AI Detection and Freelance Work: Agree on AI Disclosure

Your client ran your delivered draft through a detector and now wants to know whether you used AI. Here is the short answer: put an AI disclosure clause in the contract before the next project, decide together what counts as acceptable assistance, and prove authorship with version history rather than a detector score you cannot control. AI detection in freelance work is easier to handle in advance than to argue about after delivery.

Key takeaways

• Agree on disclosure in the contract, before the work starts, not after a flag appears.
• Clients are usually protecting confidentiality, originality, accuracy and their own reputation, not policing your keyboard.
• Version history, timestamps and edit trails are stronger authorship evidence than any detector percentage.
• A detector score is a probability estimate, not proof — say that plainly and put the burden on process instead.
• Document your process lightly as you go; a pre-delivery checklist beats reconstructing a timeline under pressure.
• If you are accused, respond with evidence and a rewrite offer, not with defensiveness.

What is AI detection in freelance work actually measuring?

Detectors look at statistical patterns in text — how predictable each word is, how much sentence length varies, whether phrasing repeats in familiar shapes. They output a probability, not a verdict. Two tools looking at the same paragraph can disagree, because they weigh different signals. If you want the mechanics, how AI detection works walks through the signals in plain language.

That matters for your contract because a client who treats a detector score as a factual finding is treating a guess as a measurement. Your job is not to win that argument on the merits of the tool. Your job is to move the conversation to something you can both inspect: your process.

We ran our own detector across two small corpora on 2026-09-19 to see how noisy this gets. On 46 hand-written DeAIze guides, the mean AI score was 27% and the median 25%, with a range from 0% to 55%. On 12 unedited model outputs, the mean was 35% and the median 34%. Those two distributions overlap. A human-written piece can score higher than an unedited machine passage, which is exactly why a single threshold should never be a contract term. This is our own measurement, not a third-party detector result.

What should an AI disclosure clause in a freelance contract actually say?

Keep it short enough that both sides will read it. A workable clause answers four questions:

1. What counts as AI use. Spell out the line: grammar and spell checkers, autocomplete, transcription, translation, research summarisation, outline generation, full draft generation. Most disputes are about the grey zone in the middle, not about whether you pasted a whole article.
2. What must be disclosed and when. A simple rule works: any tool that generated more than a phrase of publishable prose gets named in the delivery email, before the client reads the draft. Disclosure after a question is not disclosure.
3. What the client can require. Options include: no generative AI in the final text; AI allowed for research and outlining but not drafting; AI allowed with disclosure; or AI allowed with client sign-off on the workflow.
4. What happens if a detector flags the work. This is the clause most contracts miss. Write down the remedy: you get a defined window to revise, the client does not withhold payment on the strength of a score alone, and any dispute is settled by comparing drafts.

A sample sentence you can adapt: “The Writer will disclose in writing any generative AI tool used to produce publishable text in the delivered work, and will provide draft versions on request. The Client will not treat an automated AI-detection score as conclusive evidence of undisclosed AI use; where a score is disputed, the parties will review the Writer’s draft history before determining a remedy.”

That is not a trick to escape scrutiny. It is a fair process, and it protects the client too — a client who fires a good freelancer over a false positive loses more than the freelancer does.

What are clients actually trying to protect against?

Almost never the tool itself. When a client asks about AI use, they are usually worried about one of five things:

• Confidentiality. They may have shared unreleased product details, and a public model may retain prompts.
• Originality and ownership. They need to know the text is not lifted from somewhere else. That is a different failure from AI use — AI detection vs plagiarism explains why the two get confused.
• Accuracy. Invented facts, invented quotes and fake citations are the failure mode clients actually care about. If you use AI anywhere in your process, verify every reference before delivery.
• Voice. They hired you for a tone their audience recognises. A flat draft reads like a different brand.
• Their own liability. Regulated industries, government tenders and academic publishers may have disclosure rules the client must follow.

Notice that only one of those five is really about detection. If you ask a worried client which of these five keeps them up at night, you usually get a specific answer — and a specific answer is something you can solve.

If you want to see which sentences in your own draft are carrying the machine-made signal before you send it, run it through the AI detector — detection is sentence level, so you get the whole-document score and a per-sentence breakdown instead of one opaque number.

Why is a detector percentage a bad contract term?

Because it is unstable in both directions. The same text scored by two tools, or by the same tool twice, can land differently. Non-native English writers get flagged more often than native speakers for reasons that have nothing to do with AI, and that pattern is well enough documented that we wrote about why detectors flag non-native writers.

There is also a subtler problem: a detector score is not reproducible for the client. They cannot open your draft six months later and get the same reading. A contract term that cannot be reproduced is not a term, it is a mood.

How to say this without sounding evasive:

“I’m happy to talk about how the work was made, and I can show you every draft. I can’t guarantee a number from a tool that estimates rather than measures, and I don’t think either of us wants a payment decision resting on one. Can we agree that draft history is the record?”

That sentence works because it offers something concrete in place of the thing you are declining.

How does version history prove authorship?

Version history is the closest thing freelancers have to a paper trail. It shows a document changing over time in ways that look like human work: an outline that got reorganised, a paragraph that was cut and rewritten, a fact that was checked and corrected, a sentence that got shorter on the third pass.

What to keep, without turning your week into admin:

Evidence	What it shows	Effort to keep
Cloud document version history	Sequence of edits with timestamps and authors	None — it accumulates on its own
Two or three saved drafts (outline, mid, final)	Progression of thinking, not one paste	Low — save copies at natural checkpoints
Notes file with sources and quotes	Where facts came from and when you checked them	Low — you are already researching
Interview notes, call recordings, brief emails	Raw material that predates the draft	None — it exists already
Commit history or tracked changes in Word	Line-level authorship of edits	Low if you already use track changes

A client reviewing that set is not being asked to trust you. They are being shown a process. In practice this settles most disputes faster than any explanation of how detectors work, because it answers the real question — did a person think about this — with artefacts rather than assurances.

One thing to be honest about: version history proves a document evolved. It does not prove no AI was involved at any step. Do not oversell it. Present it as what it is — evidence of authorship and process — and pair it with a straight answer about which tools you used.

What do you do when a client has already accused you?

You have probably already searched for “client accused me of using ai to write.” Here is the sequence that works.

1. Do not argue about the score. Thank them for raising it and ask which tool and which passage.
2. Ask what specifically concerns them. Confidentiality, accuracy, voice, originality or their own compliance. Get the actual worry on the table.
3. Send your draft history unprompted. Outline, mid-draft, final, plus your sources file. Do not make them ask twice.
4. Offer a targeted revision, not a defensive rewrite. Say which sections you will rebuild from your own notes, and when they will have it.
5. Disclose accurately. If you used a tool for outlining or cleanup, say so now. A late admission costs far more than an early one.
6. Fix the contract for next time. Add the clause above, or amend the current one if the relationship is worth keeping.

If the client is an agency or publisher with a formal policy, ask to see it. Many of these policies are vaguer than the people enforcing them realise, and reading it aloud together often resolves the disagreement. We covered the same dynamic from the student side in false positive AI detector: what to do when you’re accused — the structure of the response is identical.

A pre-delivery checklist that costs you five minutes

Run this before you hit send on any substantial deliverable. It documents your process as you go, which is the only way this stays cheap.

• Save a copy of the outline as its own file.
• Save a mid-draft copy at roughly the halfway point.
• Keep your sources in a single notes file with links and dates.
• Write one line in the delivery email naming any AI tool you used and what for.
• Read the draft aloud once; anything you would not say out loud gets rewritten.
• Check every factual claim and every citation.
• Keep the client’s brief and your reply in the same thread as the deliverable.

That is the whole system. It takes a few minutes per project and it converts an argument about a probability estimate into a conversation about files the client can open.

Where DeAIze fits, and where it does not

DeAIze is a web app with two halves: an AI detector and an AI humanizer. Detection is sentence level — you get a document score and each sentence highlighted by AI probability, so you can see which lines carry the signal rather than guessing. The humanizer runs a closed loop: it scores the text, rewrites only the flagged sentences, keeps a rewrite only when semantic similarity says the meaning survived, then re-scores and reports before and after. On the example paragraph on the site, a 92% AI score becomes 4% AI, and 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.

Where it helps a freelancer: as a pre-delivery pass on your own writing, to catch the flat, machine-made patterns that make a human draft read as if it were generated. Where it does not help: as a shield in a contract dispute. No tool can promise a particular score on someone else’s detector, and DeAIze does not either. It also is not a way to pass off someone else’s work as yours. The ideas, data and conclusions in a deliverable have to be the author’s own — that part is not a formatting problem, and no rewrite fixes it.

If you want to understand what a client’s tool is actually reporting before you reply to them, what each detector reports is a fair starting point, and why two tools disagree explains the disagreement you are about to be asked about.

Ready to see which sentences read machine-made?

Before your next delivery, run the draft through the AI detector to see the sentence-level breakdown, then use the draft editor to rework only the lines that carry the signal — keeping your meaning, your citations and your voice intact. If you are working from a Word file, rewriting preserves fonts, styles and tables, so you are not rebuilding formatting afterwards. Detection and rewriting cover English and 9 other languages, the interface ships in 10 languages, and the free tier gives you 200 words on signup with no credit card. Pricing is pay-as-you-go with credits that never expire and no subscription: Starter at $20.99 for 1,000 words, Standard at $44.99 for 3,000, Pro at $99.99 for 10,000, Max at $199.99 for 30,000. Text is processed for the current task only, never used for training and never resold. Start at register or check pricing first.

None of that replaces the clause in your contract. Do both: agree on disclosure up front, keep your draft history, and use the tool as a final pass on your own work rather than as a defence you will need later.