AI Licensing Pays the Archive Before It Pays the Byline

The important line in Condé Nast’s AI deal is not in its announcement. It sits inside the contributor agreement a freelancer may have signed years earlier, in the section granting the publisher permission to reuse or sublicense the work.

That clause was drafted to cover syndication, databases, foreign editions and formats nobody wanted to enumerate. Public reporting on Condé Nast’s freelance terms has described broad rights grants that let the company exploit commissioned work across platforms. The freelancer received the agreed assignment fee. The contract did not need to predict a future market for machine training or an AI answer assembled from archived articles.

It only needed language broad enough to reach one.

In August 2024, Condé Nast announced a partnership allowing OpenAI to use content from publications including Vogue, The New Yorker, Wired and Vanity Fair. The announcement described attribution and links in OpenAI products, along with work on ways to support journalism. It did not announce a royalty pool for contributors, a per-use payment or a process through which freelancers could learn whether their work was included.

That silence is the mechanism. The archive can acquire a new customer without the byline acquiring a new contractual right.

The publisher owns the negotiating position

AI licensing announcements tend to present the publisher as the creator. A company says it has reached an agreement covering its journalism, archives or brands; the AI company gains reliable material and permission to use it; readers are told that licensed access is preferable to scraping. The individual article disappears into a catalog.

Some of that catalog was written by employees. Under the US doctrine of work made for hire, qualifying work created within employment generally belongs to the employer from the start. Staff writers and photographers are paid wages or salaries, not royalties each time an article is syndicated, indexed or licensed. Unless a union contract, employment agreement or new company policy creates additional compensation, an AI payment lands where earlier licensing payments landed: with the employer.

Freelance work can look different on paper. A freelancer may retain copyright while granting the publisher extensive rights, including the right to sublicense, meaning the publisher can authorize another company to use the work. Other agreements transfer copyright outright. Some grants are exclusive for a period; others are perpetual and global.

The wording varies, but the commercial result can converge. The publisher controls the bundle that an AI company wants to buy.

The sublicensing clause in that old Condé Nast agreement matters more than the current press release because it determines who can enter the room. OpenAI does not want to negotiate separately with every writer who filed a reported feature, every photographer who supplied a portrait and every illustrator whose work remains attached to an old page. It wants an archive at institutional scale, accompanied by a counterparty confident enough to license it.

Scale becomes leverage. Leverage becomes the check.

Training and retrieval are different uses with the same gatekeeper

The announcements often blur several forms of access. Training means using material to adjust a model’s internal parameters, the numerical settings that shape which outputs it produces. Retrieval means fetching relevant material from an outside database when a user asks something, then using that material to compose an answer. A deal may also cover summaries, excerpts, links or product development.

Those uses create different risks for a contributor. Training can absorb patterns from a body of work without showing the source to a user. Retrieval may surface a recognizable passage or summary while sending some traffic back to the publisher. A linked answer can still satisfy the user before the click.

Attribution is useful, but a byline displayed beside an AI response is not compensation.

The legal status of unlicensed training remains contested in US courts, and licensing deals do not settle whether every use required permission. They settle something narrower. Publishers have material AI companies value, litigation is expensive, and a contract can exchange access for money while reducing uncertainty. That bargain can make sense for both companies even when the underlying legal question remains open.

For the contributor, however, the distinction between training and retrieval may never reach the contract. An older rights grant commonly speaks in terms of reproduction, distribution, display, adaptation or sublicensing rather than model weights and generated answers. Broad language written before this revenue existed can still give the publisher room to claim that the new use belongs inside rights already purchased with the original fee.

The freelancer paid for that flexibility at the moment of signing, although nobody priced it as AI flexibility. That is the trick of future-proof contract language. It makes a new market look like an old permission.

The announcements disclose access, not distribution

OpenAI has announced licensing arrangements with publishers including Axel Springer, the Associated Press, the Financial Times, News Corp, Vox Media, The Atlantic and Hearst. The terms differ, and many financial details remain confidential. Across the public announcements, there is much more information about content access, attribution and product collaboration than about money reaching individual contributors.

No public promise of a contributor share does not prove that no writer anywhere receives one. A staff member may benefit indirectly if licensing revenue supports payroll. A union could bargain over AI income. A publisher might make private arrangements with particular rights holders.

In some jurisdictions, authors have statutory or collectively managed claims that do not exist in the same form under standard US freelance practice.

But indirect institutional benefit is not the same as a payment right. The rent covering an editorial office does not tell a freelancer whose photograph was licensed whether the photograph generated revenue, how the publisher valued it or whether the creator can withdraw it. “Supporting journalism” is the industry’s preferred unit of accounting because it cannot be audited by the person who made the journalism.

The Condé Nast sublicensing clause returns here. Once the publisher has acquired broad permission, it has little financial incentive to reopen the agreement and volunteer a second payment. Doing so would raise the price of the archive, require records matching licensed material to contributors, and create a precedent for other secondary uses. The cheap option is to treat the AI deal as corporate revenue produced by an asset the company already controls.

That asset has names on it. The accounting removes them.

A workable alternative requires records and bargaining power

Contributor payment does not require inventing a perfect per-word meter. Publishers already track authorship, publication history and rights status because they need to know what they can syndicate, republish or defend. An AI revenue pool could allocate a negotiated portion of licensing income to staff and freelance contributors, with rules based on inclusion in the licensed corpus rather than unverifiable claims about which article changed a model’s output.

Retrieval offers a more granular route. If an AI product fetches particular articles to build an answer, logs can record which sources were retrieved and displayed. A contract could require reporting and attach compensation to those uses. That would not capture every form of value, but it would move payment closer to the work than a general promise to preserve quality journalism.

There are costs. Publishers would need cleaner rights databases, especially for old photography, wire copy and commissioned work whose paperwork may be scattered. Revenue-sharing systems need audits and rules for deceased contributors or disputed credits. The work is administrative rather than impossible.

Media companies already perform comparable labor when money is owed to licensors instead of creators.

The harder obstacle is bargaining power. A freelancer presented with a standard agreement can reserve AI training rights, ask for an additional fee or refuse a sweeping grant, but the publisher can assign someone else. Individual negotiation works best for people who already possess the leverage that a shared system is supposed to create. Collective bargaining, statutory remuneration rights and coordinated contract standards can change the default because they prevent the publisher from pricing each refusal as one replaceable worker’s problem.

Until then, the old clause keeps doing new work. The assignment fee paid for the article once. The sublicensing language lets the archive sell access again. The AI company gets a clean counterparty, the publisher gets a revenue line, and the person named above the text gets attribution if the product is designed to show it.

The press release arrives years after the contract. The contract has already decided where the money stops.

Questions people ask

Do freelance writers automatically get paid from AI licensing deals?

No. Payment depends on the contributor agreement, any applicable collective agreement and the structure of the publisher’s deal. If a freelancer transferred copyright or granted broad sublicensing rights without a revenue-sharing clause, the publisher may collect licensing income without owing an additional fee.

Are

AI training and AI retrieval the same thing?

No. Training uses material to adjust a model’s parameters, while retrieval fetches source material when producing a response. Both can be licensed, but retrieval can be tied more readily to specific articles because the system records which sources it fetched.

Do public publisher deals disclose how contributors are compensated?

Most major announcements emphasize archive access, attribution, links and collaboration rather than contributor payments. That leaves writers and photographers unable to determine from the announcement whether a revenue pool exists, which work is covered or how any share would be calculated.

What would make contributor payment possible?

Publishers could reserve a negotiated share of licensing revenue, maintain records connecting archived work to its creators and report retrieval use. Collective agreements or statutory remuneration systems could make those terms harder to avoid than one freelancer’s request added to a standard contract.