Assessment design
Your Coding Assessment Can Be Replayed, Not Just Read
A coding assessment can replay how you wrote your code, if the employer switched that feature on. HackerRank's knowledge base says recruiters can replay keystrokes to see why a candidate was flagged [1], and its Screen page lists full session replay among the integrity signals it publishes [2]. CodeSignal says every AI-assisted assessment comes with a session replay plus a full transcript of the candidate's AI interactions [3]. What a replay shows is the work: edits, runs, pauses, and anything that arrives in one block.
The takeIt is easy to treat the recording as a threat and stop there. It is also the part of a coding assessment where your thinking is most visible. A finished solution that arrives whole says nothing about how it got there. A wrong first attempt, run and corrected in the editor, carries something a static file cannot.
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: you work a role-grounded assignment with it open, think aloud or type, and a person writes or edits six findings describing how you framed the problem, what you delegated and what you verified. You are granted the identical report the employer reads, free.
Assess your shortlistWhat Does a Replay Actually Show?
The sequence of your work, rather than its result. HackerRank's knowledge base on AI plagiarism detection says flagged attempts are highlighted in the report a hiring team reads, and that recruiters can replay keystrokes to see why a candidate was flagged 1. The same company's Screen page lists full session replay alongside tab proctoring, copy and paste tracking, multiple-monitor detection and leaked question alerts 2.
None of it records your thinking, only what your hands did.
CodeSignal publishes a version of the same feature. Its May 2025 announcement of AI-assisted assessments says every one of them comes with a full transcript of the candidate's interactions with the AI, plus a session replay 3. The stated purpose is that hiring teams can see what a candidate asked and how the suggestions were used 3. On that product the prompts you type are part of the record, not a private scratchpad.
The same announcement says the employer picks the assistant's mode, from a full copilot down to guided help with syntax and documentation 3.
The record is longer than the file it produces.
Both companies are describing what an employer can switch on, and neither page says which pieces are on by default. A given assessment may use none of them. HackerRank states that its plagiarism feature covers coding questions only, and that it asks for the candidate's consent before the test starts when it is enabled 1.
So the honest answer is: it can be, and not always.
Hiring teams have their own written guidance for this. What employers are told to look for in a candidate's AI chat log reads a session for acts rather than volume: where the first message aimed, what got thrown out, and what was checked outside the chat. A replay of an editor can show some of the same acts, such as what got thrown out and what got run.
What Else Gets Recorded During the Session?
More than the editor, when the employer turns it on. CodeSignal's cheating and fraud page says its proctoring records video, audio and screen activity for the entire assessment session 4. Identity verification asks for a valid government-issued photo ID, matched to the person on camera, and the verification status reaches the hiring team with the results 4.
HackerRank's Screen page describes a comparable stack 2. It names a proprietary AI plagiarism model and an AI proctor that intervenes only when suspicious activity is detected 2. A third piece, image analysis, verifies candidate identity and monitors for multiple faces during an assessment 2.
None of those pages publishes an accuracy figure, a false-positive rate, or a route to contest a flag.
CodeSignal also describes a Suspicion Score, built from solution similarity, telemetry and copy-paste activity, and intended to flag patterns that may indicate AI-assisted work 4. The same page says sessions are evaluated with automated analysis, with human reviewers validating flagged activity and final decisions, and results returned within one hour 4. A score built that way is a signal routed to a person, not a verdict.
A camera is not a given, and neither is a replay.
Nothing on a vendor page tells you which of these a particular employer chose. The invitation, the setup screens and any consent prompt before the test are where to look, along with asking the recruiter, and they are worth reading slowly rather than clicking past.
The rule on AI is set per test rather than per platform. HackerRank's Screen page puts it plainly: if a test is designed to allow AI, it is fair for everyone to use it 2. If a test explicitly instructs otherwise, using AI will be flagged as suspicious activity 2. Two tests on the same platform can therefore carry opposite rules. Where a test does allow it, what an AI-open coding round is actually testing is the better thing to read before you start.
Why a Wrong First Attempt Is Worth Leaving In
Because the tools measured so far cannot reliably read from a finished file how it was made. An ICSE 2025 study tested five AI-content detectors plus GPTSniffer, a detector built for code, against four language models on standard code benchmarks 6. Accuracy was mostly below 0.6, which the authors call ineffective at detecting AI-generated source code 6.
Those benchmarks are standalone functions rather than repositories or take-home submissions, and the detectors were the ones available when the study ran. The researchers' own purpose-built model did better, reaching an F1 of 82.55, and it is not a product a hiring team can buy 6.
Nothing there supports checking a submitted file for AI authorship.
The reverse holds too: an honest submission cannot be proved honest by the file alone.
What a record of the working adds is the part that never reached the file. The closest evidence is about chat transcripts rather than editor replays. In one study, engineering students were allowed to use ChatGPT during take-home open-book exams and had to submit their interaction transcripts alongside their solutions 5. The researcher found three progressive patterns of use: answer retrieval, guided collaboration, and critical verification 5. Some of the strongest evidence of reasoning appeared where students evaluated incorrect or incomplete AI responses 5. The paper concludes that correctness of final answers alone may no longer provide sufficient evidence of comprehension 5.
That is one qualitative study of students in a course, with no control group and no comparison of grades. It says nothing about whether work read this way can be graded reliably, or whether it holds when the stakes are a job rather than a course mark.
What it shows is smaller and still useful: permitting the tool and requiring the working makes the working visible.
Employers are hearing a version of this from the other chair. How employers are told to run a coding interview with AI open describes the round as an audit rather than a build. What gets written down there is what was tested against something outside the chat, what got refused, what the candidate kept for themselves, and what changed in the final call.
How Should You Work When the Session Is Recorded?
Read the instructions first, then work in the editor instead of around it. The test's own rules and settings decide everything downstream: whether an assistant is allowed, whether pasting is expected, whether anything is recorded at all. After that, the ordinary way of working is the one a replay shows most plainly. Sketch, run it, watch it fail, fix it.
Working in the open is not a way to disguise help a test forbids. Where a test bars an assistant, the honest options are to work without one or to ask the recruiter about the rule before you start.
Leave the correction in.
A wrong approach that you ran, watched fail and replaced is direct evidence in a coding assessment that a decision was made. Deleting the trail leaves a file that could have come from anywhere, and the measured detectors cannot tell the difference either.
If the test permits an assistant, treat the prompts as part of the submission. CodeSignal's AI-assisted assessments hand the hiring team the full transcript of candidate-AI interactions along with the replay 3. A first prompt asking for the finished answer is as visible in that record as anything else in it.
If a flag comes back, the flag is not the finding. HackerRank's knowledge base puts its model's overall precision at 85%, meaning flagged sessions are correct 85% of the time, and says human oversight remains crucial because of it 1. By the vendor's own number, roughly one flagged session in seven is wrong. That figure is self-reported, with no test set, period or recall figure published beside it.
If the assistant conversation is requested afterwards, what to send when you are asked for your AI chat log covers what to send and why the corrections stay in.
Work you could describe out loud a week later is the version you can walk someone through in any follow-up.
Common questions
Can a recruiter watch me code in real time?
Not in what HackerRank and CodeSignal publish about assessments, which is playback after the fact: HackerRank tells recruiters they can replay keystrokes to see why a candidate was flagged, and its Screen page lists full session replay among the signals an employer can switch on. CodeSignal ships a session replay with every AI-assisted assessment. A live interview with someone on the call is a different format. Whether any of it is turned on for the test you were sent is the employer's choice, and neither page says how often a human actually watches one.
Does every coding assessment record a replay?
No. Replay is a feature an employer turns on, and the vendor pages describing it do not say which pieces are on by default. Some assessments record nothing beyond the submission. Others record video, audio and screen activity for the whole session, which is what CodeSignal's proctoring page describes, and check a government-issued photo ID against the person on camera. The invitation, the setup screens, any consent prompt before the test and a question to the recruiter are the places to find out which one you are in.
Can they tell if I pasted code from a chatbot?
Not reliably from the file. An ICSE 2025 study tested five AI-content detectors plus one built specifically for code and found accuracy mostly below 0.6, which the authors call ineffective. What a platform can see is different: HackerRank says its model looks at code-writing patterns, time taken, copy-paste activity and tab-switching, and that human oversight remains crucial because the model is not always right. A block that appears in one step is visible in a replay. None of that is proof of where the code came from, and whether pasting from a chatbot is allowed at all is the test's own rule, which is the thing to settle first.
What does it mean if my attempt is flagged for AI plagiarism?
On HackerRank, it means the vendor's model classified the session as suspicious, at High or Medium, and the flag goes into the report the hiring team reads. HackerRank puts the model's overall precision at 85%, meaning flagged sessions are correct 85% of the time by its own account, and tells hiring teams that human oversight remains crucial. The vendor also says recruiters can replay the keystrokes to see the reason. What an employer does with a flag is the employer's decision, and no appeal route is published.
Should I write messy code on purpose so the replay looks natural?
No. Performing a struggle wastes the time the assessment gives you, and a staged trail misrepresents the work as much as a hidden one does. Work the way you actually work: sketch something, run it, fix what breaks, and leave the correction in the history. That produces a real trail because it is one. If the worry underneath the question is that honest work looks suspicious, the better move is to read the test's rule on AI first, then follow it.
If the assessment allows AI, is my conversation with it part of the record?
On CodeSignal's AI-assisted assessments, yes. The vendor says every one of them comes with a full transcript of the candidate's interactions with the AI, along with a session replay, so hiring teams can see what was asked and how the suggestions were used. Treat the prompts as part of what you are submitting. It runs in both directions: a prompt asking for the finished answer sits in the record, and so does the exchange where you push back on something the model got wrong.
References
- 1. AI Plagiarism Detection support.hackerrank.com Supports the keystroke replay recruiters are told they can run on a flagged attempt, the 85% overall precision and the human-oversight instruction, the coding-questions-only scope, and the consent prompt before an enabled test.
- 2. HackerRank Screen (product page) hackerrank.com Supports full session replay sitting among the platform's integrity signals, the plagiarism model, AI proctor and image-analysis pieces, and the per-test rule this article tells the reader to look for.
- 3. Introducing: AI-Assisted Coding Assessments and Interviews codesignal.com Supports the claim that an AI-assisted assessment ships a full transcript of candidate-AI interactions plus a session replay, which is why this article treats the prompts as part of the submission.
- 4. Cheating Detection & Fraud Prevention codesignal.com Supports what a proctored session can record (video, audio, screen), the government-ID match passed to the hiring team, and the Suspicion Score routed to human reviewers.
- 5. Reimagining Assessment in the Age of Generative AI: Lessons from Open-Book Exams with ChatGPT arxiv.org Supports the three observed patterns of AI use and the finding that the strongest evidence of reasoning appeared where students evaluated incorrect or incomplete AI output.
- 6. An Empirical Study on Automatically Detecting AI-Generated Source Code: How Far Are We? arxiv.org Supports the claim that measured detectors, including one built for code, sit mostly below 0.6 accuracy, and that the researchers' own 82.55 F1 model is not a purchasable product.
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.