The Take-Home AI Use Score, Each Part Called Use, Disclose, Ask, or Hand-Do

This guide is the scoring standard for one decision you make before starting a take-home: which parts you can do with an AI tool, which you must disclose, which need a question to the recruiter first, and which you do by hand. It is for candidates holding a brief whose AI rules are vague or missing, with a deadline this week. Work through it once and you leave with one call per deliverable part, plus the recruiter question and disclosure note those calls require, in under 30 minutes.

What are the four calls, and why score each part separately?

Every part of a take-home gets exactly one of four calls: Use, Use and Disclose, Ask First, or Hand-Do. Score each part separately because briefs routinely mix allowed and off-limits work, and one blanket rule for the whole assignment will be wrong for at least one part.

A part is one separable deliverable in the brief. Most briefs break into some mix of five: research, analysis, code, written narrative, and slides. A defense round is a live session after submission where you walk an evaluator through the work and answer questions about it.

The per-part view is not theoretical. Anthropic's company-wide guidance tells candidates to complete assessments without its model unless told otherwise, yet its performance engineering team's take-home has permitted AI tools from the start. That is one employer with two rules, each set by what the stage is meant to measure. A brief can split the same way internally: a line allowing "any tools" for the slides says nothing about the analysis behind them.

This guide covers a different job from the library's other take-home guides. Deciding whether to accept a take-home at all, and grading the finished submission, are covered elsewhere, and live rounds where AI is allowed have their own coding-round guide. This one covers the decision in between: what tool use is allowed on each part while you do the work.

Where does the AI rule live, and which wording class is it?

The AI rule for a take-home can sit in four places: the employer's company-wide candidate guidance, the recruiter's invite email, the brief, and any instruction attached to one part. You have not found it until you have read all four. Whatever you find falls into one of five wording classes, and the class is the first thing you score.

A wording class is the category the employer's AI language falls into once you strip it down to its default.

Notes on the evidence behind each row:

• Prohibited unless told otherwise is the most common published default among large employers. Amazon asks candidates to "please do not use GenAI tools during your interview unless explicitly permitted." PwC tells candidates not to use these tools during any assessment. Microsoft asks candidates to show their own skills without outside help unless explicitly permitted.
• Permitted with disclosure comes from a public policy tracker whose entry cuts off the employer's name. Treat the pattern as real but the employer as unverified.
• AI expected at Canva applies to backend, frontend and machine learning candidates in a live round, not a take-home. Do not carry it over to a written assignment.
• Silent is described by Greenhouse CEO Daniel Chait as the common case. No published count of silent employers exists.

The default flips by company, not by industry. Anthropic builds an AI model and bans AI on take-homes by default. Canva requires AI in a live round. Each employer sets the rule according to the signal it wants from that stage, so the sector tells you nothing.

Silent is the hard class, and published advice splits on it. One line says to ask before starting a take-home, because that is where disclosure rules usually apply first. Another says you may use AI tools unless the brief says otherwise. My position sits between them: when the employer is silent, any part that carries the core signal or will be defended live goes to Ask First, and incidental parts get Use and Disclose. The scoring below follows that rule.

What are the six dimensions, and how do I score each one?

Score every part as Green, Amber, or Red on the same six dimensions: wording class, input confidentiality, what the part tests, defense exposure, explain-and-change, and workload realism. Each dimension proves one thing, and each one can mislead you in a predictable way.

Wording class

What it proves: what the employer actually said about this part. When it lies: when you read only the brief and miss a company-wide ban, or when you read "unless we indicate otherwise" as silence. Write down the quoted sentence. Write "silent" only after checking all four layers.

Input confidentiality

What it proves: whether a cloud AI tool may see this part's inputs at all. Look for datasets, starter code, and case packs. A published candidate NDA template commits the candidate to use the material only to take part in the recruitment process, and not to copy, share or publish it. Practitioner guidance adds that an NDA usually bars you from sending the material to a third party, which is what any cloud AI service is. I found no named employer's candidate privacy notice on this exact question, so the employer-level norm is not established. Read your own paperwork. When it lies: when you assume that switching off model training makes confidential input safe.

What the part tests

What it proves: where the evaluator's score sits. The core signal is the part the brief says it will judge, or the skill the posting leads with. Scaffolding is everything that makes the work presentable without being the judgement itself. The posting is the best outside evidence. If you applied with a resume Refolk tailored to that posting, its fit score already ranks which requirements the role weights most, which is a fast way to name the core-signal part. When it lies: when the brief is vague and you file the part you would like to automate under scaffolding.

Defense exposure

What it proves: whether you will be questioned on this part live. Unknown scores Amber, not Green. When it lies: when you assume there is no defense because the invite does not mention one. The share of take-homes followed by a live walkthrough has not been established publicly, so ask.

Explain-and-change

What it proves: whether you would pass the live check employers actually use, which one screening guide puts as: "Ask the candidate to change it and explain what moves." Score it honestly before you start. Can you produce this part at the required level without the tool, and change it on request? Requirements where Refolk's fit score shows your history is thin are where this dimension most often turns Red, because that is where you would be leaning on the tool for a skill you have not yet shown. When it lies: when you mistake understanding the tool's output for being able to produce or change it.

Workload realism

What it proves: whether the rule makes sense alongside the work. One documented case reads: "Do not use AI" in the candidate email, then a take-home that takes 12 hours to do without AI. Time limits also move as tools change. On Anthropic's AI-permitted performance engineering take-home, the time limit went "from 4 hours to 2 hours." When it lies: when you score Red because you would prefer to use AI. A tight deadline is Amber. Red means the rule and the workload cannot both be true.

How do the scores turn into one call per part?

Apply the rules below in order, stop at the first one that matches, and that rule's call is the part's call. The order matters: confidentiality and contradictions come first, because they decide the call whatever the other dimensions say.

Here is an illustrative case, not a real brief. It is an analyst take-home where the careers page and invite say nothing about AI, the brief says "use any tools you like for the slides", the dataset arrives under a recruitment-use-only note, and a live walkthrough is confirmed.

Four parts give three different calls. A blanket rule would have handled at least two of them wrongly. Both Ask First parts go into one recruiter message.

The 30-minute procedure, step by step

The procedure takes about 25 minutes of your time before you start work, plus the recruiter's reply time, plus five minutes at the end for the disclosure note. Each step has a definition of done, so you can tell when to move on.

Why does the defense round set the stakes?

Take-homes are distrusted as a signal, so evaluators move the real judgement into the live walkthrough. That makes the parts you will defend live the place where unexplainable AI use costs the most. How common take-homes are depends on who was measured, which is why the published figures disagree.

The last row is the one that matters for this framework. Only 12% of recruiters named work-sample or take-home projects as the format that best reflects on-the-job ability. In the same Karat survey that counted take-homes in 45% of US engineering loops, 71% of leaders said AI is making technical skill harder to assess, and 38% of loops allowed AI use. When the artifact is distrusted, the score moves to the conversation about it.

Two limits are worth stating. The share of take-homes followed by a live defense has not been established publicly; one widely shared hybrid figure has no traceable primary source, so I do not use it. And each survey edition re-measures a different population, so check the current edition before quoting any of these numbers, and quote the population along with the percentage.

What goes in the recruiter question and the disclosure note?

Send one written question that covers every Ask First part and offers fixed options to choose from. After the work is done, write a disclosure note naming the tool, where you used it, and how you checked the output. Both fit on one screen.

The question needs fixed options because the person answering it probably has no ready policy to quote.

The employers that appear most often in each segment come from 25-profile samples, so they are indicative only. They range from staffing firms to large tech companies, which means the reader of your message could be an agency recruiter just as easily as an in-house one. An open question such as "what's your AI policy?" asks that person to write a policy on the spot. A lettered list asks them to tick one box.

Subject: Quick check on AI tools before I start the [role] take-home

Hi [name],

Before I start the take-home, I want to use AI tools the way your team intends. The brief doesn't cover this for two parts, so could you tell me which option applies to each?

Parts:
1. [e.g. market research]
2. [e.g. written recommendation]

Options:
A) No AI tools on this part
B) AI tools allowed, and I note which tool and where
C) AI tools allowed, no note needed

Two related questions:
- May the provided [dataset / starter code / case pack] go into a cloud AI tool, or should it stay on my machine?
- Will I need to explain the submission live?

Until I hear back, I'll do those parts by hand.

Thanks,
[your name]

For the note, employers ask for very little. Published employer-side wording asks candidates to disclose which tools they used and roughly where, or to say what they used and how and be ready to walk through the result without them. The strictest documented format is academic rather than from an employer. It covers the tool and who makes it, what you used it for, where in the work you used it, how you checked or changed the output, and your prompts. I found no employer that publicly requires prompt logs, so keep yours and offer it on request.

Why disclose at all? Public data on how evaluators treat disclosed versus undisclosed AI on take-homes does not exist. The nearest evidence is about resumes, so this is an inference. In one survey, 52% of hiring managers accepted AI for drafting and proofreading support, while 19.6% would reject a candidate with an AI-written resume or cover letter and another 20% saw heavy AI reliance as a red flag. A second-hand report of another survey put the share of hiring managers who think candidates should disclose AI help at 79%. A note moves you out of the hidden-use group and into the accepted-use group.

## AI use

Tools: [tool name], made by [company]

Where I used it:
- [Part]: [task, e.g. first draft of the slide outline]
- [Part]: [task, e.g. debugging the data loader]

Done without AI: [parts called Hand-Do], and [any provided inputs kept out of AI tools for confidentiality]

How I checked the output: [what you verified, rewrote, or tested]

I can walk through any part of this submission without the tools, including changing it live. My prompt log is available on request.

How does this call go wrong?

Most bad calls come from reading the rule too narrowly, assuming a permission covers more than it does, or betting on detection being weak. Each failure below has a check that catches it before you start.

Misreading the wording class

The common case is reading "unless we indicate otherwise" as silence, or reading only the brief. This false positive is expensive: you assume the brief allows AI because it does not mention it, while the company-wide page bans AI by default, as Anthropic's does. Check: read the careers page as well as the brief, and keep a copy of what you read.

Carrying permission across rounds

A take-home that allows AI tells you nothing about the onsite.

The questions to ask at each later stage are covered in the AI policy questions guide for every interview stage.

Treating a setting as confidentiality

Turning off training does not make confidential input safe. In one practitioner's words: "If the NDA prohibits disclosure to third parties or taking material outside the organization, changing a setting won't solve that." Check: score confidentiality from the paperwork, never from the tool's settings.

The brief contradicts its own workload

"No AI" next to a 12-hour task is a documented pattern. Using AI quietly because the rule seems unrealistic is the failure. Check: score the affected parts Ask First (rule 2) and let the employer resolve the contradiction in writing.

Counting on weak detection

Detectors really are weak. A study by Weber-Wulff and colleagues found no tool above 80% accuracy, and OpenAI's own classifier caught 26% of AI text, falsely accused 9% of human text, and was withdrawn. But detectors are not how employers verify. In a Greenhouse survey, 65% of hiring managers said they had caught applicants using AI deceptively, and a Karat estimate puts candidates using AI in code tests even when it is banned at about 80%. Evaluators assume AI use and check for it live, by asking you to change the work and explain what moves. Check: if a part is defended live and scored Red on explain-and-change, it is Hand-Do whatever the wording allows.

Being flagged for work you wrote yourself

False flags fall unevenly. Detectors treat low-perplexity, plainly written prose as machine-made (perplexity measures how predictable each word is), so non-native writers who did their own work get flagged more often.

No audit of detectors on take-home submissions specifically has been published. The rerun found 23.1% on TOEFL essays against 0% for the native comparison group. Check: keep dated drafts, name your Hand-Do parts in the disclosure note, and be ready to walk through the work. A live walkthrough clears a false flag faster than any argument about detectors.

Citing the 68% statistic

A widely repeated claim says 68% of teams use take-homes and attributes it to a CoderPad survey. It contradicts that survey's own 21.06% figure. Check: do not cite it, in your own reasoning or to a recruiter.

Before you start: the final check, and keeping the call current

The call is finished when every part has one call, every Ask First part has a written answer or is being done by hand, and the evidence for each call is saved. Policies change, so the call belongs to this brief on this date. Anthropic, for example, has replaced its candidate guidance at least once with staged guidance.

To keep the call current, save a dated copy of every policy page you relied on, because the page you read this week may not be the page the employer cites later. Re-run the procedure at each new round of the same loop rather than carrying calls forward. If the recruiter's answer changes a call mid-assignment, update the log and the disclosure note to match.