AI Detection Take Home Assignment: What to Document

If you used AI on a take-home assignment and now worry about detection, the fix is documentation, not concealment. If you’re a hiring manager, the fix is asking better questions. Running a detector on a take-home task is the weakest evidence you can gather and the easiest to get wrong. AI detection take home assignment questions sit where two things people care about collide: getting hired and hiring well. Neither side benefits from a probability score treated as proof.

Here’s the honest version of what to do next, for both sides of the desk.

Key takeaways

• Candidates: document your process as you go — prompts, drafts, edits, decisions — so your reasoning is visible.
• Hiring managers: ask about the work itself (trade-offs, dead ends, why a choice was made) instead of running a detector.
• Detector scores are probability estimates, not authorship verdicts. They flag human writing too.
• A clear take-home policy up front prevents almost every dispute later.
• Disclosure is usually safer than silence. Most rubrics reward judgement, not tool abstinence.
• If you do run a detector, treat it as one weak signal among many, never as a gate.

Why is AI detection in hiring so unreliable?

Detectors don’t read minds. They estimate how statistically predictable a piece of text looks, then convert that estimate into a score. Clean, uniform, well-structured prose tends to read as more predictable, which is exactly what a competent candidate produces under deadline pressure. That’s the trap.

We ran our own measurement on 2026-09-19 using the DeAIze detector (local ensemble, lite tier) at a 30% flag threshold. On a human-written sample of n = 46 hand-written DeAIze guides, the AI score mean was 27%, median 25%, range 0-55%. At that threshold, 17 of 46 were flagged. These were real guides written by people, not model output. More than a third crossed the line.

The same run scored unedited model output: n = 12 passages generated from fixed prompts and not edited. Mean 35%, median 34%, range 25-49%.

That’s the whole problem in two lines. Human writing and unedited model output overlap so heavily on a detector’s scale that a single score can’t separate them reliably. If you’re a hiring manager considering a detector as a screening gate, you’re about to reject roughly a third of honest candidates to catch some fraction of dishonest ones. If you’re a candidate who wrote every word yourself, you can still be flagged.

What should candidates document about a take-home assignment?

The goal isn’t to prove you’re innocent. It’s to make your thinking legible, so a reviewer can see the work behind the artifact.

Document these as you go, not after:

1. Your starting point. What did you read, review, or research before writing? Name the sources, the competitor examples, the internal docs you were given.
2. Where you used AI, if you did. Be specific: “used a model to outline section three,” “asked for a critique of my first draft,” “generated three alternative headlines and picked one.” Vague disclosure reads as evasion.
3. What you changed and why. Keep the messy middle. A draft that shows a paragraph moving from a generic claim to a concrete one is evidence of judgement.
4. Your dead ends. The approach you abandoned, the data that didn’t support the angle you wanted. This is the single hardest thing to fake and the most convincing thing you can show.
5. Your open questions. What you’d ask the team on day one. It signals you treated the task as real work, not an exam.

A short process note — a paragraph or two, plus a draft history if you have one — is enough. You don’t need to narrate every keystroke.

If you’re writing in a second language and worry that your clean, formal English will look machine-made, the risk is real and worth understanding. Why an AI Detector Flags Non Native English Writers explains the specific patterns that trip detectors, and AI Detectors Penalise Non Native Writers: What to Document covers exactly the evidence trail that protects you.

What should hiring managers ask instead of running a detector?

Replace the score with a conversation. Fifteen minutes of structured questions tells you more than any detector output, and it can’t be gamed by rewriting a paragraph.

Ask questions that only the person who did the work can answer:

• “Walk me through your first draft. What did you cut?”
• “Which part of this are you least confident about, and why?”
• “What would you do differently with another day?”
• “You chose approach A over approach B. What would have changed your mind?”
• “Where did you get stuck, and how did you get unstuck?”

Then ask the direct question, without accusation: “Did you use AI anywhere in this, and if so, where?” How someone answers that is more informative than the answer itself. A candidate who says “yes, I used it to pressure-test my structure, and here’s what I rejected” has just demonstrated exactly the judgement you’re hiring for.

What you should not do is run a detector and then present it as evidence. A score from GPTZero, Turnitin, Copyleaks, Originality.ai or Sapling is a probability estimate about text patterns. It is not a statement about who wrote what. Treating it as one is how hiring processes end up in disputes they can’t win.

If you want to see how a detector actually reads a piece of writing — sentence by sentence, not as a single verdict — run the take-home text through the DeAIze AI detector before you draw conclusions from it. The sentence-level view shows which specific lines carry the signal, which is usually the moment people realise how blunt a whole-document score really is.

What are the limitations of AI detectors in hiring contexts?

Four limits matter most.

They don’t measure authorship. They measure text predictability. A human who writes cleanly and a model that writes cleanly look similar on that axis.

They disagree with each other. Run the same passage through two detectors and you’ll often get two different verdicts. Why Two Tools Disagree walks through the reasons — different training data, different thresholds, different definitions of “AI-like.”

They’re sensitive to length and genre. A short, structured take-home task gives a detector very little to work with. Business writing, which favours plain sentences and parallel structure, reads as more predictable than a personal essay.

They can’t be audited the way a reference check can. If you reject a candidate on a score, you have no defensible record of why. If you reject them because they couldn’t explain their own trade-offs, you do.

Our own measurement illustrates the overlap: human guides averaged 27% while unedited model output averaged 35%. Those distributions sit close enough together that a threshold line drawn anywhere sensible will catch plenty of both.

Signal	What it tells you	Reliability in hiring
Detector score	How predictable the text looks	Weak — flags human writing, misses edited AI
Process note from candidate	What they did and why	Strong, if specific
Live walkthrough of the work	Whether they understand their own output	Strongest
Follow-up questions on trade-offs	Depth of judgement	Strong
Draft history	Real iteration vs one-shot generation	Strong
Writing style alone	Very little	Misleading

How do you set clear expectations for a take-home task?

Put the policy in the brief. Not in a handbook nobody opens — in the document the candidate receives.

A workable policy answers four things:

1. Is AI allowed at all? Yes, no, or yes-with-disclosure. Pick one and say it plainly.
2. If allowed, where’s the line? “Use it for research and editing. Don’t generate the final deliverable with it” is a clear line. “Use your judgement” is not.
3. What should be disclosed? Ask for a short note: which tools, which parts of the task, what you changed.
4. How will the work be assessed? Say that you’ll discuss the submission live, and that the discussion is part of the evaluation.

That last point does most of the work. When candidates know they’ll be asked to defend their choices, the incentive to submit unexamined output drops sharply — without anyone needing to police anything.

On the candidate side, read the brief for those four answers. If the policy is silent, ask before you start. One email — “Should I disclose any AI assistance, and in what format?” — removes the entire risk. Hiring managers overwhelmingly read that as diligence.

If you’re already past that point and worried about how your submission reads, cleaning up machine-made patterns in your own drafts is the honest route: it makes your actual voice clearer rather than disguising someone else’s.

What are the ethical considerations for both sides?

For candidates, the line is straightforward: the ideas, analysis and conclusions have to be yours. Using a model to sharpen a sentence or challenge your structure is a tool decision. Submitting a generated deliverable as your own reasoning is misrepresentation, and it will surface the moment someone asks you a follow-up question in the interview.

There’s a practical argument here too. If you can’t explain a paragraph, you don’t want it in your submission, because you’ll be sitting across from someone asking about it.

For hiring managers, the ethical obligation runs the other way. You owe candidates a process that doesn’t punish them for writing well. That means no silent detector screening, no rejection without a conversation, and no treating a score as a finding. It also means consistency: if you run a detector on one candidate, run it on all of them, and apply the same scepticism to every result.

One more consideration. Detector output can carry bias against non-native English writers, whose formal, textbook-correct prose often reads as statistically smooth. A process that quietly filters those candidates is a process that narrows your pipeline for reasons nobody wrote down.

A worked example

Here’s an illustration, not a real submission. A candidate submits a two-page market analysis. The prose is clean, the structure is tidy, the recommendations are sensible. A detector returns a high score. The hiring manager pauses.

The weak move is to reject. The strong move is a 15-minute call:

• “Your recommendation was to enter the mid-market segment. What made you rule out enterprise?”
• “You cited three competitors. Which one worried you most?”
• “If the budget were halved, what would you cut first?”

The candidate who did the work answers these with specifics and hesitations. The candidate who generated the document stalls, generalises, or repeats the document back. Fifteen minutes resolves what a score cannot, and it produces a record you could defend to anyone.

Where does DeAIze fit into this?

It fits as a diagnostic, not a verdict — and it’s most useful to the candidate, not the employer.

DeAIze is a web app with two halves: an AI detector and an AI humanizer. Detection is sentence level: you get a whole-document score and each sentence is highlighted by AI probability, so you can see which specific lines carry the signal. That matters here, because “your document scored high” is useless feedback and “these four sentences read as machine-made” is something you can act on.

Humanizing runs a closed loop: score the text, rewrite only the flagged sentences, keep a rewrite only when semantic similarity says the meaning survived, then re-score and report before and after. The site reports an average AI-score reduction of 60%+ after humanizing, best case 71% — those are the product’s own predicted scores, not an official verdict from any third-party detector. On the example paragraph used on the site, a 92% AI score becomes 4% AI.

The free tier gives you 200 words on signup, no credit card. Uploads cover .txt, .docx and .pdf, and rewriting a Word document preserves fonts, styles and tables. Detection and rewriting cover English and 9 other languages; the interface ships in 10 languages. Text is processed for the current task only, never used for training and never resold. Cross-checks run the output against real third-party detector APIs where configured — GPTZero, Turnitin, Copyleaks, Originality.ai and Sapling — instead of trusting an internal guess.

To be clear about what it isn’t: DeAIze does not guarantee any particular score on any detector, and it is not a tool for passing off someone else’s work. If your take-home submission isn’t your own thinking, no amount of rewriting fixes that.

The realistic use is narrower and more useful. You wrote the submission yourself, you used AI somewhere in the middle, and you want to know whether the final text still sounds like you. Run it, look at the flagged sentences, and decide whether each one says what you meant in the way you’d say it. Often the answer is a small edit, not a rewrite.

If you want to understand the signals before you act on them, How AI Detection Works covers the mechanics, and Is Bypassing AI Detection Cheating? A Disclosure Test gives you a clean way to decide where your own line sits.

Ready to see which sentences read machine-made?

If you’re a candidate with a submission due, the fastest useful step is a sentence-level read of your own draft. Run it through the DeAIze AI detector and look at the highlighted lines rather than the headline number. Then decide, sentence by sentence, whether each one sounds like you.

If you’d rather work on the text directly — rewriting the flagged lines while keeping your meaning intact — the draft editor is built for exactly that pass, with before-and-after scoring so you can see what changed.

If you’re a hiring manager, the useful move is different: read your take-home brief and check whether it answers the four policy questions above. If it doesn’t, fix the brief before you fix the screening process.

And if you want to know what you’d be paying for before you commit, pricing is pay-as-you-go, credits never expire, and there’s no subscription — Starter $20.99 for 1,000 words, Standard $44.99 for 3,000, Pro $99.99 for 10,000, Max $199.99 for 30,000. You can also create an account and start with 200 free words.

Detection is a probability estimate, not proof. Use it as a mirror on your own writing, not as a verdict on anyone else’s.