Your Old Magazine Article Is in the AI Deal. Your Check Is Not.

In May 2024, The Atlantic announced a licensing and product partnership with OpenAI. The deal would make Atlantic material discoverable through OpenAI products with attribution and links, while the magazine would develop an experimental site using OpenAI technology. The financial terms were not disclosed.

The announcement concerned an archive built partly by freelancers who had been paid once, article by article, under agreements written before publishers began treating generative AI companies as customers. The Atlantic did not publicly announce a freelancer compensation pool alongside the deal. That absence does not prove no contributor received money through some private arrangement. It does expose the operative assumption: the company could negotiate access to the archive as company property.

Return to the old contributor agreement. Somewhere near the payment amount and the promise that the work is original, many contracts contain the phrase that keeps earning long after the writer stops: use in all media now known or later developed. The wording varies. Its purpose does not.

It prevents a new format from forcing a new negotiation.

AI licensing has made that boilerplate newly expensive for the people who signed it.

The deal happens above the byline

OpenAI’s publisher agreements arrived in quick succession. The company announced arrangements with Axel Springer, the Financial Times, News Corp, Vox Media, The Atlantic and other media owners. Public descriptions differ, but they commonly refer to access, display, attribution, product development or the use of publisher material in AI systems. News Corp’s announcement explicitly covered current and archived material from publications including The Wall Street Journal, Barron’s and the New York Post.

The checks go to the companies negotiating the package. Those companies may employ reporters, commission freelancers, license wire copy and host decades of work acquired under several generations of contract language. An archive page makes those distinctions look tidy. Copyright does not.

A freelancer normally begins as the copyright owner unless the work qualifies as a work made for hire, a category that can make the commissioning party the legal author, or the writer signs ownership away. Most disputes begin after that sentence, because a publisher does not always need to own the copyright outright. It only needs a license broad enough to authorize what the AI company wants.

That license may be perpetual, meaning it has no end date; worldwide; transferable; sublicensable, which allows the publisher to grant rights to another company; and broad enough to cover reproduction, distribution, adaptation or use in later-developed media. A writer can retain copyright in theory while giving the publisher most of the commercially useful permissions in practice.

The byline survives. The bargaining power does not.

The missing trigger

Freelance payment clauses often describe the original fee as full compensation for the rights granted. If a later use falls within that grant, no fresh payment event occurs. The publisher has not ordered another article, and it may not classify the AI agreement as syndication, anthology publication or a conventional reprint, categories that older contracts sometimes address more clearly.

This is where reuse compensation disappears. The money is real, but the contract contains no trigger that routes any of it back to the writer.

Consider the later-developed-media clause attached to The Atlantic archive. Generative AI did not need to be named when the work was commissioned if the agreement already anticipated unknown delivery systems, broad digital use or sublicensing. Whether a specific deal falls within a specific contract depends on its wording and the undisclosed licensing terms. The structural advantage is plain without pretending those private documents are public: the publisher has the files, the contract records and the negotiating relationship, while each freelancer has one contribution inside a bundle.

Bundling matters. An AI company is unlikely to price a licensing agreement as a row of individual articles with a visible value beside every byline. It is buying access to a corpus, meaning a body of material assembled for search, display, analysis or model development, along with the publication’s reputation and technical delivery. Once thousands of works, trademarks and product commitments enter one agreement, management can say there is no practical way to identify what any single article earned.

That difficulty is designed into the transaction. Publishers already track pageviews, subscriptions, syndication and royalty-bearing uses when those numbers affect their own revenue. They can build a contributor ledger if payment is part of the deal. Without a contractual obligation, accounting inconvenience becomes policy.

Training is only one permission

Public debate tends to collapse every AI use into training, the computational process through which a model adjusts its internal parameters using large collections of material. A publisher deal can cover more than that. It may authorize article summaries, quotations, retrieval inside a chatbot, links, search features, product experiments or access to an archive through a technical feed.

Those uses engage different rights and create different exposure. A contract that does not support model training might still permit syndication-like display. A broad adaptation clause might be invoked for summaries. A sublicensing provision may let the publisher pass certain permissions to an AI vendor even though the freelancer never dealt with that vendor and would not recognize its systems as a publication.

The distinction also helps explain why the industry’s preferred promise, attribution with a link, is inadequate. Credit can matter. Referral traffic can matter. Neither substitutes for payment when the publisher receives consideration for access to work it did not create.

The Atlantic archive is useful here because its public announcement emphasized citation and links while leaving the financial arrangement private. The reader sees a source label. The writer sees no accounting statement. Both features can coexist with a profitable license.

Ownership settles leverage before the deal begins

Publishers have presented licensing as the responsible alternative to unauthorized scraping. There is a defensible point buried inside the sales pitch. A negotiated license can set limits, require attribution and establish that journalism has monetary value. The problem is that media companies often define journalism’s value at the corporate border, then stop counting once the money crosses it.

Freelancers remain outside many protections won by staff unions. A collective bargaining agreement may regulate how employee work is used, require consultation or establish AI rules, but it generally does not rewrite every independent contributor contract in the archive. The freelancer may also have little practical ability to negotiate. Refusing a broad rights clause can mean losing the assignment before reporting starts, after time has already been spent pitching and corresponding.

AI did not create this imbalance. It found a well-labeled cabinet.

Magazine and newspaper contracts have long sought permissions for databases, electronic editions, syndication and formats that did not exist when the first agreement was drafted. Court fights over digital archives pushed publishers to tighten language. Later-developed-media wording became a form of insurance: the company pays the known fee today and reserves room for unknown revenue tomorrow.

Now the unknown revenue has a buyer with a vast appetite for text. The clause attached to an old feature, review or investigation can do more economic work than the article’s original fee ever suggested.

A payment system is possible

The alternative does not require valuing every adjective. A publisher could reserve a defined share of licensing revenue for contributors whose work sits inside the licensed archive, then allocate it by article count, usage records or another disclosed formula. A deal could exclude freelance work until contributors opt in. Contracts could reserve machine-training rights to the writer, require a separate fee for sublicensing to AI companies or create an audit right so contributors can learn where their work went.

Each model has flaws. Article count ignores differences in use and labor. Usage-based payment depends on records supplied by companies with incentives to minimize what counts. Opt-in systems burden writers with monitoring old clients.

Still, these are administrative objections, not physical laws. Publishers manage complicated royalty and rights systems whenever a powerful licensor demands them.

Better contract language would separate ordinary editorial reuse from machine uses and prohibit the publisher from treating an unspecified future medium as consent to every future commercial purpose. That change would cost publishers leverage. It would also force AI companies to confront the fragmented ownership beneath the polished archive they want to buy.

For now, the later-developed-media clause keeps doing its quiet work. The AI company pays for scale. The publisher supplies it. The freelancer’s old check remains the only check because the contract made the future part of the original bargain, even though only one side knew it was buying insurance.

Questions people ask

Do freelancers automatically get paid when a publisher licenses an archive to an AI company?

No. Payment depends on the contributor agreement and any later arrangement. If the original fee covered a perpetual, sublicensable or broadly defined reuse license, the publisher may have enough permission to make the deal without creating another payment obligation.

Does retaining copyright protect a writer from an AI licensing deal?

Not by itself. A writer can keep formal copyright while granting a publisher extensive rights to reproduce, adapt and sublicense the work. The crucial issue is the scope of that license, including whether it covers digital systems, later-developed media or third-party commercial use.

Why can’t publishers divide AI licensing revenue among contributors?

They can. The obstacle is usually the absence of a contractual duty and a payment formula, not an inability to track works. Publishers could create a revenue pool, use article or usage records, and disclose the method, but doing so would redirect part of the deal away from the corporate rights holder.

Can attribution and links replace reuse payment?

No. Attribution identifies the source, and links may send some readers back to the publication, but neither gives the freelancer a share of the licensing fee. Credit addresses provenance. Compensation addresses who supplied the asset being sold.