How HackerRank Spots AI-Written Code, Signal by Signal

The short answer: HackerRank detects behaviour, not “AI text”

Searches like “does HackerRank detect ChatGPT” assume a detector that recognises AI-written code the way some essay checkers claim to recognise AI-written prose. HackerRank does not describe anything like that. Its own AI Plagiarism Detection article says the model predicts suspicious behaviour from coding and typing patterns, which is why it is useful “for identifying the use of AI tools like ChatGPT when their usage is restricted”. In other words, it infers outside help from how a solution appeared, not from the solution alone.

The layers, according to HackerRank’s Test Integrity overview:

• On for every test: MOSS code similarity, copy-paste tracking and tab proctoring.
• Optional, chosen per test: AI Plagiarism Detection, webcam snapshots (Image Proctoring), AI analysis of those snapshots (Image Analysis) and photo identification.
• One integrity mode, if any: Secure Mode, Proctor Mode or Desktop App Mode. The last two are part of HackerRank’s paid AI Add-on.

Live interviews (the product once called CodePair) have their own, lighter set of signals, covered further down. If you want the general picture of the test and interview format first, read what to expect from a HackerRank interview; this page stays on detection.

To compare HackerRank with CoderPad, CodeSignal, Codility and Karat, see what interview platforms actually log.

Layer 1: MOSS code similarity, on for every test

HackerRank’s MOSS article states that MOSS (Measure of Software Similarity) plagiarism detection “is turned on by default for all HackerRank tests”. It compares your code with other submissions in HackerRank’s database by splitting it into tokens and looking for substantial overlap. Because it works on structure, the article says that renaming variables or adding whitespace is “generally ineffective” against it.

Matches are graded by percentage: High at 90% or more, Medium at 80–90%, Low below 80%. A flagged candidate gets a red marker in the employer’s candidate list, and the reviewer can open the matching submissions side by side and mark the flag as confirmed or not. The article is candid about one limit: the tool cannot identify the exact source of the copying, and it recommends reviewing the code playback before making a decision.

What MOSS can see: code that is structurally close to something already in the database, including another candidate’s answer or a widely shared solution. What it cannot see: a fresh solution that does not resemble anything stored. HackerRank said as much itself when it launched its AI model in a June 2023 blog post, writing that GPT-4 “can also bypass MOSS code similarity” and that similarity checking alone was no longer enough. That gap is the reason the second layer exists.

Layer 2: AI Plagiarism Detection, HackerRank’s answer to ChatGPT

The AI Plagiarism Detection model is a per-test setting the employer turns on under the test’s integrity settings. When it is on, candidates must consent before starting, so you will normally see that it is active. The article lists the signals it uses, and says the list is not exhaustive:

• Copy-pasting – significant portions of code copied from external sources.
• Coding patterns – structural similarities in code and distinct coding styles.
• Iterations – how the code evolves over multiple attempts on layered tasks.
• Tab-switching – leaving the active test tab.

It also weighs how the code was written and how long the problem took. Flagged attempts are labelled High or Medium depending on the model’s confidence, and the reviewer gets a keystroke code player to jump to the relevant part of the session. The 2023 launch post describes the same idea at a higher level: “dozens of signals” across coding behaviour, the submission and the question itself, and a claim that the system “repeatedly detects ChatGPT-generated solutions, even when those results are typed in manually”.

On accuracy, HackerRank gives two figures and they are not the same measurement. The launch post claims 93% accuracy. The current help article states an overall precision of 85%, meaning that 85% of flagged sessions are correct, and adds that “human oversight remains crucial”. Neither page publishes the dataset or method behind the number, and we found no independent evaluation, so treat both as the vendor’s own claims. The help article also lists limits worth knowing: the model only covers coding questions, and it works less well on problems that need very little effort or very few lines of code, because there is not enough behaviour to analyse.

Layer 3: copy-paste and tab tracking

Two signals need no setup at all, according to the Test Integrity overview. Copy-paste tracking records each copy and paste in coding, subjective and sentence-completion questions, keeps the pasted content in the report and shows it on a timeline. Tab proctoring records how often and for how long you leave the test window, reported as the number of window exits and the out-of-window duration.

The useful thing to understand is the boundary of each. Copy-paste tracking sees what lands in the editor through the clipboard; it does not see code you type by hand, which is exactly why the AI model looks at typing patterns instead. Tab proctoring sees that the test window lost focus and for how long; in a plain browser test it does not, by itself, show what you were looking at. Recording the rest of the screen is the job of Proctor Mode, below. HackerRank’s candidate guidance on malpractice lists switching tabs, exiting full screen, copying code from online sources and outside help as malpractice, so the employer reads these numbers in that light.

Layer 4: Secure Mode, Proctor Mode and the Desktop App

Secure Mode. According to the Secure Mode article, the test runs full screen, a multiple-monitor check runs before you start and continues during the test, copy-paste from outside the editor is disabled, and you get a warning when you switch tabs or leave full screen. The page describes warnings and logging rather than hard blocks. Violations end up in the report with an integrity result of High or Medium.

Proctor Mode. The Proctor Mode article describes an AI feature that “simulates live human proctoring”. It includes everything in Secure Mode and adds:

• Screenshot analysis. You are asked to share your entire screen. Screenshots are taken every 15 seconds, and every five seconds after suspicious events such as a window switch. A model labels each screenshot and looks for answer-sharing sites, external AI assistants, collaboration tools and screen recorders.
• Webcam. On tests that collect candidate details, webcam images every five seconds, with checks for a missing face, multiple faces or a secondary face, and object detection for phones and tablets, even partly in frame.
• Conversation detection. Type-and-delete patterns in the editor that may indicate someone is communicating with you through it.
• Gaze detection. Repeated looking away from the screen and returning to type. HackerRank lists this as limited availability.
• Session replay. A video-style timeline of the browser screen with flagged events and labelled screenshots.

The Image Proctoring and Image Analysis article adds a “static face” category aimed at static image projections, that is, attempts to simulate a live presence on the webcam.

Desktop App Mode. The strictest option. The Desktop App Mode article says the test runs inside a native app for Windows and macOS that includes all of Proctor Mode, blocks other applications “including tools such as Cluely, video conferencing, messaging applications, or LLM applications”, detects and closes programs opened during the test, prevents screen sharing and remote access, refuses to run in a virtual machine, and requires a single, non-mirrored display. Here the control is at the operating-system level, not just inside a browser tab.

Live HackerRank interviews: integrity signals and a human watching

A live interview is a different situation: there is a person on the call. HackerRank’s Interview Integrity Signals feature, which an admin enables in interview settings, flags four kinds of activity in coding, project, code-repository and whiteboard questions: leaving the coding environment for another tab, pasting from an external source, resizing the browser window, and connecting or disconnecting monitors. The monitor check needs your permission and works only in Chrome; if you decline, the interviewer is told.

Rather than pinging the interviewer on every event, the system raises a grouped “Multiple Suspicious Activities” alert when it sees a cluster, and the interviewer can then ask you to share your screen. The article itself notes that pasting “does not always indicate misconduct”, since copying from forums, AI tools or your own editor may be acceptable in some contexts. A separate AI Add-on option, Screen-to-Interview Identity Match, compares your identity from the screening test with the person in the interview.

The page does not describe any signal that reads AI authorship in a live interview. In practice the stronger signal there is the interviewer: follow-up questions about why you chose an approach tend to separate understanding from recitation. Some employers now allow HackerRank’s built-in AI Assistant during interviews, with the conversation visible to the interviewer; the rules for that are in our post on AI-assisted interviews on HackerRank and CodeSignal.

What the employer sees, and what you do not

HackerRank’s guide to reviewing integrity issues lists what a recruiter can open after a proctored test: an Integrity Issues indicator (High, Medium or no issues detected) with a summary, tab-switch and full-screen exits with times and durations, webcam flags, screenshot analysis, phone detection, conversation detection, high similarity with other candidates shown side by side with a percentage, and a session replay. On any test, the report also carries copy-paste frequency, window exits, the MOSS flag and the code playback.

Two points matter for how this is used. First, the language is hedged throughout: behaviour “may indicate the use of external resources”, and each item is described as suspicious activity “that requires review”. Nothing in the documentation says a flag rejects a candidate automatically; the decision belongs to the employer. Second, the information is one-sided. HackerRank’s candidate help centre says it “does not share plagiarism check results with candidates” and points you to the recruiter if you have questions. If you believe a flag is wrong, for example because a common algorithm produced a high similarity score, a short factual note to the recruiter is the only channel.

What each signal can and cannot see

Put side by side, the coverage looks like this, based only on what HackerRank documents:

• MOSS sees structural matches with stored submissions. It does not see an original solution, wherever it came from.
• AI Plagiarism Detection sees how code was produced: pastes, typing, iterations, time, tab switches. It is weaker on short, low-effort problems and covers coding questions only.
• Copy-paste and tab tracking see clipboard input and focus loss, not what happened outside the test window.
• Proctor Mode sees what screen capture shows, what the webcam shows, and editor behaviour.
• Desktop App Mode controls which applications may run at all.
• In live interviews, the signals are tabs, pastes, window size and monitors, plus a human asking questions.

This is where tools like Interview Copilot often come up, so it is worth being precise. Interview Copilot is a desktop overlay whose window is excluded from screen capture at the operating-system level. On Windows this kind of exclusion is a documented API, SetWindowDisplayAffinity, and Microsoft itself notes it is not a security guarantee and does nothing about someone photographing the screen. What it changes is narrow: the window does not appear in screen shares or screenshots taken through the operating system’s capture APIs. It does not change the webcam, gaze detection, typing and paste patterns in the editor, the similarity check, or what an interviewer infers from your answers. HackerRank’s Desktop App Mode is also built to close other applications during the test and names this category of tool explicitly.

The honest summary is that no tool makes a rule disappear. Proctoring can rely on webcam images and behavioural signals as well as the screen, every employer sets its own policy on outside help for each test or interview, and the candidate is responsible for knowing and following it. If AI is allowed, use the assistant the employer provides. If you use Interview Copilot, the AI coding interview assistant page explains what it does; check first that the round you are in permits it. For other platforms, see can HireVue detect AI, can CoderPad detect AI, can CodeSignal detect AI, can Codility detect AI.

Sources

All pages were checked on 31 October 2026. HackerRank’s help articles show only relative “last updated” dates, and features change, so check them again before relying on details.

• HackerRank Test Integrity – default features, integrity modes, optional add-ons.
• Plagiarism Detection Using MOSS – default MOSS check, thresholds, review advice.
• AI Plagiarism Detection – signals, consent, 85% precision, limits.
• HackerRank launches AI-powered plagiarism detection (June 2023) – MOSS bypass, vendor accuracy claim.
• Secure Mode, Proctor Mode, Desktop App Mode – integrity modes.
• Image Proctoring and Image Analysis – webcam snapshots and analysis categories.
• Interview Integrity Signals – live interview signals.
• Review Integrity Issues in Proctor Mode – what the employer sees.
• Malpractices in Tests – candidate-facing rules; results not shared.
• SetWindowDisplayAffinity (Microsoft Learn) – capture exclusion on Windows and its limits.

Related articles