What the October 2026 Guild guidance says

The Authors Guild’s publisher best practices, dated October 13, set out seven principles:

1. Provide a clear written AI policy acknowledged with the publishing agreement.
2. Avoid AI manuscript editing beyond spelling and grammar corrections; obtain express written permission before inserting generated content, including artwork.
3. Disclose publisher AI and detector use. Before third-party screening, obtain permission and confirm that submitted text will not be retained or used for training.
4. Do not base AI accusations or cancellation solely on detector results.
5. Discuss concerns with authors and allow correction through the contractual approval process.
6. Account for false-positive risks affecting second-language, conventional-genre, and deliberately plain writing.
7. Do not require authors to prove authorship through writing surveillance.

The Guild also ties cancellation to clearly prohibited use that breaches a material signed term. After publication, it recommends against cancellation for AI use alone unless the contract clearly allows it. These are the Guild’s positions; they are not independent termination powers or statutory rules.

The rest of this guide develops an operational framework. Its records, suggested timelines, decision categories, and sample language are editorial proposals, not additional Guild requirements.

Separate voluntary recommendations from enforceable requirements

A publisher should identify the authority behind each decision. An industry recommendation describes a preferred practice. A signed agreement may create obligations between the parties. Applicable law can impose duties independently of either document. A retailer or distributor may also impose conditions for its own service. Calling all four things “compliance” without identifying the source makes disputes harder to resolve.

For a concrete U.S. distinction, the Copyright Office’s January 2025 copyrightability explanation says that AI assistance does not automatically defeat protection for human authorship. Protection depends on sufficient human expressive contribution; prompts alone are insufficient. Its registration guidance instructs applicants to disclose AI-generated material and exclude more-than-de-minimis generated material from the claim. That registration process is separate from a publisher’s house policy and from a public-facing AI label.

Do not reduce an entire mixed work to a single copyright conclusion because one asset was generated. Record the human contributions and generated components separately. Likewise, author permission for an upload does not necessarily resolve rights in quotations, confidential interviews, licensed illustrations, or personal information about other people. Escalate the specific issue rather than assuring staff that a checkbox cures every legal concern.

This guide provides general publishing information. The sample clauses require review against the governing agreement and applicable law before adoption. No universal conclusion about infringement, privacy liability, or lawful cancellation follows from the presence of AI alone.

Design screening around a defined editorial question

Before buying or enabling a screening service, decide what problem the house is trying to solve. Factual accuracy, plagiarism, undisclosed generated passages, and poor writing are different review questions. A manuscript can fail a factual check regardless of its production method. A source match requires examination of quotation and permission. An AI-use concern requires attention to the agreed policy. Staff should record the question before selecting a tool.

Use ordinary editorial review as the starting point. Check sources, read for continuity, verify quotations, and identify passages that require clarification. A reviewer who encounters a fabricated reference can document the missing source directly. There is no need to turn that finding into a separate assertion about the author’s technology use.

Make a vendor register before creating an upload habit

For every proposed service, maintain a short approval record showing its purpose, account tier, contractual terms, access controls, and permitted input. Review the actual account configuration as well as the marketing page. A product name alone is an inadequate approval because different plans or features may handle data differently.

Ask specifically about temporary processing, logs, backups, human support access, subprocessors, deletion, and training or evaluation. Record what the provider can demonstrate. If a service cannot meet the intended confidentiality conditions, leave the manuscript out of that service and use another review method. Do not turn an unresolved retention question into a promise that nothing is stored.

Link vendor approval to a responsible staff member and a review date. Reopen the approval when the provider changes its terms, the publisher changes plans, or a feature begins sending text to another processor. A spreadsheet can handle this for a small list of vendors; a professional house can connect the same fields to procurement and security systems.

Control how manuscripts leave the editorial system

Give staff an approved route for manuscript handling and a clear stop rule for unapproved tools. Include browser extensions, document add-ins, shared accounts, and freelance workflows in that instruction. An editor may otherwise believe that text stays inside the word processor even when an enabled feature sends it elsewhere.

Limit access to people working on the title. Keep permissions separate from general marketing folders. Use approved accounts, access logs where available, and a documented retention schedule. For an authorized external task, send only the material needed for that task. Removing the author’s name may reduce exposure, but the prose itself remains unpublished material and can still reveal identities or confidential facts.

Use the existing rights-provenance guide for the title’s broader ownership and licensing file. Add the review decision and any external-transfer permission to that record rather than creating a competing archive of contracts.

An actionable manuscript-review procedure

The following ten-step procedure is designed for intake, acquisition, and prepublication review. It can also organize a backlist concern, although decisions affecting a published book need separate contractual and legal review. Suggested response periods are internal service targets, never substitutes for negotiated deadlines or legal requirements.

1. Open a case tied to the submitted version

   Assign a title ID and record the submission date, manuscript version, responsible editor, and applicable policy version. Preserve the submitted file without silently replacing it. Note whether the concern arose at intake, revision, acceptance, or after publication. Those stages can have different approval provisions and practical consequences.

2. Record a specific observation

   Describe the passage, page or section, and observable problem. Separate facts from hypotheses. “The citation cannot be located in the named journal” is a fact to investigate. “The author must have generated the chapter” is a hypothesis. If the only input is a complaint from a reader, preserve the complaint as a report rather than treating it as an established finding.

3. Find the actual policy question

   Identify the term that might apply and what it permits. Ask whether the question concerns generated published expression, an undisclosed production aid, or simply an editorial defect. If the contract and policy are silent or ambiguous, refer that uncertainty to the decision owner. Do not rewrite an old agreement through an editor’s interpretation of a newly adopted policy.

4. Choose the least intrusive useful review

   Begin with a second editorial read, source checking, and a discussion of the passage. If anyone proposes external automated analysis, pause until the vendor conditions, permission, and input scope are resolved. Record a decision not to run a detector just as clearly as a decision to use one. A tool is optional within this proposed workflow.

5. Interpret automated output cautiously

   Preserve the report, its date, the exact input, service version when available, and the account settings. Do not translate a displayed percentage into a percentage of the book that was generated or a probability that the author lied unless the service actually defines and validates that interpretation. A second automated score should not be counted as independent corroboration without examining how the systems and inputs relate.

6. Contact the author privately

   Explain the particular concern, the relevant policy provision, and any review already performed. Share the passages and relevant report information in an accessible format. Invite an explanation without demanding access to personal devices or unrelated accounts. Offer a conversation or written response, and include the author’s agent or representative when appropriate.

   A workable administrative target is to acknowledge the concern within two business days and offer ten business days for the author’s response. Adjust for accessibility needs, translation, travel, manuscript complexity, and existing contractual notice provisions. Record the agreed date so silence before that date is not mischaracterized as refusal.

7. Evaluate the explanation and contrary evidence

   Assign a reviewer who can assess the relevant genre and language context. Consider the author’s explanation, source materials, editing history voluntarily supplied, and any identifiable evidence bearing on the specific issue. Do not infer wrongdoing from missing draft files. A writer who composes offline, dictates, or overwrites working drafts may have a different record from one who uses cloud version history.

   Where language background or formulaic structure may explain a flag, record that alternative explanation in the case file. The reviewer’s task is to test the concern, not to assemble only material that supports it. If the evidence remains inconclusive, the disposition should say so.

8. Separate the finding from the remedy

   Use explicit outcomes: no supported policy concern; editorial defects requiring revision; disclosed use permitted under the agreement; established noncompliance requiring contractual review; or unresolved concern. A book can need revision without a finding of prohibited AI use. Define the actual correction, responsible person, review date, and contractual basis for any consequence.

9. Provide a second review before a serious adverse decision

   Have someone other than the original complainant review a proposed rejection, cancellation, payment demand, or public statement. In a one-person press, this may mean an independent editor and, for contractual consequences, qualified counsel. Confirm notice requirements, opportunities to cure, the evidence supporting the decision, and whether the author received a meaningful chance to respond.

10. Close the case and repair incorrect records

    Send a factual written outcome to the author. Record any accepted revisions and the version approved for publication. If the concern was unsupported, correct internal labels that could prejudice later decisions. Restrict the retained case file, apply its retention rule, and preserve material needed for an actual dispute or legal hold. Do not circulate a cleared allegation as informal background about the author.

Example: a technical handbook with repetitive prose

Suppose a handbook repeatedly uses the same warning structure and an automated service flags it. The editor identifies no separate evidence of prohibited generation. Under this proposed procedure, another editor reviews the warnings for technical clarity and checks the relevant sources. The author explains the repeated structure as an intentional safety convention. If that explanation fits the evidence, the case closes without an AI violation. Any inaccurate warning still receives an ordinary technical correction.

If the same author instead confirms that prohibited generated passages remain in the submission, the editor identifies those passages and applies the agreed revision process. The admission answers a different question from a stylistic hunch. The resulting record should explain what happened and how the manuscript was brought into compliance.

Make author consent specific enough to operate

A consent record should let an editor identify the authorized action without searching through an entire email thread. Use separate permissions for separate purposes. Permission to run a limited review should not silently authorize training, synthetic narration, translation, promotional reuse, or a new derivative product.

Give the author the information before requesting a decision. If the provider’s terms later change, stop new transfers and reassess the authorization. Withdrawal can govern future use, but do not promise that a completed transfer can always be reversed. State any practical limits accurately.

An author’s refusal should route the title to an agreed alternative review, not automatically become evidence about authorship. If the parties have an actual dispute over a contractual cooperation duty, address that duty directly and obtain advice before attaching consequences.

Tell authors what the publisher’s editorial tools actually do

The proposal here is to maintain a title-level editorial tool record. Identify the function, input, reviewer, proposed output, author authorization where applicable, and final disposition. Distinguish a spelling correction from a rewritten paragraph. A feature labeled “polish” may require closer examination than its friendly product label suggests.

For a conservative house implementation, keep substantive manuscript editing with human editors. Treat any proposal to add generated text or imagery as a separate creative decision requiring specific written approval and rights review. Disclose operational uses that touch the work, including marketing preparation, through a clear description of purpose and data handling.

Author communication and reader disclosure serve different audiences. An author needs to understand processing and creative changes. A reader may need an accurate description of material synthetic contributions. Use AI Content Disclosure: A Practical Publishing Standard for 2026 for the broader labeling framework. This article adds a permission and review process rather than repeating that guide’s taxonomy.

For terminology, refer staff to AI-Assisted Is Not AI-Generated. Do not use a broad production label to evade a narrower manuscript agreement: the title’s executed terms still need to be checked.

Write contract language that an editor can apply

The Guild also maintains separate AI-related model publishing contract clauses covering rights reservations, negotiated AI licensing, author use, and publisher use. Read those provisions in their own context. A model clause is a drafting option, and different documents may contain different exceptions.

For house drafting, define the work covered, the effective policy version, permitted author uses, prohibited uses, disclosure timing, publisher processing permissions, and the review route. Address how amendments become effective. Avoid a floating promise to comply with whatever policy happens to appear on a website in the future.

The following original discussion clauses illustrate the structure. They are not Guild text and are not a complete agreement. Counsel should reconcile definitions, warranties, cure rights, payment, acceptance, termination, and dispute provisions with the full contract.

Policy incorporation

The parties adopt the AI Policy version identified in the signed policy schedule. Later revisions do not modify this agreement unless accepted in a written amendment signed by both parties. The schedule identifies permitted and prohibited uses for this work and the disclosures required at delivery and revision.

Processing permission and reserved rights

Publisher will not submit the unpublished work to an external AI analysis service without a separate written authorization identifying the service, purpose, permitted material, and agreed data-handling conditions. Such authorization grants no training, translation, narration, adaptation, or other exploitation rights beyond the stated task. Any further rights require a separate express agreement.

Review and response

Before asserting an AI-policy breach, Publisher will provide confidential written notice identifying the applicable provision, relevant passages, and supporting information, and will give Author a reasonable opportunity to respond. Proposed corrective steps and deadlines will follow the agreement’s acceptance and cure provisions. Automated screening output alone will not determine breach.

Negotiate what happens if correction is incomplete, the parties disagree about scope, or a concern emerges after publication. A new house procedure cannot itself create a cancellation right. Specify the escalation route, authorized decision-maker, available review, and relationship to existing dispute-resolution terms. Do not assume an indemnity clause answers the underlying question of whether a breach occurred.

Manage a live dispute with a controlled record

Keep a chronology of the submission, governing terms, concern, evidence, author response, proposed correction, and decisions. Separate legal advice from routine editorial correspondence according to counsel’s instructions. Limit outside communications to authorized people and factual statements. If the publisher has already made an inaccurate statement, seek advice on a prompt correction and communicate the corrected outcome to the people who received the original allegation.

Where confidentiality may have been compromised by an upload, treat that as its own incident. Stop further transfers, preserve the relevant settings and records, contact the provider through the approved channel, and assess deletion and notification obligations. Resolving the manuscript’s editorial quality does not automatically resolve an exposure of the manuscript.

Sample author-facing AI policy

Original model for adaptation: The text below is a proposed house policy, not an Authors Guild policy or certification. Before adoption, place the publisher’s identity, policy version, effective date, and designated editorial contact in the signed policy schedule. Connect it to the publishing agreement and use only promises the publisher can fulfill.

Publisher compliance checklist

Use this as a self-audit of the proposed workflow, not a legal certification. Mark each item complete, incomplete, or not applicable, and identify an owner and evidence location.

• The executed agreement identifies the applicable AI policy version.
• Staff can distinguish correction, generation, translation, and operational processing.
• Permitted author uses and required disclosures are explained before delivery.
• The publisher’s substantive editing and generated-content rules are documented.
• Approved vendors have current, account-specific data-handling records.
• External-processing permissions identify purpose, scope, and conditions.
• Contractors, add-ins, and shared accounts follow the same handling rules.
• Review records separate observed facts, hypotheses, and automated output.
• Author notices identify the relevant policy and allow a meaningful response.
• Alternative explanations and contrary evidence receive documented consideration.
• Voluntary process evidence is not converted into a surveillance requirement.
• Findings distinguish editorial defects, permitted use, noncompliance, and uncertainty.
• Serious adverse decisions receive independent and contractual review.
• Final records link to the accepted manuscript version and agreed correction.
• Access, retention, incident handling, and correction procedures have named owners.
• Reader-facing disclosures and distribution records reflect actual production.

Implement the process without building a bureaucracy

A small press can begin with one policy schedule, one vendor register, one consent form, and a restricted case log. The publisher may own all four, while an outside reviewer handles contested decisions. Start by reviewing current tools and pending submissions; resolve unclear permissions before further transfers.

A larger house can assign editorial ownership of policy, procurement ownership of vendor terms, security ownership of access, and legal ownership of contractual escalation. Keep one accountable coordinator per title so an author does not have to reconcile conflicting instructions from several departments.

For the wider governance framework, use What a Modern AI Publishing Standard Should Include. The operational test for this guide is concrete: can the publisher retrieve the relevant policy, explain the actual concern, identify authorized processing, and show how the author’s response affected the outcome?

Frequently asked questions

Are the Authors Guild’s October 2026 recommendations legally binding?

No. They are voluntary recommendations. Applicable law, executed contracts, and relevant service terms must be assessed separately.

Does this guide require publishers to buy an AI detector?

No. The proposed workflow allows ordinary editorial review, source checking, and author discussion. Vendor approval and permission are prerequisites when external automated analysis is proposed.

What should an editor do when a detector flags a manuscript?

Open a restricted review record, identify the actual policy question, examine the relevant passages, and invite the author’s explanation. Under this proposed procedure, a score alone does not establish breach.

Can an author decline a proposed third-party upload?

This model policy provides an alternative review route. If the parties dispute an existing contractual obligation, review the actual agreement rather than treating refusal as evidence of generated writing.

What if the author has no saved drafts?

The proposed procedure does not treat missing draft history as misconduct. Evaluate the specific concern and available evidence without demanding surveillance records or private account access.

Does AI involvement make an entire book uncopyrightable?

Not automatically. U.S. Copyright Office guidance distinguishes human authorship from generated material. A mixed work requires analysis of its protectable human contributions and appropriate registration disclosures.

Can publishers adopt the sample policy immediately?

Use it as a drafting starting point. Identify the publisher, version, effective date, and contact; confirm operational promises; and reconcile the text with existing contracts and applicable law before adoption.

Verified references and related reading

Sources checked October 21, 2026. The first item is the primary source for the October recommendations. The contract resource is separate guidance. Copyright Office materials address U.S. copyright and registration, not a universal publishing AI policy.

1. Authors Guild: AI Best Practices for Publishers: Using AI Detectors and Dealing With Undisclosed AI Use. October 13, 2026. Seven publisher principles.
2. Authors Guild: AI-Related Model Publishing Contract Clauses. Separate drafting resources.
3. U.S. Copyright Office: Copyright Office Releases Part 2 of Artificial Intelligence Report. January 29, 2025.
4. U.S. Copyright Office: Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence. March 16, 2023. See the application guidance in section III.
5. Publishers Weekly: Authors Guild Issues AI Guidelines for Publishers. Jim Milliot, October 16, 2026. Contemporaneous reporting.