How CodeSignal spots AI use: Suspicion Score, replay and proctoring

Can CodeSignal detect AI? What CodeSignal itself claims

CodeSignal answers the question directly on its cheating and fraud page: asked whether it can detect AI-assisted cheating, the answer is "Yes", with the Suspicion Score described as identifying "patterns that may indicate AI-assisted behavior, which are often missed by basic plagiarism checks". An older CodeSignal post, Can ChatGPT pass a CodeSignal assessment?, makes the same point in different words: using generative AI "often leaves footprints of detectable indicators" that CodeSignal tracks and reports through the Suspicion Score.

Two things are worth reading carefully in that wording:

• It is pattern detection, not provenance. CodeSignal does not say it can identify which model wrote a piece of code, and it has not published an accuracy rate, a false-positive rate or a list of the exact features it uses. Anything you read claiming a precise detection percentage is not coming from CodeSignal.
• It produces a signal for a human. Every official description ends with a person reviewing the evidence: a hiring team looking at a flagged report, or CodeSignal's proctoring reviewers looking at a recording.

So "does CodeSignal detect ChatGPT?" has an honest answer: it detects the behaviour that usually comes with using an outside tool, and it hands that evidence to people who decide what it means.

How the CodeSignal Suspicion Score works

The Suspicion Score is the layer that runs on the code itself. According to CodeSignal's Suspicion Score article, it lives in the test-taker's Coding Report in two forms:

• Integrity Flagged: one summary field per result with four possible values: Yes, No, N/A (not run, for example because the question type or certified assessment is not supported) and Pending.
• A Suspicion Score per question, ranging from None to High, with a breakdown of the events behind it and buttons such as "View Diff" to inspect them.

The signals CodeSignal documents fall into four groups:

1. Similarity. A Similarity Score with a diff view. CodeSignal's post on preventing and detecting cheating says it compares answers "to all submissions on our platform" and "searches the web for similar solutions posted online".
2. Suspicious paste events. The same post says CodeSignal tracks when candidates copy and paste text from another window, "and what was copied".
3. Description copy. Whether the task description was copied out of the IDE, and how many characters.
4. Telemetry. CodeSignal says it uses telemetry to identify "activities correlated with use of AI assistance, such as unusual typing or speaking patterns". The example dashboard on its fraud page lists "typing linearity" next to paste events, description copy and similarity, and its February 2026 press release explains the idea: complex solutions produced "with minimal pauses or debugging".

Employers can tune part of this. When a company creates an assessment, the cheating detection tab, per CodeSignal's guide to creating a certified assessment, flags copies or pastes of 100 characters or more by default; the threshold can be lowered or raised, and CodeSignal warns that a lower threshold means more false positives. The same guide lists optional fraud prevention toggles: IP monitoring, disabling description copy, and proctoring.

Note what is missing from the documentation: a separate "AI-generated code detector" that scores text for machine authorship. AI use is inferred from the combination of these signals, not from a classifier CodeSignal describes publicly.

Code replay: the keystroke record behind every flag

Most "keystroke tracking" talk about CodeSignal really means the code replay. CodeSignal's article on accessing assessment results says the employer can replay how a test-taker wrote each solution, at up to 32x speed, with idle stretches skipped by default and the view following the cursor. Suspicious paste events appear on that replay as filterable markers, timeline events and entries in a Replay Activity list.

The Suspicion Score article tells reviewers what to look for in the replay. It is useful to know, because it shows how a flag is interpreted rather than just raised:

• the size and number of paste events, with larger and more frequent pastes treated as more suspicious;
• a paste after a long period of no activity, or after the description was copied;
• repeated pastes followed by submissions that error;
• a switch of programming language combined with pastes;
• a description copied near the start, followed later by typing that looks like retyping a solution.

The same article is equally clear about innocent explanations. A simple problem produces similar solutions from everyone, which is what the diff view is for, and a paste may come from an approved resource. Its own conclusion: a high Suspicion Score should not be treated "like a guarantee that a test-taker was cheating".

Does CodeSignal detect tab switching?

It depends on the product, and the documentation is more specific than most forum answers.

• Live CodeSignal Interview: yes. CodeSignal's activity indicator article describes a light near the top of the IDE that is green while the candidate works in the IDE and red when they click away. It "cannot indicate where the candidate is navigating", and it is visible both live and in the replay.
• Assessments (GCA, ICA, custom tests): the support articles do not document a tab-switch counter in the Coding Report. What is documented is closely related: paste events from another window, copying of the description, and, if the employer enables them, IP monitoring and a block on copying the description.
• Proctored assessments: here the question changes, because the whole screen is recorded. Other tabs and windows do not need a counter if they appear in the recording.

What CodeSignal proctoring records and reviews

Proctoring is configured per organisation and assessment, so not every CodeSignal test has it; the basics of how to tell are in our guide to the CodeSignal GCA, and what is and is not monitored during practice is in the CodeSignal practice test guide. When it is on, this is what the official pages describe:

• Camera, microphone and the entire screen for the whole session. CodeSignal's proctoring article lists all three, and the screen-sharing help page tells candidates to choose "Entire Screen" and notes that the screen will be recorded.
• Identity: a photo of your face and of a government-issued photo ID. The fraud page describes matching "the candidate's name and photo ID to the person on camera".
• A setup step that asks you to close unrelated tabs, browser windows and applications before the timer starts.
• Review after the session. The fraud page describes "multi-agent analysis, with human reviewers validating flagged activity and final decisions", with results within one hour. The support article for hiring teams still gives 1 to 3 business days for verification, so expect either.

The reviewers check that camera, microphone and screen share stayed on, that the ID is valid, and that the rules were followed. CodeSignal's hiring-team article lists the common reasons a result is not verified: someone else in the room, leaving the camera's view, using other devices, using an external IDE, and searching online for more than syntax.

CodeSignal's February 2026 press release gives its own breakdown of the signals behind flagged 2025 assessments: 35% involved frequent off-screen referencing, 23% showed unusually linear typing, and 15% showed elevated similarity to known answers or leaked content. These are vendor-reported figures without a published methodology, but they show which signals CodeSignal leans on, and the largest one is about where the candidate looks, something no screen recording captures on its own.

What the employer actually receives

This is the part candidates most often get wrong. The company does not watch your webcam recording. Per the proctoring article, identification and proctoring data is reviewed by CodeSignal only, is not shared with the company, and is deleted within 15 days. Proctoring also has no effect on the score, which is calculated automatically.

What the company does see in its CodeSignal account:

• Your score and Coding Report, including the diff of each final solution against the starting code.
• The code replay for each question, with paste events marked.
• Integrity Flagged and the per-question Suspicion Scores, with the breakdown behind them.
• A proctoring status. A verified result appears under Results; a result that failed review is shown as "Proctoring rejected", and CodeSignal's cheating-detection post says the company receives an explanation for any rejection.
• An AI transcript, if AI was allowed. When an employer enables CodeSignal's own assistant, the conversation is logged in the report. Our post on AI-assisted interviews on HackerRank and CodeSignal covers those rules.

In other words, the employer gets conclusions and code-level evidence, while CodeSignal keeps the video. A rejected proctoring result is still a strong message to a hiring team, even without the footage.

What detection can and cannot see

Put the layers side by side and the limits are clear:

• The Suspicion Score is built from the browser IDE: keystrokes, pastes, copies of the description, the final code and how it compares with other solutions on the platform and the web. It does not see your desk or other devices.
• The screen recording sees what the operating system hands to screen capture. It does not see a phone, a second person, or anything off-screen. Those are what the camera, the microphone and the "not alone in the room" rule are for.
• The camera and microphone catch things code analysis never could: looking away repeatedly, leaving the frame, another voice. CodeSignal's own figures put off-screen referencing at the top of its flagged signals.
• Unproctored tests have only the code-level layer. That is less monitoring, not different rules; the employer's rules still apply.

This is also where tools like Interview Copilot come in. It is a desktop overlay that listens to a live interview and can solve a coding task from a screenshot, and its window asks the operating system to exclude it from screen capture on macOS and Windows (not on the Linux beta). That changes one thing: the overlay window does not appear in a screen share or screen recording. It changes nothing else. It does not hide your eyes from a webcam, change how your typing looks in a replay, make a solution less similar to others, or alter what the rules permit. No tool can honestly promise to be "undetectable" against a session that combines camera, telemetry and human review.

Be honest with yourself about the rules. Proctoring can rely on webcam footage and behavioural signals as much as on the screen, each employer sets its own policy on outside help and AI, and the candidate is responsible for following the rules shown before the session. The GCA rules, for instance, allow web searches for syntax only and do not allow AI at all.

When AI is allowed on CodeSignal

Detection only matters when AI is off-limits, and that is not always the case. Employers can switch on CodeSignal's built-in assistant for a specific assessment or interview, and the safeguards panel pictured on CodeSignal's fraud page includes an "Allow AI Usage" toggle. When the assistant is on, using it is part of the test and the transcript goes into your report. When it is off, the defaults apply. The GCA rules say web searches are for syntax reference only and that "the use of AI is not allowed, including syntax search".

So read three things before you start: the invitation email, the rules screen at the end of setup, and whether an AI tab is present in the IDE. Company-specific processes vary too; our Capital One CodeSignal assessment guide shows how one employer describes its own test. Outside formal assessments, AI help is a different conversation, but for a proctored certified test the answer from CodeSignal is clear.

Employers set these rules themselves. See which companies allow AI in interviews, with sources.

Taking another platform? See also: can HireVue detect AI, can CoderPad detect AI, can Codility detect AI, can HackerRank detect AI.

Sources

• CodeSignal Support — Using Suspicion Score
• CodeSignal Support — Accessing and sharing a test-taker’s assessment result
• CodeSignal Support — How to Create a Certified Assessment
• CodeSignal Support — What is proctoring and how does it work?
• CodeSignal Support — I can’t share my screen. What do I do?
• CodeSignal Support — What happens after I send a proctored assessment invitation to a test-taker?
• CodeSignal Support — If I’m interviewing a candidate remotely, can I see if they click away from the IDE?
• CodeSignal Support — General Coding Assessment (GCA) Rules and Setup
• CodeSignal — Cheating Detection & Fraud Prevention
• CodeSignal blog — Prevent and detect cheating in recruiting
• CodeSignal blog — Can ChatGPT pass a CodeSignal assessment?
• CodeSignal press release — assessment fraud more than doubled in 2025 (25 February 2026)

Every claim above was checked against these pages on 31 October 2026. CodeSignal updates its products often, so check the rules shown in your own assessment before you start.

Related articles