How Can Authors Negotiate Fair AI Rights With Publishers and Platforms?

The direct answer

Authors who negotiate with publishers, AI companies, and content platforms should treat AI rights as a bundle of permissions, not as a single yes-or-no question. The central issue is whether a work may be copied for model training, stored, retrieved, quoted, displayed, converted into synthetic material, or used to compete with human authorship. A useful AI rights negotiation guide therefore separates licensing for development, commercial reuse, attribution, payment, deletion, audit rights, and future enforcement. As of 24 September 2026, there is still no universal copyright settlement for these activities, and rules differ between countries, platforms, and jurisdictions. The best strategy is to prepare a written request before opening commercial discussions, identify the exact work and proposed use, and ask for language that can survive changes in corporate ownership or model deployment. Authors should not rely on a promise that an AI company “will credit authors” unless the contract defines where, how often, and for how long credit appears. The goal is a deal that pays fairly, explains the use, preserves author control, and avoids giving away rights that the author may need years later.

Also worth reading: AI Publishing Disclosure Rules for Authors and Publishers in 2026: What Must You Declare? · How do AI manuscript detection tools compare for authors and publishers in 2026? · How are authors and publishers protecting creative works from AI scraping and training in 2026?

What AI rights are actually being negotiated?

An AI rights negotiation usually involves at least four different rights. Reproduction rights determine whether a publisher or model builder can copy text, images, recordings, or datasets. Training rights concern whether those copies may be used to train or fine-tune an AI system. Output rights determine whether the system may reproduce recognizable passages, styles, characters, voices, or substantial substitutes. Commercial rights concern whether generated or model-assisted material may be sold, licensed, advertised, or used in products serving the same market as the original. These rights should not be treated as interchangeable: permission to scan a book for a search index does not automatically authorize training a generative model, and permission to train a model does not automatically authorize publishing its outputs.

The negotiation must also address records, transparency, and remedies. Authors can ask which works are included, what data was removed, whether the use is for research or commercial services, where processing occurs, whether human reviewers can access the material, and how complaints are handled. They can request notice before new model training, a prohibition on creating datasets from works covered by an opt-out list, and a commitment not to use the content to produce competing books. Payment may take several forms, including an upfront license, revenue share, per-use royalty, donation, or contribution to a collective fund. A higher upfront payment can look attractive while being worse than a modest royalty if the AI product becomes highly profitable and the contract permanently limits accounting rights. The strongest approach combines money with enforceable boundaries and information.

Why publishers and AI companies are negotiating now

The pressure comes from overlapping legal, commercial, and technical changes. News organisations worry that AI summaries can replace the visit to an original article, while authors worry that their books, papers, artwork, and recordings can become training material without payment or visible attribution. Reports in 2025 and 2026 about AI licensing in academic publishing illustrate the dispute: some authors reacted strongly after learning that access to their research had been sold to a large technology company. These incidents show why authors should not assume that an ordinary publishing contract, database agreement, or institutional policy already settles AI exploitation. The Reuters Institute has documented news-industry efforts to coordinate a response to AI companies, reflecting the fact that many organisations face similar questions but do not have identical bargaining power.

Regulation makes the timing more complicated rather than simpler. The European Union's AI Act entered into force on 1 August 2024 and is being implemented in stages, with many provisions applying in 2025 and 2026 and some high-risk-system obligations scheduled for later dates. The UNESCO Recommendation on the Ethics of Artificial Intelligence, adopted by 193 countries in November 2021, is not itself a copyright law, but it gives governments a reference point for transparency, human rights, consent, and accountability. The Universal Declaration of Human Rights and the UN Guiding Principles on Business and Human Rights provide related language about dignity, participation, and responsible business conduct. None of these instruments automatically creates a royalty for an author. They can nevertheless inform a negotiation position: a contract that prevents attribution, exploits labor without consent, or exposes people to unlawful processing may be commercially and ethically difficult to defend.

How to prepare before the first meeting

Preparation begins with an inventory of the author's work and the rights that already exist. Record the title, publication date, publisher, contract term, territory, language, subsidiary-rights clauses, and any provisions concerning electronic editions, databases, reprints, translations, or machine-readable copies. Identify whether the work is published under a work-for-hire agreement, whether copyright ownership has transferred, or whether the author retains some rights. In many publishing contracts, the author grants a broad exclusive licence while retaining copyright, but the precise wording matters. A rights request sent to the wrong party can delay negotiations, and a request sent to the publisher may not bind an AI company that later licenses content from a distributor.

Next, convert complaints into tradeable proposals. Instead of writing that the company is “stealing” the author's work, ask for a named licence covering defined uses, a stated duration, a geographic scope, and a payment mechanism. Authors can ask for a pilot covering 50,000 books or a six-month evaluation, rather than an undefined licence for the entire catalogue. They can propose a per-title opt-in, a minimum guarantee, revenue reporting, an audit right, and a termination right if the system creates outputs that substitute for the original. These numbers are negotiating examples, not legal requirements, but specific limits make a proposal easier to evaluate. The author should also prepare a fallback position showing which permissions are essential, which are negotiable, and which are unacceptable.

## Comparing the main negotiation options

Authors generally face four routes: a direct licence, a publisher-led collective agreement, an industry-wide collective scheme, or a refusal combined with enforcement. A direct licence can produce the clearest terms, but it requires identifying the actual user and negotiating individually. A publisher-led agreement may be faster because the publisher already controls a catalogue, yet it can give an author less visibility into downstream licences. A collective scheme can spread legal costs and create a standard price, but it may accept compromises that one author would not accept. Refusal preserves leverage, but enforcement can be expensive and technically difficult when evidence about training data, model outputs, or revenue is hidden.

Hybrid structures often work better than a single option. An author might first accept a limited research licence, then negotiate a commercial phase only if the company supplies dataset details, impact measurements, and payment records. A publisher could license a defined subset of works while preserving an author's right to object to new generative uses. A collective scheme could establish a minimum payment, while individual authors retain control over voice, likeness, characters, or highly sensitive works. The commercial result will depend on bargaining power, but a documented alternative gives the negotiator something better than a simple refusal.

What to say about payment, attribution, and control

Payment should be tied to measurable events. Authors can request an advance, a per-work fee, a revenue share, or a fixed fee for a limited period and territory. They should ask what counts as revenue: licence income, advertising revenue, subscription revenue, API income, or the entire product revenue attributable to the licensed material? If the company refuses to share figures, the author may prefer a larger guaranteed payment rather than an unverifiable royalty. Contract language should also address minimums, payment dates, currency, taxes, late fees, reporting frequency, and audit access. A royalty clause without a clear accounting method is not a real promise, because the author may never learn whether the system produced income from the work.

Attribution should be concrete. “Credit the author” could mean a credit in training documentation, a bibliography, a generated answer, a product interface, or nowhere visible to users. The contract should identify the permitted forms, such as name, title, publisher, ISBN, URL, and trademark, while acknowledging that a system may sometimes be unable to attach attribution to every output. A useful compromise is a visible citation when a source is displayed, a machine-readable source record in the dataset, and a public licence statement describing the category of works used. Authors should avoid requesting a guarantee that every answer will be accurate, because attribution can be present even when the model misstates the content. The right to correct or challenge harmful output is more realistic than perfect factual performance.

Control rights need separate treatment from copyright. Authors may want approval for voice cloning, face or likeness use, adaptation of characters, explicit material, or works used to create a competing title. They can set age limits, audience restrictions, geographic limits, and rules for confidential manuscripts or unpublished work. A clause requiring “reasonable efforts” to prevent misuse is often weaker than a clause requiring specified security measures, staff access controls, incident notification within a defined period, and compensation for losses caused by a breach. A 24-hour notification requirement may be unrealistic, but 30 days, 10 business days, or another clear period can be proposed. The author should ask what happens after termination: when will copies be deleted, will derived models be retired, and will existing products be allowed to continue using outputs permanently?

Common mistakes that weaken the author's position

The first mistake is negotiating from a slogan rather than a proposed contract. Language such as “I want AI to stop stealing” may be morally understandable but commercially difficult to convert into permissions. A useful proposal names the work, user, purpose, model, term, territory, payment, and exit conditions. The second mistake is assuming that copyright ownership alone determines the outcome. The publisher may have electronic rights, database rights, or subsidiary rights that it can license, while an author's contract may contain reversion provisions that are difficult or slow to exercise. The third mistake is accepting “exposure” as the only benefit. A model card, social-media post, or conference appearance may create awareness, but it does not replace payment when a commercial product directly uses protected expression.

Another mistake is signing a broad agreement without checking who can bind the company and whether the agreement covers affiliates, contractors, and later acquisitions. A contract with one AI developer may not control a separate cloud provider, model host, or corporate customer. Authors also make the error of treating human editing as automatic copyright protection. Some jurisdictions and courts examine whether human contribution is sufficiently original; merely adding a cosmetic change to generated work may not create protectable authorship. Finally, waiting too long can weaken a practical position if a company has already used the work in a product. That does not remove legal options, but it can reduce the chance of a clean negotiated licence and may require a complaint about ongoing use rather than a request for permission before deployment.

When to act, and what it may cost

Authors should act before signing a new publishing contract, granting a new digital licence, or allowing a work to enter a large commercial dataset. A review is also sensible before an exclusive AI licence is renewed, when a publisher announces an AI training programme, or when a platform offers a settlement that waives claims against multiple companies. The urgency can be expressed in a 30-day response period for information, a 60-day period for a proposed licence, and a defined date for resolving objections. Authors should not sign a rights waiver that says they will not challenge future uses unless the payment and scope are unusually clear. If the work is already used without permission, gather contracts, invoices, screenshots, model outputs, correspondence, and evidence of revenue impact before sending a formal complaint.

Costs vary by catalogue size and dispute complexity. A rights-clearance review by a publishing or media lawyer may cost roughly US$2,000 to US$15,000 for a moderate catalogue, while detailed negotiation or an audit-clause review can run from US$5,000 to US$50,000 or more. Hourly legal fees may range from about US$150 to US$600 depending on the lawyer and jurisdiction, but these are planning ranges rather than quoted rates. A small author can reduce expense by preparing a one-page rights summary, selecting the most important works, and using a collective organisation or recognised advocacy group. The economic decision should compare the likely licence value with legal fees, delay, reputational exposure, and the value of preserving future options. An AI publishing consultant can help organise the commercial case, but legal advice remains necessary when copyright ownership, privacy, publicity rights, or litigation risk is involved.

A workable negotiation sequence

The sequence usually works best in stages. First, identify the counterparty and verify the chain of title. Second, send a concise demand for the dataset description, intended use, model category, territory, term, payment proposal, and contractual clauses. Third, ask for a written response within 30 days and require silence to mean no permission unless the contract expressly says otherwise. Fourth, negotiate from a defined option sheet with a preferred deal, a fallback deal, and a walk-away position. Fifth, store the final agreement, licence scope, payment records, and any objection notices in a permanent rights file. This process does not guarantee a favourable deal, but it prevents a casual business conversation from becoming an accidental rights transfer.

The strongest outcome is not simply the highest one-time payment. It is an agreement that identifies what was licensed, separates research from commercial deployment, pays for defined uses, allows reliable accounting, protects author identity and control, and provides a practical way to exit. Authors should remember that AI technology changes faster than contracts. A 2026 clause may need review after a model becomes a subscription service, an API, an advertising system, or an acquisition target. The safest negotiating principle is therefore to negotiate the use, not only the product name.