Who Owns AI-Written Text, and What Happens When a Client Asks

The question "can I use this?" has three separate answers depending on who is asking, and they are routinely confused. A vendor's terms of service, a copyright office, and a client contract can all give different answers to the same document, and only one of them can cost you money.

Take them in order.

1. The vendor's licence: what you are actually granted

Most AI writing tools grant you rights to the output on paid plans and restrict or reserve rights on free plans. That sounds simple until you read the clauses that differ:

• Some grant a broad, transferable licence — you can use it commercially, assign it to a client, and not think about it again.
• Some grant use but not ownership, which matters if your contract with a client says you assign ownership of deliverables. You cannot assign what you do not own.
• Some reserve the right to use your inputs to improve their models. If your content includes client confidential information, this is the clause that matters and it is easy to miss.
• Some require attribution or disclosure, which is a real constraint if your marketing implies human authorship.

The practical check is narrow: read the section on output ownership, and the section on training data, on the tier you are paying for. Not the marketing page. Those two sections are the entire answer to "can I use this commercially", and they are usually a screen and a half of prose.

2. Copyright: the honest state of play

Copyright in purely AI-generated text is unsettled in most jurisdictions, and the direction of travel has been toward not granting protection to output with no meaningful human authorship.

What that means in practice is narrower than the headlines suggest:

• You can publish it. Nothing stops you publishing AI-assisted text.
• You may not be able to stop someone else republishing it. If there is no copyright in the output, there is no infringement when a competitor copies it. This is the risk that actually bites: your differentiated content is not protectable.
• Human authorship in the selection, arrangement, and substantial editing is the part that carries protection. A heavily edited, restructured, fact-checked piece is a different legal object from a raw generation, and the difference is measurable in how much of it a human actually wrote.

The workaround that holds up is not a disclaimer. It is sufficient human contribution — original research, specific numbers, quotes, structure, and edits that change the argument rather than the wording. The more of the document's value comes from things the model could not know, the less the copyright question matters.

This is also, incidentally, the same thing that makes content rank. The legal answer and the SEO answer point the same direction.

3. The client contract: the clause that costs money

This is the one people skip, and it is the one that generates invoices.

Standard freelance and agency contracts often contain one of these:

• A warranty that the work is original and does not infringe. If the deliverable is substantially AI-generated and the client's own counsel takes a narrow view, this warranty is where the argument starts.
• A requirement of human authorship, sometimes explicit ("all copy to be written by a named writer"), more often implied by a statement of work describing a writing service.
• An IP assignment clause that assumes you own what you are assigning.

None of these are automatically fatal. All of them are things to resolve before delivery, not after a client asks a question you cannot answer cleanly.

The practical move is a one-line disclosure in the engagement: "Deliverables may be produced with AI assistance and are reviewed and edited by a human before delivery." Put it in the contract, or in an email confirming scope, before you start. Clients who would object almost always object at that point, which is when you can still have the conversation. Clients who would not object do not care.

If a client has an explicit no-AI clause, that is a hard constraint and no licence fixes it.

4. AI detection: unreliable in both directions

You will be asked whether a piece is AI-written, possibly by a client, possibly by a publisher, possibly by a university.

Detectors produce two kinds of error, and both matter:

• False positives on human writing, especially from non-native English speakers and from people who write in short, declarative sentences. There are documented cases of students and writers being accused on detector evidence alone.
• False negatives on lightly edited AI text, which is most AI text in professional use.

Treat a detector score as a conversation starter, never as proof. If you are on the receiving end of an accusation, the answer is not the detector result — it is your drafts, your research notes, and your revision history. Keep them.

If you are the one checking a freelancer's work, the same logic applies in reverse: a clean detector score is not evidence of authorship, and a flagged score is not evidence of anything.

5. What to actually do

A short checklist that covers the realistic cases:

1. Read the output-ownership and training-data sections of your tool's terms, on your tier. Two minutes.
2. Add the AI-assistance disclosure line to your standard contract or scope email. Once, permanently.
3. Keep drafts and research notes for anything client-facing. This is your evidence and it costs nothing to keep.
4. Edit substantively — change the argument, not the adjectives. It improves the output and it strengthens your position.
5. Never submit a raw generation to a client with an originality warranty attached.

The uncomfortable summary

The tools are licensed generously enough that ordinary commercial use is fine. The exposure is not in the licence — it is in the warranty you signed and the assignment you promised. Those are two clauses in documents you already have, and they are worth ten minutes of reading before the next delivery rather than after it.

If you are choosing a writing tool primarily for client work, the AI writing tools reviews note where each one sits on commercial licensing and what the free tier permits. For the detection side, Originality.ai is reviewed with the accuracy caveats attached rather than left implied.