The Take-Home AI-Use Call, Scored to Use Openly, Disclose, or Do Unaided

You have a take-home or an online assessment in front of you, and you need one decision: can you use AI on it, and what exactly do you have to say about it. This guide is for candidates mid-assignment who want a repeatable call, not a shrug. It turns the stated policy and the assessment type into a scored judgement that lands on one of three actions and gives you the disclosure wording to go with it.

Generic advice stops at "use AI transparently" and leaves the two hard cases unhandled: a silent policy, and a policy that permits AI yet demands you own every line in the debrief. Below is a procedure for both.

What the five dimensions are and why they decide the call

The call turns on five dimensions: the stated posture, the assessment type, how much authorship AI would take, whether a debrief is scheduled, and any legal or data constraints. Score each, and the scores point to one of three actions: use openly, disclose, or do unaided.

Two of these dimensions do most of the work. The stated posture tells you the rule you are living under. Whether a debrief is scheduled tells you how high the ownership bar sits regardless of posture, because the conversation about the take-home is the test, not the artifact. The other three are tie-breakers that pull the call toward caution when the first two are ambiguous.

That gap is why this is a judgement call and not a lookup. The information you need is usually not published, so you score what you can read and ask for the rest.

The five postures a stated policy can take

Published policies cluster into five postures, and your first job is to decide which one you are reading. Each maps to a different starting point for the score.

GitLab says transparent AI use is welcome in technical assessments and that a take-home may use AI for research if you cite it, answer in your own words, and disclose the tools you used. Code.org permits candidates to use AI on take-homes and interviews provided you disclose your usage with the interview team. Anthropic asks candidates to complete take-home assessments without Claude unless indicated otherwise, and says it will be clear when AI is allowed. PwC tells candidates they should not use these tools during any assessment.

The silent posture is the common one and the one blog posts duck. On a sample of 1,765 job descriptions, only one company explicitly allowed AI in take-homes and none explicitly banned it. An open tracker of 20 companies found 10 ban candidate AI use and 8 allow or require it, which tells you that among companies that bother to publish, the split is roughly even and you cannot guess your way to the right posture.

The default when the instructions say nothing

When the policy is genuinely silent or ambiguous, the documented default is conservative: ask first, and if you cannot get an answer, produce work that reflects your independent capability. This protects you against the risk that AI use would violate an unstated policy, and you can note in the submission that you produced the work independently.

The inverse rule governs what you volunteer. When nobody has asked and your use was light and bounded, describe what you built and what you can defend, and leave the authorship confession out. Over-disclosing a prep-only use reads as either naivety or a flag you planted on yourself.

So the "ask first" step and the "do not over-disclose" step are not in tension. Ask when the rule is unknown. Disclose when the rule or the brief requires it. Stay quiet when neither the rule nor a request is in play and the help was minor.

The one message that resolves a silent policy

When the policy is silent, send the recruiter a single calibration line before you write anything. It is the highest-leverage move in the whole procedure, because the information genuinely is not published and asking signals professional behaviour rather than doubt.

Hi [name], I want to make sure I approach this appropriately. What is your policy on AI tool use during the assignment, and is there anything you would like me to note about how I worked?

Recruiter guidance beats guesswork. Before writing code, confirm whether AI use is allowed and in what form. Use the company's wording back to them, not your own assumptions, and do not start the clock on a timed assessment until you have asked where asking is possible.

Refolk tailors your resume and cover letter to each posting from your own history, which frees the hour you would otherwise spend rewriting application material and lets you spend it on the calibration and the debrief rehearsal that actually move this decision.

Scoring the five dimensions and landing on an action

Score each of the five dimensions, then read the dominant signal. The stated posture and the debrief carry the most weight; the other three pull toward caution. This table is the scoring sheet to keep open.

How to read it: if any dimension points to "do unaided," that signal dominates, because a single breach cannot be cured later. If no dimension points to unaided and at least one points to disclose, the action is disclose. Only when every dimension points to "use openly" do you use openly without a line.

The debrief deserves its heavy weight. Across 19,368 live interviews analyzed between July 2025 and January 2026, 38.5% of candidates were flagged for AI-assisted answers, a rate that tripled from 9% to 45% in one three-month stretch. Take-homes feel safer than live rounds, but a scheduled follow-up is a human detector, and it does not need software to catch an answer you cannot explain.

The procedure, start to finish

Run these seven steps in order. The disclosure step is conditional: you write and place a line only if the rule or the brief asks for it, otherwise you hold it.

Step five is where discipline pays off. If use is permitted, deploy AI for tightening tests, improving naming, drafting documentation, and checking edge cases, not for asking for the whole answer. The decision log is your insurance for step seven: it records where you accepted, rejected, or modified a suggestion, which is exactly what a debrief probes.

How to write the disclosure line

A compliant disclosure names three things in a sentence or two: which tool, which decisions were yours, and what you checked before you shipped it. A line missing any of the three reads as vague and invites the questions you were trying to prevent.

I used [tool] to help with [specific bounded support, e.g. drafting the test scaffold and suggesting edge cases], then I [your human work, e.g. wrote the core logic, rejected two suggested approaches, and refactored the data layer] and checked the final version for [accuracy, correctness against the brief, and readability].

Vague fails on purpose-built grounds: a line like "I used AI to help" raises more questions than it answers, because a marker cannot tell whether you mean spell-check or a ghost-written paragraph. Specific wins because it demonstrates the judgement the employer is actually scoring. GitLab, for one, evaluates how you appropriately leverage AI and expects you to discuss how you collaborate with it, so a precise line is evidence in your favour, not a confession.

A specific disclosure line is a cheap differentiator, because almost nobody writes one.

The behaviour is rare where it is required. A survey of 500 US job seekers found 54.2% would never bring AI up with an employer, and only about one in eight always discloses. Set against near-universal adoption, a precise line marks you as the candidate who can be trusted with the tool.

The two adoption figures differ because the surveys differ in sample and wording, so read them as a range rather than a single truth. Either way, the share who disclose is a small fraction of the share who use.

Why employers increasingly permit-with-disclosure rather than ban

Employers testing genuine AI-collaboration skill are assessing a scarce trait, which is why the permit-with-disclosure posture is spreading. In Refolk's index of professional profiles, the talent that can do this well is thin and concentrated.

In Refolk's index, 55,005 US profiles titled Software Engineer list Python, but only 1,707 list Artificial Intelligence and 1,618 list Prompt Engineering, each roughly a 3% slice. The AI and prompt-engineering pools are nearly equal, a ratio of about 1.05, which suggests the two skills travel together. The scarcity is also geographic: Refolk's index shows 1,707 US Software Engineers tagging AI against 173 in the UK, a derived 9.9x gap. When the skill is this rare, banning the tool in assessment throws away the signal an employer most wants, so they ask you to use it openly and show your judgement instead.

That is the strategic read behind the posture table: a prohibited policy is often about integrity on a specific role, while a permit-with-disclosure policy is the employer actively trying to measure how you collaborate with the tool.

How this call goes wrong

The call fails in predictable ways, and most failures are a dimension misread rather than a bad intention. Each one below has a check you can run before you submit.

• Reading silent as allowed. No ban is mentioned, so you use AI freely. Check: absence is not permission; on real banks most postings never mention policy, so default unaided or ask.
• Vague disclosure. "I used AI to help" feels honest but tells a marker nothing. Check: does your line name the tool, your decisions, and your review? If not, rewrite it.
• Over-reliance you cannot defend. A clean submission paired with a hollow debrief. Check: can you explain every line unaided? Submitting AI-generated work you cannot explain is a fast path to rejection at any company that runs a thorough follow-up.
• Over-disclosing when unasked. A volunteered confession on prep-only use. Check: was disclosure required or requested? For light, bounded help you do not need to proactively tell a recruiter.
• Disclosure as a fix for a breach. Using AI where banned, then disclosing. Check: a disclosure sentence does not fix a rule breach; if the posture is prohibited, do the work unaided.
• Treating all rounds alike. Applying take-home rules to a live or proctored round. Check: the rules may differ by type, so ask before you start and use the company's wording.
• Assuming permitted means no ownership expected. Even GitLab requires you to explain your prompts and technical decisions and to avoid presenting generated content as your own. Check: permitted raises your ownership bar, it does not remove it.

The through-line of these failures: the artifact is not where candidates lose. The take-home is not the test, the conversation about the take-home is the test, and anything you cannot explain becomes a liability the moment somebody asks a follow-up question.

Before you submit

Run this checklist against the assignment in front of you. If any item fails, do not submit yet.

Keeping the call current

Re-run the posture read for every employer, because policies are published unevenly and change without notice. The structural gap is the thing to remember: 80% of US candidates report employer policies as vague, rare, or absent, and 63% of job seekers have now been interviewed by an AI, up 13 points in six months. The direction of travel is more AI in the process and more scrutiny of how you use it, not less.

Two mechanisms are worth watching rather than any fixed number. First, detection in live rounds is climbing fast, so a concealment strategy that works today degrades quickly; the defensible-disclosure strategy does not. Second, state law is moving: Illinois already requires advance notice, a general explanation, and consent before AI analyzes a recorded video interview, which means a video round can carry legal constraints a take-home does not. Re-check the posture, re-check the type, and re-score. The five dimensions do not change; the values you plug into them do.