In a live Codility interview where AI is allowed, the AI you're using may have been secretly told to hand you broken code, and the interviewer is watching to see if you catch it. Every prompt you type and every suggestion you accept goes into the report. So the test isn't whether you use AI. It's whether the record shows you checking what it gave you.
What does Codility record when AI is allowed?
All of it. For every task where AI is turned on, Codility saves which models you used, how many interactions you had, every prompt and answer, and every command and code change. Autocomplete usage now goes into AI Usage Insights too, so accepted inline suggestions count, not just chat messages.
So the plain answer to "can interviewers see my AI prompts" is yes, every one. The report captures:
- Models used. Codility's environment offers Claude Code and its own AI Copilot, with models including Claude Sonnet 5.5 and GPT-6.1 Sol.
- Interaction count. How many times you went to the AI on each task.
- Full transcript. Every prompt you typed and every answer you got.
- Commands and code changes. What you actually ran and edited afterward.
- Autocomplete acceptance. How often you hit Tab on a suggestion.
AI can also be on for one task and off for the next in the same assessment. Codility pitches this as a way to see how a candidate builds with AI and how they handle the fundamentals without it. Expect at least one stretch where you're on your own.
How does the hidden-bug feature work?
Interviewers get an AI Instructions tab where they can tell Claude Code or the AI Copilot to slip a small bug or a poor coding practice into what it gives you. Candidates never see that tab. The interviewer then watches whether you review the output or paste it.
A "poor coding practice" covers more than bugs. Think of an unhandled edge case, a swallowed exception, an O(n²) loop where a hash map was the obvious choice, or a hardcoded value that should be a parameter. The code might pass the visible tests and still be the thing you were supposed to flag.
Codility isn't alone here. According to a secondary write-up of Meta's AI-enabled round, the AI tools provided might introduce subtle errors on purpose, testing what Meta reportedly calls "critical verification." AI-allowed rounds are spreading more broadly too. One review read the official candidate AI guidance of 32 employers on September 29, 2026: 24 restrict AI during live interviews, and 10 describe at least one round or task where AI is allowed or expected. Two examples:
- Canva replaced its computer science fundamentals screen for backend and frontend roles with an AI-assisted coding interview.
- Google announced it will allow Gemini in its code comprehension round starting in H2 2026.
24 plus 10 is more than 32, so some employers run both kinds of round. Before each stage, check which one you're in. Don't assume the last round's policy carries over.
Why the next new chat is the one to distrust
Claude Code and the AI Copilot pick up hidden instructions at different times, and that difference sets a trap late in the session. Per Codility's support docs, Copilot applies the interviewer's instructions immediately as context. Claude Code applies them as rules starting from the next new conversation.
| Tool | When hidden instructions apply | What that means for you |
|---|---|---|
| AI Copilot | Immediately, as context | Any answer can carry the planted issue |
| Claude Code | From the next new conversation | Early chats can be clean, later ones not |
| Either tool | Instructions tab is hidden from you | You never get told a trap is active |
Here's how that plays out. Claude Code gives you three solid answers in your first conversation, then you start a fresh chat for the last task. That new chat may be the first one running under the interviewer's rules. It's easy to review less once the tool has earned your trust, and that's exactly when the risk goes up.
The fix: review late answers line by line, just like the first one. Being right earlier tells you nothing about whether the next answer is clean.
Treat every AI answer like a pull request from a coworker you've never worked with. Especially the last one.
Why the log is your alibi, not the trap
A full prompt log protects you more than it exposes you, because it replaces guessing with evidence. Kyle & Co's October 14 report, "The State of Interview Intelligence," shows how unreliable the guessing is.
Kyle & Co surveyed 257 employers and 174 job seekers in August, nearly all in the U.S. Employers rated their own ability to spot secret AI use in live interviews at 8.1 out of 10. Meanwhile:
Kyle & Co, October 2026, from a sample of 174 job seekers and 257 employers.
The gap shows up elsewhere too. Employers gave their own interview experience a net promoter score of +48. Job seekers gave it +3. The report also notes that employers often lack clear rules separating preparation from inappropriate live help.
Put that together. An interviewer who rates their detection at 8.1/10 and works without clear rules is making a judgment call about you. On an AI-allowed task, the log takes that call away. The question stops being "did they cheat?" and becomes "how well did they use the tool?" The second question is one you can win.
That's also why going quiet on an AI-allowed task is the riskier play. A candidate who never touches the AI but produces polished code fast can look more suspicious than one whose transcript shows a sensible back-and-forth.
What reviewers see first: the summary, not your transcript
Reviewers form their first impression from the summary stats, so the shape of your AI use matters before anyone reads a single prompt. Codility's own guidance to reviewers is to scan the summary first and open the transcript only when something stands out.
Two candidates with the same final code can look completely different at a glance:
| Pattern in the summary | What a reviewer infers |
|---|---|
| 1 interaction, huge prompt, answer accepted whole | Copy-paste; opens transcript to confirm |
| High autocomplete acceptance, few edits after | Typing on autopilot |
| Several scoped prompts, edits between each | Directing the tool, checking output |
| AI-off task solved cleanly, AI-on task fast | Fundamentals are real; AI is leverage |
You want the summary to look like the third row. That doesn't mean padding your interaction count with throwaway prompts. It means breaking the problem down the way you would anyway, and leaving visible edits in the record between AI turns.
Code review is the skill under test, and the one engineers list most
The hidden-bug feature tests code review, and code review is the skill engineers already claim most. In Refolk's index of professional profiles, US profiles titled Software Engineer list Code Review far more often than any AI coding tool.
| Skill listed | Profiles | Top employers in a 25-profile sample |
|---|---|---|
| Code Review | 3,471 | Google (3), Meta, Dropbox, Adobe, Zoox |
| GitHub Copilot | 496 | Microsoft (8), Google, Intuit, Meta (2 each) |
| Claude Code | 244 | Apple (5), Meta (4), Microsoft, Google (2 each) |
These counts cover the slice of the index matching each query, so read them as ratios rather than market size. The employer columns come from small samples and only point in a direction.
Refolk's index. Code Review is also listed 14.2x as often as Claude Code.
"Code Review" is easy to claim and hard to verify from a resume. Copilot and Claude Code together come to 740 profiles, about 21% of the Code Review count. In this slice, engineers say they review code far more often than they say they use AI tools. A planted bug inside an AI tool tests both claims in one session.
If Code Review is on your resume, the assessment is where it gets checked, so the resume a company sees has to match what you'll actually show in the session. Refolk writes your resume from your own work history and tailors it to each posting, so the review and AI-tooling claims a given employer sees are ones you can back up on a recording, not boilerplate.
How to use AI in a technical assessment: a working protocol
Work like a reviewer directing a junior engineer: small asks, read everything, change what's wrong, and say why. Here's the sequence I'd use in any Codility AI Copilot interview or Claude Code task.
- Read the problem without AI first. Spend two or three minutes writing down the inputs, outputs, and edge cases yourself. That gives you something to check the AI's output against.
- Prompt for pieces, not the whole solution. Ask for a parser, then a helper, then a test. Scoped prompts make defects easier to spot and produce the summary pattern reviewers trust.
- Put your constraints in the prompt. "Handle empty input and duplicate keys, O(n) time" is a prompt that shows judgment. "Solve this" is not.
- Read every line before running it. Look for off-by-one bounds, missing null checks, swallowed errors, mutable defaults, and complexity that doesn't match the constraint.
- Test the boundary, not the happy path. Edge cases are where code can pass the visible tests and still be wrong.
- Edit in the editor, visibly. Fixing the code yourself leaves a code-change trail. Asking the AI to fix its own bug without saying what the bug is leaves a weaker one.
- Narrate the catch. In a live interview, say it out loud: "This loop skips the last element, fixing that." In an async test, a short code comment does the same job.
- Go easy on Tab. Autocomplete acceptance is now counted. Accept suggestions you've read, not every grey line that shows up.
- Start new chats skeptical. Because of the Claude Code timing above, a fresh conversation late in the session gets your hardest look.
Stay inside the recorded environment. Meta says only the tools inside the interview environment may be used in its AI-enabled CoderPad rounds, and that's the safe assumption anywhere. A second AI tab outside the recorded editor turns an allowed task into a disallowed one.
AI for Business: sales, marketing, and finance candidates face the same test
Codility's AI-on assessments now go beyond engineering. AI for Business assesses AI skills for sales, marketing, and finance roles, and in live interviews the interviewer watches you work with the AI Assistant in a text editor you both share.
This group usually has the least preparation, because nobody told them a technical-style recording was coming. The rule is the same, though. What scores is visibly checking the AI's output, not polished text you never questioned. Your version of a planted bug looks like:
- A wrong number. A growth rate that doesn't match the figures, a margin calculated on the wrong base, a total that doesn't add up.
- An unsupported claim. A market stat with no source, or a customer benefit the product doesn't actually offer.
- Wrong tone for the audience. A cold email that reads like a press release, or a CFO summary buried in adjectives.
- A missed constraint. A word limit ignored, or a required call to action left out.
Fix these in the shared editor where the interviewer can see it, and say what you changed. A finance candidate who catches a bad formula in the AI's draft has shown more than one who submits a clean-looking memo.
FAQ
Will I know if the interviewer told the AI to add a bug?
No. The AI Instructions tab is hidden from candidates. With the AI Copilot the instructions apply right away. With Claude Code they apply from the next new conversation, so a fresh chat late in the interview can be the first one carrying the trap. Assume every answer might contain a planted bug or a poor practice, and review accordingly.
Is it safer to skip the AI on a Codility assessment where AI is allowed?
Usually not. If the task allows AI, part of the evaluation is how you use it, and an empty log doesn't show that skill. Kyle & Co found 39% of job seekers already feel unfairly suspected of using AI, and a recorded, well-directed session is the clearest evidence against that suspicion. Save the no-AI approach for tasks where AI is turned off, and expect at least one.
What if I miss the planted bug?
One miss does less damage than a pattern of checking nothing. Reviewers are looking at whether you reviewed the output at all: scoped prompts, tests aimed at edge cases, visible edits, and comments explaining your changes. A candidate who catches two real issues and misses the planted one still leaves a record of judgment. A candidate who pastes everything leaves a record of nothing.