Assessment design
What a Paste Looks Like to a Coding Assessment, and What a Flag Does
Often, yes: HackerRank publishes what its coding screens can collect, and copy-paste activity sits on the list next to tab switching, code-writing patterns and time taken [1][2]. Which signals run depends on what the employer switched on [2]. What that produces is a flag on your report, not a verdict. HackerRank puts its own model's precision at 85 percent and tells hiring teams human review is still needed [1]. A flag means a person is meant to look, and walking through your code is how you answer one.
The takeThe interesting part of a self-reported 85 percent is the rest of it [1]. By that number, roughly one flagged session in seven is a model's mistake sitting in somebody's report, and no page here publishes a recall figure, an appeal route, or a test set anyone outside the company has seen. That is not a scandal. It is an ordinary young signal, and it is why the vendor tells recruiters to look for themselves.
Where Olive fits
Open a role and see what the work shows
If the assessment you are sent is Olive, using the AI assistant is the point: a person writes or edits six findings, among them how you framed the problem, what you delegated and what you verified, in words rather than a number. You are granted the identical report the employer reads.
Assess your shortlistWhat Does a Coding Platform Say It Records?
Copy-paste activity is on the published list, next to tab switching, timing and code-writing patterns. HackerRank's knowledge base says its plagiarism model analyses code-writing patterns, time taken, copy-paste activity and tab-switching behaviour 1. The same page says suspicious attempts are classified as High or Medium, and that candidates consent before a test starts when the feature is enabled 1. Its Screen page adds tab proctoring, multiple-monitor detection and full session replay 2.
Two of those signals are worth separating. Copy/paste tracking records an event. The plagiarism model reads a pattern across the whole session, including how the code took shape and how long it took 12. A single paste is one input to that pattern.
- CodeSignal describes a Suspicion Score assembled from solution similarity, telemetry and copy-paste activity. Its proctoring can record video, audio and screen activity for a whole session, with a government-issued photo ID matched to the person on camera 3.
- Vervoe disables copying the question text and lets an employer switch off pasting into answer fields, in which case an extra guidelines screen appears before you start 4.
None of that says which of these your test uses.
What runs depends on what the employer switched on, and neither the HackerRank nor the CodeSignal page says which are defaults 23. The invitation and the instructions screen in front of you are the only places that name what is running on this particular assessment. If they are silent, the recruiter can answer the question, and asking is not an admission of anything.
These pages are also sales material. Apart from a single precision figure on one of them, they describe what a product can do rather than how accurately it does it 1. None publishes a false-positive rate a candidate could check 23.
Why a Flag Is a Probability, Not a Verdict
Behind the label is a model's guess, and one vendor publishes how often it lands. HackerRank states that its model's overall precision is 85 percent, meaning flagged sessions are correct 85 percent of the time, and that human oversight remains crucial because of it 1. That is the vendor's own figure, with no test set or period published beside it. Flagged attempts are highlighted in candidate reports, and recruiters can replay keystrokes to see what triggered one 1.
Turned around, the same number says roughly one flagged session in seven is not what it looks like.
A High or Medium label is a sorting decision inside a report.
Precision is also only half a measurement. It says how often a flag is right and nothing about how many sessions the model never flags, and no recall figure appears on any of these vendor pages 13.
The consequence for you is narrow. A flag is a reason for a person to look at your session, and a person looking is something you can answer. What to say when you're wrongly accused of using AI covers the same problem one step later, once the looking has already produced an accusation.
What Happens After a Flag?
A person is meant to review it, and may ask you to explain the work. CodeSignal says its sessions are evaluated with automated analysis and human reviewers validating flagged activity and final decisions, with results returned within an hour 3. HackerRank puts the flag in the candidate report for the hiring team to read 1. What an employer does with it after that is the employer's decision, and no vendor page here commits them to anything.
Running the finished code through a detector does not settle it either.
An ICSE 2025 study evaluated five AI-content detectors plus GPTSniffer, which is built specifically for code, across three benchmarks. Accuracy was mostly below 0.6, which the authors call ineffective 5. Those were benchmark functions in Java, C++ and Python rather than take-home submissions, so the result speaks to the tools, not to your repository. It still means a finished file does not answer the authorship question on its own.
Ask for the specifics in writing before you defend anything, because you cannot answer a finding nobody has described.
So the check moves to a conversation, which is where employers are being pointed. The employer-side guidance for running a coding interview with AI open treats the round as an audit of flawed AI output rather than a race to produce an answer. On the candidate side, being asked to redo the exercise live without AI is a stage some loops give every finalist, so whether everyone at your stage gets it tells you whether it is routine or aimed at you.
Either way, what is being checked is whether you can explain what your solution does and why.
Read the Test's Own Rule Before You Paste
The rule is set per test, and one platform says so in as many words. HackerRank's Screen page states that if a test is designed to allow AI, it is fair for everyone to use it, and that a test explicitly instructing otherwise will flag the use as suspicious activity 2. Same platform, opposite rules, set by whoever wrote your assessment, and where it says no, a pasted solution you did not write breaks it, registered or not.
That makes the instructions screen the document that matters, more than any list of signals a vendor advertises.
A paste is also not one act. Pulling in a utility you wrote last month, copying a snippet out of documentation, and submitting a solution somebody else produced are three different things, and none of them is settled by how the characters arrived in the editor. A person is what tells them apart, and that is what a precision figure short of certainty concedes.
Screenshot the instructions before the timer starts, so the rule you followed is still readable later.
The format is changing underneath all of this. Evaluated against 115 Python problem statements taken from HackerRank, OpenAI's Codex solved 96 percent of them zero-shot in 2022, with clear signs the model was reproducing memorized code rather than writing it 6. That is an old model on public problems, so read it as a floor and not a headline.
Employers looking at the same finding are being told to replace an assessment a model solves in thirty seconds with work where the model's answer is only the starting point.
If you are flagged for work you did yourself, ask what was flagged and offer to walk through the solution line by line.
Common questions
Does a coding assessment record when I paste?
The platforms say they can. HackerRank lists copy/paste tracking among the signals on its Screen page, and its plagiarism model reads copy-paste activity alongside code-writing patterns, time taken and tab switching 12. CodeSignal describes copy-paste activity as one input to a Suspicion Score 3. Vervoe lets an employer switch pasting off in answer fields entirely 4. Which of these is running depends on what the employer turned on, and the pages do not say which are defaults.
Does a plagiarism flag mean I am rejected?
The vendor pages do not say so. HackerRank puts a flag in the candidate report for the hiring team to review and says human oversight remains crucial, since the model's precision is 85 percent by HackerRank's own self-reported figure 1. CodeSignal says human reviewers validate flagged activity and final decisions 3. What an employer does with a flag is the employer's decision, and none of the vendor documentation commits them to a process or gives a candidate an appeal route.
Can anyone prove the code was written by AI?
Not from the file alone. An ICSE 2025 study tested five AI-content detectors and GPTSniffer, a detector built for code, across three benchmarks, and found accuracy mostly below 0.6, which the authors call ineffective 5. That study used benchmark functions rather than take-home submissions, so it is evidence about the tools, not about any one repository. It is why the useful check is a conversation about your solution instead.
Is pasting my own earlier code allowed?
Your own earlier code is still governed by the test's stated rule, which is set per assessment. HackerRank's Screen page shows the same per-test logic for AI: if a test is designed to allow it, it is fair for everyone to use, and a test explicitly instructing otherwise will flag the use as suspicious activity 2. Read the instructions screen, and if it does not cover reusing code you wrote before, ask the recruiter before the timer starts instead of guessing.
What should I do if I am flagged for my own work?
Ask what was flagged and what produced the finding, in writing, before you argue anything. Then offer the check the flag was standing in for: a walkthrough of the solution, or your version history. HackerRank says recruiters can replay keystrokes to see why somebody was flagged, so on HackerRank you can ask what the replay showed 1. Some employers will not reopen a decision, and knowing that going in is part of the answer.
References
- 1. AI Plagiarism Detection support.hackerrank.com Supports what the plagiarism model reads (code-writing patterns, time taken, copy-paste activity, tab switching), the High/Medium classification, keystroke replay, and the stated 85 percent precision with human oversight still required.
- 2. HackerRank Screen (product page) hackerrank.com Supports the listed integrity signals (tab proctoring, copy/paste tracking, multiple-monitor detection, session replay), the per-test rule that AI is fair game where a test allows it and flagged where a test forbids it, and the absence of any published accuracy or false-positive rate on the page.
- 3. Cheating Detection & Fraud Prevention codesignal.com Supports the Suspicion Score built from solution similarity, telemetry and copy-paste activity, the video, audio and screen recording with government ID matching, human reviewers validating flagged activity and final decisions, and the absence of any published accuracy or false-positive rate for the score.
- 4. Vervoe's Anti-Cheating Features help.vervoe.com Supports the claim that copying the question text is disabled and that an employer can turn off pasting into answer fields, with an extra guidelines screen shown to the candidate.
- 5. An Empirical Study on Automatically Detecting AI-Generated Source Code: How Far Are We? arxiv.org Supports the claim that measured AI-content detectors, including the code-specific GPTSniffer, scored accuracy mostly below 0.6 on code benchmarks and were called ineffective by the authors.
- 6. Codex Hacks HackerRank: Memorization Issues and a Framework for Code Synthesis Evaluation arxiv.org Supports the claim that Codex solved 96 percent of 115 HackerRank Python problems zero-shot in 2022 with clear signs of reproducing memorized code, which is why public exercise banks are being replaced.
6 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.