Can Codility Detect ChatGPT? Typing Patterns and AI Matches

Codility does not see ChatGPT. It looks for what ChatGPT leaves behind. As of September 9, 2026, Codility documents two default-on checks aimed at AI-generated code: Typing Pattern Detection, which its documentation says is "designed to identify candidates who may be retyping solutions generated by external AI coding assistants like ChatGPT or Claude," and a Similarity Check that cross-references submissions "against AI-generated solutions." A third layer, Device Integrity, targets hidden AI desktop apps but requires the employer to move the test into the Codility App. None of these produces a verdict; Codility says each flag should be reviewed by a person.

Layer one: how the code was typed

Codility's Typing Pattern Detection documentation is explicit about its target: "It addresses the increasingly common practice of using AI tools on secondary devices to bypass assessments." The mechanism is keystroke structure, analyzed after submission: "It tracks the sequence and structure of keystrokes to determine whether the solution was written organically or in an unnatural coding pattern consistent with copying from an external source. Authentic coding is typically iterative, involving edits and rewrites, which differs from the sequential input seen when retyping."

Three facts about it matter for candidates. It "is on by default but can be turned off per assessment." It is silent: "The candidate is entirely unaware that this signal is being collected or evaluated." And it has documented gaps. It does not run on multiple-choice, debugging, essay, file-upload, or VS Code tasks, and "Solutions that require less than 10 lines of code are automatically skipped to reduce the potential for false positives."

Codility also bounds the claim: "The flag suggests a high likelihood, not certainty, of retyping," and "it should not be used alone as grounds for automatic candidate rejection."

Layer two: what the code looks like

The Similarity Check documentation describes a comparison against "over 12 million assessments" and web-scraped leaks, robust to "changing identifiers, reformatting code, or making small changes in the code structure." Its AI section says: "the submitted solutions are cross-referenced against AI-generated solutions. If they meet a certain set of parameters and indicators, they get flagged for review, as each AI model tends to converge on a similar answer for the same question. Similarity check operates with dozens of such solutions for each task and technology."

Two limits apply. The check runs only on candidates who "reach your Passing Score," or 40% where none is set. And Codility states, "We do not use AI to decide whether a solution should be flagged as potentially plagiarized," and that a warning "doesn't necessarily suggest cheating."

Pasting and retyping trade one layer for the other. A paste skips the typing signal and shows up as a paste event and a similarity candidate; retyping avoids the paste event and produces the sequential keystroke pattern the first layer is built to catch. How paste events appear on the report is in whether Codility detects tab switching and copy-paste.

Layer three: the tool on the machine

Codility's integrity features overview describes the gap the first two layers leave: "A growing category of AI tools runs as hidden desktop applications that feed candidates answers in real time — and they're built to be invisible to screen sharing. Proctoring can't catch what it can't see."

Its answer is Device Integrity, in Preview. The Device Integrity documentation says the test runs inside the Codility App for macOS and Windows, which "checks the device for known cheating tools," and that "Any detections appear in the integrity widget on the candidate report." Codility also limits its scope: the app "checks only for the presence of known cheating tools. It does not record the candidate's screen, monitor keystrokes, or collect personal data beyond what the standard Codility assessment already collects."

"Known" is doing work in that sentence. Codility does not publish the list, and a signature check by definition covers the tools it has signatures for. Control, a desktop AI interview assistant for interviews, assessments, and screen-share workflows, makes no claim about how the Codility App classifies it.

The signals around the code

Even without proctoring, one behavioral event is named as an AI tell. The Proctoring documentation says copying the task description "could indicate a candidate's attempt to look up the task or employ AI platforms like ChatGPT," and the IDE blocks the copy so the attempt itself is logged when proctoring is on.

After submission, an employer can add AI Follow-Up Questions. Codility's AI "generates three targeted questions unique to that submission. Each question is time-boxed to 2 minutes, with copy-paste prevention and keystroke monitoring active during answers." The transcript goes into the report. A solution you cannot explain in two minutes is the case this feature exists for.

When AI is allowed

Codility Interview can include AI by design. The Interview overview documents an AI Assistant and AI Copilot that account admins toggle on or off, so some interviews expect you to use the built-in tool. For Screen tests the rule is whatever the intro page says; Codility's candidate FAQ tells you to "Review the assessment policies included in your test link." How to confirm permission in writing is covered in when you can use AI during a technical interview.

What this means for candidates

If the test prohibits AI, Codility has two default checks on the finished code and a possible third on the device, and none of them is switched off by the absence of a camera. Prepare with AI beforehand and type your own solution in the editor; iterative editing is what the typing model expects to see. If AI is permitted, use the sanctioned tool and be ready for follow-up questions. The full inventory of Codility's checks, including the older IP and similarity layers, is in what Codility detects.