Writing With AI Without Misrepresenting Authorship

We build a tool that edits AI-generated text, so we have an obvious interest in this question. That is exactly why it is worth answering carefully rather than skipping.

There is a real line here, and it is not "did a model touch this". It is whether anyone is being misled about something they have a legitimate reason to know.

Where the line actually sits

Most writing tasks are not assessments of your unaided ability. An email, a product description, internal documentation, a first draft — nobody is evaluating whether you produced those keystrokes personally. Using a model there is like using a spellchecker or a calculator. No disclosure is owed because no one is relying on a claim you have not made.

Some tasks are different, and they share one feature: the process is the point.

• Assessed work. An essay exists to demonstrate that you can construct an argument. Outsourcing that defeats the purpose, whether to a model or a friend.
• Professional attestations. Signed legal, medical or engineering work asserts that a qualified person exercised judgement.
• Work sold as original authorship. A commissioned article where the client is paying for your voice and expertise.
• Anywhere disclosure was required and skipped. Many journals, employers and clients now have explicit AI policies. The rule exists; following it is not optional because you disagree with it.

The test that actually works

Ask: if the person receiving this knew exactly how it was produced, would they feel misled?

If no, you are fine. If yes, the problem is not detection — it is the misrepresentation, and it remains a problem whether or not anyone catches it. Notice that this test does not depend on any tool's capabilities, which is why it stays reliable as those capabilities change.

Using AI in ways that hold up

Plenty of heavy AI use is entirely defensible:

• Thinking out loud. Interrogating an idea, finding counterarguments, testing whether your reasoning survives pressure.
• Structure. Reorganising material you have already produced.
• Language help. Getting a second language to read fluently. This is arguably the highest-value legitimate use, and — bleakly — the one most likely to be falsely flagged.
• Editing. Tightening, cutting, fixing rhythm in your own draft.
• Getting unstuck. Generating a bad opening paragraph so you have something to react against.

What these share: the substance, judgement and claims remain yours. The model helped you express your thinking rather than replacing it.

Keep evidence of your process — before you need it

The best protection against a false accusation is a record, and it costs almost nothing to maintain.

• Write somewhere with version history. Google Docs, Word with autosave, or git. A document that grew over two weeks in 40 revisions is powerful evidence; a single paste event is the opposite.
• Keep your notes and sources. Reading annotations, outlines and dead ends are hard to fabricate after the fact.
• Keep drafts. Do not overwrite. The messy version is the useful one.
• Note what you used and how. A line in your own notes — "used a model to restructure section 3" — costs nothing and settles arguments.

This matters because classifiers produce false positives at rates that guarantee innocent people get flagged. Process evidence is what you fall back on, and you cannot create it retroactively.

If you are accused and you did the work

1. Ask what the finding is based on. A perplexity classifier is a guess with a documented bias; a watermark z-score is near-proof. Establish which one.
2. Produce your process evidence. Version history, drafts, notes.
3. Ask about the false-positive rate. Many institutions have adopted these tools without examining published accuracy. The Stanford finding that detectors flagged over half of non-native TOEFL essays while barely misclassifying native speakers is directly relevant.
4. Offer to discuss the content. Someone who wrote a piece can talk about the choices in it. This is more persuasive than any statistic.
5. Escalate on process, not on tooling. Argue that the evidentiary standard is inadequate, not that the detector is technically flawed.

Where our tool fits

We built it because statistical watermarking is genuinely interesting and because targeted editing is a better approach than blanket rewriting. Legitimate uses include understanding how these schemes work, keeping AI assistance out of text where you have already disclosed it, and reducing false-flag risk on your own writing.

It is not a way to make plagiarism undetectable, and we would rather say so than pretend the question does not come up. Our terms ask you to use it on text you own or are authorised to edit, and to follow the rules of whoever you write for. The tool changes some word choices; it cannot change who did the thinking.

If you are weighing whether editing will get you past a detector, the more useful question is whether you would be comfortable explaining your process. That answer does not change with the technology.