A critique of the Authors Guild’s “AI Best Practices for Publishers”

Many authors today fear being wrongly accused of using AI by editors, readers, and imperfect AI detectors. Publishers have already canceled several books for ostensibly containing AI-generated text. We ask authors to be transparent about their use of AI, and publishers should do the same, including with respect to their use of AI detectors.

To ensure mutual trust between authors and publishers, the Authors Guild proposes the following principles and guidelines for transparency and disclosure regarding publishers’ AI use:

This is exactly the right approach. Here are the specific guidelines with my analysis of where they’re helpful and where they fall short:

1. Publishers should provide authors with clear and transparent AI use policies.

Every publisher should adopt a written AI policy that spells out what types of AI uses it will permit or prohibit. The policy should either be made part of the publishing agreement or, if in a separate document, be provided to the author for acknowledgement when signing the publishing agreement.

Whether included in the publishing agreement or a separate document, the AI policy should encourage the author to disclose AI use and commit the publisher and author to working together to bring the manuscript in compliance with the publisher’s author AI policy. The policy should clarify the publisher’s expectations with respect to the “originality” and “sole author” representations in the author contract, in addition to explaining that AI-generated text is not original and cannot be copyrighted.

No publisher should ever cancel a contract due to AI use unless the author violated a signed contract or policy that clearly prohibits the particular use and it is a material term of the agreement.

This is a badly needed guideline. Based on news coverage, multiple authors have had contracts revoked based on policies that were not clearly stated ahead of time. Who knows how many cancellations of this kind happened quietly, out of the spotlight? The publisher has no right to invoke some vague and undisclosed guideline not clearly stated in a contract.

2. Publishers should not use AI to edit.

Either in the written AI policy or in the publishing agreement, the publisher should commit to not using AI to edit a manuscript, other than for mere spelling and grammar corrections. Further, the publisher should agree not to insert AI-generated content, including artwork, without the author’s express written permission.

This guideline needs revision. I agree that publishers should not insert AI-generated content, which would be outrageous. But the term “edit” here is vague. Publishers and editors (including me) use AI to analyze and identify problems with manuscripts that go beyond grammar and spelling. For example, you could ask an LLM, “What concepts are explained multiple times in this manuscript and where would it make sense to delete those duplications?” But in the end, editorial feedback should come from editors, whether or not they use AI for help, rather than from the AI systems themselves.

I have some experience with this. The publisher of a book I ghostwrote sent back an editorial review that was clearly done by AI. It was worse than useless. It made stupid suggestions and missed obvious problems; in the end, my client and I had to hire our own outside copy editor to do the job the publisher didn’t. AI can be helpful, but is no substitute for an editor’s judgment.

3. Publishers should be transparent about their use of AI and AI detectors.

Publishers should be fully transparent about when and why they use AI detectors and should include this information in their written AI policies. Publishers should likewise be transparent about any use of AI in the marketing or sale of a book, including whether and when a manuscript may be uploaded to an internal AI system and for what purpose; and, as we noted in our April 2026 statement, publishers should never allow employees to upload a manuscript to any public AI system (as doing so is copyright infringement), but only with the author’s permission to closed systems operated by the publisher.

The same considerations apply to AI detectors. Using a third-party detection service requires uploading that manuscript to a third party’s AI system. Before doing so, a publisher should know how the service stores and uses submitted text, should confirm that the text will not be retained or used for training, and should obtain the author’s permission. While some detectors state that they never train on user data, be aware that policies and practices could change at any time.

It certainly makes sense for publishers to be clear about which AI detectors they use and when.

The statement “Publishers should likewise be transparent about any use of AI in the marketing or sale of a book” needs more precision. For example, suppose a publisher uses AI to analyze sales data and identify which bookstores to target or which Amazon ads to place. There’s no reason that needs to be disclosed to every author. In general, I think authors, not publishers, should write the marketing materials for their books; if the publisher does so, the author should be able to approve or edit them. This is true regardless of how they were originally created.

Publishers should never upload author content to any system that uses it for training purposes. This is certainly a best practice. I think that once the use of AI systems becomes more commonplace in publishing, there will be corporate guidelines about this. Editorial management should treat an editor uploading a manuscript to ChatGPT as just as unethical as if that editor posted excerpts of an author’s manuscript publicly without permission.

4. Neither publishers nor agents should rely solely on AI detectors to determine if parts of a manuscript were AI-generated.

No publisher, agent, or other person should rely solely on an AI detector to determine if a book contains AI text. AI detectors—even the best—are not always accurate. Publishers and agents should not rely upon self-reported error rates of any detector, as they may not be accurate over time and among different works.

Given the possibility of false positives (and false negatives for that matter), the Authors Guild objects to the use of AI detectors alone by publishers, editors, and agents to accuse a writer of using AI to generate text, much less to cancel a contract. One false accusation can ruin an author’s career. The harm to the author is too great to tolerate any false positives. If AI detectors suggest that an author has violated a publisher’s written AI policies, the publisher must consult with the author regarding the matter.

AI detectors are useful. AI detectors are also fallible. This is the difficult challenge that publishers must now navigate.

As a guideline, this statement is fine. But it doesn’t give publishers any real clarity about what to do, and that’s going to make life difficult for acquisitions editors deciding which books to sign.

It makes sense to use AI detectors to identify when parts of a manuscript seem to be AI-generated. Tools like Pangram are pretty good at this, but not perfect. If an AI detector indicates a suspicion that parts of a manuscript are AI generated, that should be the starting point of a discussion about the topic. Often, a manuscript that Pangram marks as AI-generated has other problems — repetition, misattribution or missing attribution of quotes and ideas, or just plain boring prose. Those are fine reasons to reject a manuscript, regardless of whether the author used AI to generate it.

5. Editors should be educated to detect AI use and engage in conversations with authors where AI use is suspected.

If an editor suspects AI was used to generate text or other creative elements of the manuscript, the editor may use a non-infringing and generally reliable AI detector under the conditions noted above and should discuss the results with the author as part of the review and approval process. The author should be given the opportunity to rewrite the AI-generated portions in accordance with the editor’s feedback.

If the author refuses or neglects to correct the text pursuant to the approval process laid out in the agreement and AI policy, depending on the terms of the contract, the persistent AI use could be a basis for rejection of the manuscript. However, once a manuscript is approved and published, AI use alone should not be used as a basis for cancellation of the contract unless that is clearly spelled out in the contract.

The bottom line is that if an editor suspects undisclosed AI use, they should first contact the author and have a frank conversation. Results from an AI detector should be used as a basis for discussion—not to summarily cancel a contract or publicly shame any author.

There’s a lot of passive voice in both these recommendations and my critique. But this passage is the place where that passive is most problematic — in the phrase “Editors should be educated . . .” The desire to promote frank conversations is exactly the right approach, but the people in the industry don’t yet have the training to do this, and it’s a moving target. This is going to continue to be a problem.

6. Publishers should take into account special cases that may be especially susceptible to false AI accusations.

AI detectors are prone to flagging writing by non-native English speakers as AI-generated at higher rates since many detectors work by measuring how predictable a text’s word choices and sentence structures are. Writers working in a second language often use more common vocabulary and more regular syntax, and that can read as “predictable” to a detector. The same can be true of writers in genres with strong conventions and of writers whose style is plain by design. Publishers should treat an AI detector’s flagging such writing as weak evidence and be aware that plain or formulaic prose is not, on its own, a sign of AI use.

Where an author’s linguistic or cultural background may bear on why their writing was flagged, the publisher should approach the conversation sensitively and rigorously, and may include staff responsible for diversity, equity, and inclusion in the conversation.

This is a challenge. The people who use AI for help the most are, as you might suspect, people for whom writing is the most difficult. That includes non-native English speakers. So this statement from the Authors Guild may seem to be supportive of “nontraditional” writers, but all it does is create a new category of loophole. “[T]he publisher should approach the conversation sensitively and rigorously” is good advice, but the lack of specifics in this passage is going to continue to be a problem. Publishers and editors don’t really know exactly what to do, regardless of diversity characteristics of the authors they’re dealing with.

7. Publishers should not require authors to prove that they wrote their books.

Professional writers have usually spent years developing their craft; it is the author’s voice, intellect, personality, and creativity that make their work unique. They should not be required to prove that they are doing what they have always done—writing their own books. Many writers are rightfully deeply offended by such notions and feel as though they are subject to a witch hunt for merely practicing their craft as they always have.

As such, the Authors Guild opposes any requirement that authors use technologies that track or log their writing process in order to demonstrate that they wrote the work themselves.

At the same time, we fully support authors who choose to use such technologies or processes to preserve their drafts or writing in the event they are later accused of using AI to write for them.

The challenge here is the burden of proof. If a publisher says to an author, “Prove you wrote this,” how is the author supposed to respond?

Writing inevitably generates masses of research, source material, and drafts. As an author, it’s a good idea to keep track of that material, because you may need it.

It’s fine that “the Authors Guild opposes any requirement that authors use technologies that track or log their writing process.” But publishers who suspect AI-generated manuscripts are, at some point, going to need some sort of proof. Authors whose text tests as AI-generated and don’t have any of that sort of proof are going to have trouble getting published. The Authors Guild can’t change that.

My takeaways from this set of guidelines

I critique this set of guidelines from a position of respect. Publishers as an industry have abdicated their responsibility to create their own guidelines. I commend the Authors Guild for attempting to fill that void.

That said, these guidelines include a lot of vague and hopeful statements that, in the end, put the burden on editors in the publishing industry to master technologies that are changing by the minute. The risks of publishing AI-generated, uncopyrightable material are huge. The risk of false accusations is also huge. And there’s not yet a solution to resolving those risks.

I was the lead author of the AI Guidelines for Ghostwriters, a document that was initiated by Gotham Ghostwriters and developed with input from a variety of contributors. The process was inclusive, and the result clear and straightforward. I’d love to see publishers undertake the same sort of effort. If the Authors Guild guidelines prod them to do so, that’s would be a great step forward.