Meta’s removal of the word “train” from its commercial terms looks like a concession. Ars Technica reports that it will not add language respecting robots.txt, and that is the signal website owners actually control, so the edit protects publishers very little.
Key takeaways
- Ars Technica reports that Meta has removed the word “train” from its commercial terms of service following criticism over how it gathers data for artificial intelligence.
- According to the same report, Meta has declined to add wording that commits it to respecting robots.txt, the file websites use to tell automated crawlers what they may access.
- Changing a single word in a contract does not by itself show that the underlying data practices have changed, and the replacement wording has not been confirmed in the material available.
- For website owners, a public commitment to honour machine-readable opt-out signals would offer clearer protection than changes to how a terms document is phrased.
The argument: wording is not the same as restraint
The core claim of this article is simple. Changing the vocabulary of a terms of service document does less for the people whose content feeds AI systems than a written commitment to obey the technical signals those people already publish. Ars Technica reports that Meta has made the first kind of change and refused the second.
The word “train” has become shorthand in disputes over AI. When a company says it may use material to train models, readers understand that their text, images or data could shape a commercial system they have no stake in. Removing the word takes away the most visible flashpoint. Whether it removes the practice is a separate question, and the information available does not answer it. The source material does not say what wording, if any, replaced the term, or whether the rights Meta holds under the terms have become narrower, stayed the same or simply been described differently.
Robots.txt matters here. It is the long-established way for a website to tell automated visitors which parts of the site they may crawl. A company that commits in writing to honour it gives every website owner a way to opt out that they can use without negotiating, suing or reading a legal document. A company that declines to write that commitment down leaves the matter to its own discretion. The rest of this article sets out why that difference matters more than the edit, then gives the strongest objections to that view.
Terms of service govern the relationship, not the crawler
A terms of service document is a contract between a platform and those who use its products. In a commercial context it typically covers businesses and developers who build on or alongside the platform’s tools. It sets out what each party may do with the other’s material within that relationship.
That scope is the first piece of evidence that a wording change has limited reach. The concern behind much of the backlash over AI data collection is not only what a company does with material users hand over. It is also what the company’s automated systems collect from the open web, from sites whose owners have never agreed to any terms at all. A news publisher, a small business or a hobbyist forum is not a party to Meta’s commercial terms. Editing those terms does not change Meta’s legal relationship with them.
There is also a practical point about how contracts are drafted. Legal language can grant broad rights without naming specific activities. Phrases about using, analysing, improving or developing products and services can cover a wide range of processing, including the kinds of model development that the word “train” describes more bluntly. Without the new text, nobody outside the company can say whether the change narrows Meta’s rights or just rewords them. The honest position is that it is not known. That uncertainty is a reason to be cautious about treating the edit as a policy reversal.
None of this means the edit is meaningless. Contracts are interpreted by their words, and deleting an explicit grant could matter in a future dispute. But a wording change inside a contract that most affected website owners never signed is a narrow tool for the problem it is being presented as answering.
Robots.txt is the signal publishers actually control
The second piece of evidence concerns the instrument Meta has reportedly chosen not to adopt. Robots.txt is a plain text file at the root of a website. It lists rules for automated agents, usually identified by a name they announce when they request pages, and says which paths they may or may not visit. The convention is known as the Robots Exclusion Protocol. It has been used across the web for decades, and well-behaved search engines and other crawlers generally follow it.
What makes robots.txt important in the AI debate is that it puts the decision with the website owner. A site does not need lawyers, a contract or any relationship with the company running the crawler. It publishes a rule, and a compliant crawler obeys it. In recent years several AI developers have published the names of their crawlers so that site owners can block them specifically, which turns robots.txt into a rough but workable opt-out from AI data collection.
The protocol has a well-known weakness: it is voluntary. Nothing in the file itself forces a crawler to comply. That is why a written commitment matters. When a company states in a public document that it honours robots.txt, it creates an expectation that outsiders can check and hold it to. It also gives regulators, courts and journalists a benchmark against which to measure observed behaviour. When a company declines to make that statement, compliance remains a choice it can revisit without notice.
Ars Technica’s report that Meta changed its AI scraping policy while refusing to add robots.txt language therefore points to an imbalance. The visible, symbolic change has been made. The change that would hand control to the owners of the content has not. The source material does not explain Meta’s reasons for declining, and this article does not speculate about them.
Platforms already treat scraping rules as binding when they are the target
The third piece of evidence comes from how large platforms generally approach scraping of their own services. Major social networks, Meta among them, have long restricted automated collection of content from their sites without permission, both in their terms and through technical measures. Platforms have treated unauthorised scraping as a security and privacy problem, partly because bulk collection of user profiles can enable fraud, harassment and other abuse.
That stance is reasonable. It also shows that large technology companies do not regard the rules on automated access as trivial. When their own data is at stake, they write restrictions into contracts, build systems to enforce them and pursue those who break them.
Set beside that history, a refusal to commit to other websites’ published crawling rules is hard to explain away as a technicality. The principle that a site owner should decide how automated agents use its content applies whichever way the crawler is pointed. Consistency would suggest that a company which defends its own boundaries should state plainly that it respects the boundaries others publish. The edit to the word “train” does not touch that question.
The strongest case against: robots.txt is a blunt and imperfect tool
There is a serious argument that this article overstates the importance of robots.txt and understates the value of the terms change.
First, robots.txt was not designed for the AI era. It tells a crawler whether it may fetch a page. It does not say what the crawler may do with the page afterwards. A site that wants to appear in search results but not in AI training data cannot always express that through robots.txt alone, especially if one company runs a single crawler for several purposes. A contractual promise to honour robots.txt could give a false sense of protection while leaving the main question, how collected material is used, unresolved.
Second, robots.txt is not a legal standard in most places. Writing it into commercial terms might create obligations whose scope is unclear, or invite disputes over badly written or contradictory files. A company could reasonably prefer to follow the protocol in practice while avoiding a contractual commitment to a convention it does not control and that may be replaced. Standards bodies and regulators have been discussing more expressive ways for content owners to state their preferences on AI use, and a company might decide to wait for those rather than commit to an older mechanism.
Third, the terms change may matter more than it appears. If the explicit grant to use material for training has been removed and not replaced with equivalent wording, the commercial customers covered by the terms may now hold stronger rights over their data than before. For those parties the contract, not robots.txt, is the instrument that counts, and a narrower grant is a real gain.
Finally, refusing to write something into a contract is not the same as refusing to do it. Meta may already follow robots.txt in practice. The source material does not say whether it does or not, and that absence of information cuts both ways.
These points are fair. They show that robots.txt is not a complete answer. They do not show that a written commitment to it would be worthless, only that it would be one layer among several that are needed.
Evidence that would change this conclusion
Several developments would justify revising the view taken here.
The most direct would be publication of the full revised terms alongside the previous version. If the comparison showed that Meta had substantially narrowed its rights to use commercial customers’ material for model development, the edit would carry more weight than this article gives it.
A second would be clear public documentation from Meta of the crawlers it operates for AI purposes, the names they use and how they treat robots.txt. Independent analysis of server logs by website operators or researchers, showing consistent compliance with published rules, would also weaken the concern, even without contractual language.
A third would be regulatory or legal developments that make machine-readable opt-outs enforceable regardless of what any company writes in its terms. In the European Union, copyright rules already recognise reservations against text and data mining expressed in machine-readable form, and AI-specific legislation has tied general-purpose model providers to respecting them. Wider and clearer enforcement of such rules would make a voluntary commitment in Meta’s terms less significant.
Conversely, evidence that the new wording preserves the same rights under different words, or that Meta’s crawlers ignore published restrictions, would strengthen the conclusion that the change is mainly cosmetic.
Sources and further reading
- Ars Technica, technology policy coverage reporting Meta’s removal of the word “train” from its commercial terms and its refusal to add robots.txt language.
- Documentation of the Robots Exclusion Protocol, describing how robots.txt files instruct automated crawlers.
- European Union copyright and AI legislation, covering machine-readable reservations against text and data mining.
- Published crawler documentation from AI developers, explaining how website owners can identify and block AI-related crawlers.
Surfaced from the rss:arstechnica signal “AI scraping terms change”. AI-assisted draft, editorially reviewed.
