Free to Read Does Not Mean Free to Train On

Reading about the NY Times – OpenAI/Microsoft suit, I read a quote by Satya Nadella from his deposition that I found almost offensive.

According to the Times’ reporting, Nadella testified that “anything that is paywalled should be licensed by anyone who wants to use it.”

What about the rest of us, Satya?

I run a publishing business. We pay people to report, interview, write, edit and publish. Readers can access our journalism without buying a subscription, but that does not make our work product available for another company to take and use however it pleases.

Why should putting up a paywall make a difference? The work is still copyrighted. It still costs money to produce. And making it free to read does not mean we gave OpenAI, Microsoft or anyone else permission to use it to build their products.

To be clear, Nadella did not say everything outside a paywall was fair game. But his emphasis on paywalled content struck a nerve. Publishers who fund their work through advertising, sponsorships and other revenue sources have just as much reason to object to its unauthorized use as publishers who charge readers directly.

The Reader’s Price Is Not the Cost of the Work

At Techstrong, the article a reader sees is the result of work that begins long before somebody hits “publish.” Finding the right source, getting beyond a prepared talking point and figuring out whether a claim holds up all take time. So does developing enough knowledge of an industry to ask a useful follow-up question.

We make that investment because we want people to read the work and come back for more. The audience is part of how the business supports itself. Giving readers access without a subscription is a decision about how we build that relationship.

The U.S. Copyright Office explains that copyright protection generally begins when an original work is created and fixed in a tangible form. A subscription payment is not part of that requirement. Neither is a paywall.

The facts reported in an article are not themselves copyrighted. Its original expression can be. That distinction applies whether you paid to read the article or opened it free on your phone.

Our objection also goes beyond wanting somebody to acknowledge our ownership. We are talking about our work product being collected and used to build another company’s product without our permission. Calling it “training data” describes its role in that company’s process. It does not explain where the permission to use it came from.

A publisher can invite the public to read an article without issuing an unrestricted license for every use a technology company might subsequently find profitable.

People in the software business should understand this.

I Have Been Making This Distinction for Almost 15 Years

On July 4, 2012, I published a Network World column titled “Free as in freedom, not free as in beer.” I was writing about a confusion that still follows the technology industry around: Treating something that costs nothing as though its price tells you everything about your rights to use it.

That was almost 15 years ago, before some of the people weighing in on today’s AI debate had begun their careers. I have been writing about, speaking about and refining my thinking on these questions for almost 15 years. My support for open source and broader access to technology is well established. So is my understanding that freedom, price, ownership and permission are different things.

That history is part of why Nadella’s statement bothered me. I believe in making work accessible. I understand the commercial value that can grow from allowing others to build on it. But the terms under which that happens matter. Open source gives us a mature, practical example of how to grant substantial freedoms while retaining copyright.

Consider the MIT License. It begins with a copyright notice, then grants broad permissions to use, copy, modify, distribute and even sell the software. Those permissions are free of charge. They also come with a requirement to preserve the copyright and permission notices in copies or substantial portions of the software.

The license does the work. The zero-dollar price does not.

We should be equally clear about what this analogy does not establish. Open-source licenses permit commercial use. A company complying with a permissive license generally does not have to negotiate a new payment or seek individual approval every time it uses the software. That freedom is part of the point.

A freely accessible news article, however, does not automatically arrive with an MIT-style grant of permissions. The reader’s ability to open the page tells us about access. It does not establish an equivalent right to copy and repurpose the work.

The software industry has built an enormous commercial ecosystem around understanding licenses. Somehow, when the raw material becomes somebody else’s reporting, we are asked to accept public availability as an adequate explanation.

I would like a better explanation.

Fair Use Requires More Than “We Found It Online”

There is a legitimate answer AI companies can offer: Fair use. It deserves a serious discussion.

Copyright does not give me an absolute veto over everything somebody does with an article I publish. Quotation, criticism, research and other uses can qualify for protection without my permission. Whether a particular AI training use qualifies requires examining the circumstances.

In its May 2025 prepublication report on generative AI training, the Copyright Office anticipated that some training uses would qualify as fair use and others would not. That is the Office’s analysis, not a binding court ruling. It recognizes a range of uses, purposes and effects that cannot be resolved with one blanket answer.

A paywall can matter to questions about how material was obtained. Deliberately getting around an access restriction may raise issues beyond those involved in visiting an unrestricted page. But the absence of that barrier does not eliminate the copyright in the article.

Nor does commercial use automatically defeat fair use. The argument has to be more careful than that. A company asserting fair use should defend its particular use on the relevant facts, rather than treating an open page as proof that permission was unnecessary.

The economics deserve equal care. Publishers have reason to ask what happens when products built using their work become substitutes for visiting their publications.

The Times reported that OpenAI employees anticipated increasing substitution, while Microsoft researcher Brent Hecht warned about damage to the supply of content on which models depend. Microsoft said Hecht’s comments did not represent the company’s views. These are reported internal discussions, not findings that the companies violated copyright law.

Still, the business question is real. Who pays for the next round of reporting if the audience increasingly gets its answers somewhere else?

A source link does not necessarily solve that problem. It can identify who did the reporting while leaving that publisher without the visit that helps support the work.

Pew Research Center examined browsing data from 900 U.S. adults. Users clicked traditional search results on 8% of visits with an AI summary, compared with 15% of visits without one. They clicked links within the summaries on just 1% of visits where a summary appeared.

Those findings concern Google search behavior. They do not prove that training caused a particular publisher’s losses, and training a model is distinct from retrieving an article or presenting an answer. We should keep those distinctions intact.

But the findings support a practical concern: Giving readers a satisfactory answer can reduce their reason to visit the source. For a publisher whose revenue depends on reaching an audience, that is a substantial business issue even when the original article has no subscription price.

Must We Close the Door to Have a Say?

Follow the paywall distinction to its potential consequences.

If putting an article behind a subscription screen makes an AI company more likely to seek permission, publishers have another incentive to put articles there. Readers get less access. Smaller publications face more pressure to erect barriers, whether or not subscriptions fit their audiences.

That would be a miserable outcome for an industry that spent decades using the web to make specialized knowledge more accessible.

There are difficult questions on the other side. Licensing enormous collections of work can be expensive and complicated. Requirements that only the largest AI companies can afford could strengthen their position against smaller competitors.

Those concerns deserve workable answers. Licensing arrangements for uses that require permission should be practical, with room for smaller developers and smaller publishers. Legitimate fair use should remain available. Some owners will grant broad permissions freely; others will seek compensation or decline uses they are entitled to control.

The difficulty of designing that market does not establish that everybody who publishes without a paywall has already agreed to participate on whatever terms an AI company chooses.

I support AI. I use it. I also believe in making journalism accessible. Those positions are perfectly compatible with objecting when the work we finance is used without our permission to build products that may compete for the same audience.

Nadella recognizes a principle worth extending into a fuller conversation. Publishers outside the subscription business have work product, ownership and economic interests, too. We should not have to put up a payment screen to make those interests visible.

We made our journalism free for people to read. That does not mean we gave AI companies permission to use our work product to build theirs.