I Almost Told the Reporter There Was Nothing Here. Then I Read Meta’s Terms Again.

By Babak Akhlaghi on November 13, 2026.  When Stephen Beale first reached out to me, I was polite but skeptical.

He is a seasoned reporter, and he wanted my take on an update Meta had pushed to its business terms. My honest first reaction: this is routine. Companies update commercial terms constantly. It is standard operating procedure, and most updates are housekeeping.

Then I read the actual language.

Buried in the update was a provision that was anything but housekeeping. By using Meta’s platforms, business users would consent to Meta crawling their content and using it for training. The language was broad. Broad enough that it would reach material that, in an open licensing market, would cost a fortune. As I put it at the time, it was remarkable to bury something like that in terms of service, given that this material would likely cost many millions of dollars in a licensing market.

That outreach turned into one of the more instructive episodes of my career. Here is what actually happened, and what it means for you if you run a company that either signs these agreements or writes them.

What Meta Was Really Doing: Moving the Fight to Friendlier Ground

To understand why this clause mattered, you need to understand where AI training law stands right now.

Every major fight over AI training data today runs through copyright, specifically fair use. And fair use is a mess for defendants. It is fact-intensive, it is case by case, and the law is being developed as we speak.

Courts and regulators are splitting hairs in real time. The U.S. Copyright Office concluded there will be no single answer on whether training AI on copyrighted works is fair use. Whether training is transformative is, in the Office’s words, a matter of degree, and where a model produces content that shares the purpose of the original, the use is at best modestly transformative.

Meanwhile, in Bartz v. Anthropic PBC case, the court called the training use transformative, and fair use, yet still ruled that maintaining libraries of pirated books was not fair use. Having read the copyright cases on AI and fair use, it is fair to say that unauthorized or pirated use of content doesn’t cut in favor of finding a fair use but it is not determinative alone as there are other factors in play.

So the fair use questions are hard. Is the use transformative. How much is copied. Was the data pirated or obtained without authorization. That last one cuts hard against fair use, and the AI companies know it.

Meta’s clause was an attempt to exit that entire arena.

If Meta holds a contractual right to use your data, the fair use analysis becomes irrelevant to your dispute with Meta. The question stops being copyright and becomes contract. And contract law is fairly clear. Either you granted the rights or you did not. There is nothing stronger in litigation than a contractual relationship in which the customer has agreed to training on their data.

As a legal strategy, I respect it. It is smart. The problem was the execution.

The Weak Point: Notice and Assent

Updating your commercial terms does not automatically create a binding agreement. Two requirements do the real work: proper notice and proper assent.

One of the leading cases is Douglas v. U.S. District Court ex rel. Talk America, 495 F.3d 1062 (9th Cir. 2007). Talk America added an arbitration provision, a class-action waiver, and other new terms to its customer agreement, then posted the revised version on its website without telling customers. When Douglas filed a class action, Talk America tried to compel arbitration. The Ninth Circuit refused. The revised terms were a new offer that required notice and acceptance, and because Douglas never knew about the changes, he could not have assented to them.

Here is where Meta’s approach could get shaky. I don’t know exactly how Meta notified users of the update, or whether the businesses affected by it ever saw the change. If the notice amounted to little more than a revised terms page, Douglas suggests it would not hold. Even a general announcement that terms are changing may fall short when the most consequential change, a new right to crawl business websites and use that content for training, sits buried inside it. Posting something quietly and hoping continued use counts as agreement is a gamble.

Meta can argue that by continuing to use the product, you assented. That argument can work. But there is a complication: when crucial terms are hidden inside a routine update, the other side could argue that the company owes users specific assent for those provisions. Practice guidance already points this way. For complex agreements, companies ask for affirmative assent at multiple points, a checkbox or initials, to call attention to specific clauses and confirm agreement to each one individually.

A clause that grants training rights over your entire web presence is exactly the kind of provision that deserves its own checkbox. Burying it was the mistake.

Why Meta Wanted This Clause at All

People asked me whether Meta planned to go after small creators, small startups, and individual inventors with this contract. The answer is likely no. Meta is not in that business, and enforcement against small parties would be a public relations disaster for minimal gain.

The real value is defensive. It is a bargaining chip.

If somebody comes after Meta for training on their data, Meta gets to say: this is a contract question, and we have a valid contract. That single sentence changes the entire posture of the dispute. The plaintiff arrives ready to fight about transformativeness and market harm, and instead finds themselves arguing about whether they clicked a button in 2026.

For a company facing years of unsettled copyright litigation, that is worth a great deal. Certainty is the product here. The contract manufactures certainty in a legal environment that offers very little of it.

What Happened Next Surprised Even Me

Steve published. Other publications picked up the story. Then it reached Washington.

Senators Adam Schiff and Marsha Blackburn, from opposite parties, wrote to Meta asking it to withdraw the terms. Meta retreated in stages. First it removed the word “train” from the list of permitted uses and narrowed data applications to advertising and business products. Removing that one word eliminated the specific language that let critics call the terms an AI training rights grab.  Eventually, however, Meta paused the terms indefinitely.

Sit with that for a second. A reporter who broke the story first and an IP attorney who almost wrote off his email as nothing, working with very limited resources, forced one of the largest companies on earth to shelve its commercial terms. That is the beauty of effective journalism. It does not require an army. It requires someone who reads the fine print and someone who can explain why the fine print matters.

Why Founders Should Care: You Sit on Both Sides of This Clause

If you run a company, this story applies to you twice.

As the party signing agreements: read the updates. The most consequential provisions now arrive as quiet revisions to terms you accepted years ago. Continued use of a product can bind you to terms you never consciously reviewed. Assign someone to track terms changes from your critical vendors, especially platforms that touch your content or data.

As the party drafting agreements: learn from Meta’s execution failure. If you want enforceable rights to customer data, the clause needs prominence and specific assent. A separate checkbox for the data provision costs you some conversion. A clause a court refuses to enforce costs you the entire strategy. Hidden terms feel clever until a judge or in this case our seasoned reporter reads them out loud in an opinion or an article.

The underlying lesson is simple: contract law rewards transparency. The strongest data rights clause in the world is the one your customer knowingly agreed to.

FAQ

Did Meta do anything illegal by proposing these terms?

Proposing terms is lawful. The legal question was whether the terms would be enforceable given how they were introduced.  Did they amount to a proper notice to customers and a proper assent by them.

If I kept using Facebook or Instagram during this period, did I consent to training?

No.  The proposed business terms were paused before taking effect.

Is a contract really stronger than a fair use defense?

Yes, against the contracting party. Fair use requires a court to weigh multiple factors on facts specific to each case. A valid contract granting data rights ends the copyright inquiry between those parties before it starts. That is precisely why Meta pursued this route in my opinion.

Does the Anthropic ruling mean AI training is legal now?

No. Judge Alsup found the training use transformative but held that acquiring and maintaining pirated source material was unlawful. How the data was obtained matters independently of what you do with it.

What should my startup do about its own terms of service?

Put consequential data provisions in plain sight and collect specific assent for them. Checkbox, initials, separate acknowledgment. Courts enforce clickwrap agreements that show clear notice and affirmative assent. Build those two elements into your process and your terms will hold.

Can individuals really influence a company like Meta?

This case proves it. One reporter’s question, one attorney’s analysis, coverage by other outlets, letters from senators, and an indefinite pause. Each step was small. Together they worked.