent evaluation of it as of August 2026. Treat any specific number you see quoted about gaze accuracy in interviews as unsourced until a vendor or researcher publishes the methodology.

The legal line between gaze detection and emotion inference

Gaze analysis and emotion recognition are different things, and in the European Union the difference is now a legal boundary. Article 5(1)(f) of the EU AI Act prohibits placing on the market or using AI systems to infer emotions of a natural person in the areas of workplace and education, with narrow medical and safety exceptions. The Article 5 prohibitions have applied since 2 February 2025 (EU AI Act, Article 5).

Analysts reading the Act alongside Recital 18 conclude that detecting readily apparent expressions, gestures, or movements is outside the prohibition unless those observations are used to identify or infer emotions (Future of Privacy Forum). On that reading, a system that flags "looked away repeatedly" is describing a movement; a system that concluded you were anxious, evasive, or deceptive from the same footage would be inferring an emotional state. If a vendor markets confidence, nervousness, or engagement scoring on interview video, that is the claim to ask hard questions about.

What you can actually do about it

Most of the useful action happens before the timer starts.

Read the consent screen rather than clicking through it. HackerRank's candidate AI notice states that AI features "may analyze your coding, behavioral signals, webcam images, and screen activity," and lists candidate rights that include requesting reasonable accommodations or an alternative selection process, requesting human review of decisions made or substantially assisted by AI, and receiving an explanation of how AI was used if an adverse decision is made (HackerRank, Candidate AI Notice). Those rights are exercised through the employer, not the platform, and their availability depends on your location.

Ask the recruiter which integrity mode is enabled before the assessment. Configuration is per-employer, so "does this test use webcam proctoring" is a fair and answerable question. The same logic applies to the browser-side signals covered in our guide to whether interview websites detect tab switching.

Request an accommodation early if a documented condition affects your eye or head movement. Asking after a flag has been raised is a worse position than asking before.

Reduce ambiguity in your setup. Sit centered in frame with the light source in front of you rather than behind. Disconnect a second monitor if the test requires a single display — HackerRank's Proctor Mode flags new monitor connections during a test regardless of gaze. Keep the problem statement on the assessment screen so that long reads happen inside the frame.

Know what happens after a flag. A medium-severity signal is an input to a human review, and reviews have process attached to them. Our guide to what happens when a coding assessment flags you covers how those reviews typically proceed, and our summary of what candidates are entitled to know about AI in hiring covers the disclosure obligations that apply in several jurisdictions.

Where this leaves the signal

Gaze detection is real, narrow, and comparatively weak. The strong integrity signals in assessment platforms remain the deterministic ones — paste events, typing linearity, solution similarity, monitor and full-screen state — because they measure something the system directly observes rather than something it estimates through a webcam at an unknown angle. Gaze analysis produces a timestamp and a probability, and the vendor shipping it has said as much by classifying it at medium severity.

For a candidate, the practical response is not to perform stillness for a camera. It is to know which mode your assessment runs in, to fix the parts of your setup that would generate ambiguity, and to raise accommodation needs before the session rather than after a flag. If you want the broader map of how assessment environments are configured and what each mode implies, start with our desktop AI interview assistant guide for 2026, which covers where AI assistance is permitted, prohibited, and graded.