Meta drops the word "train" from its commercial TOS after AI backlash

Facing mounting public pressure, Meta has made a seemingly minor but significant change to its commercial terms of service: it deleted the word “train.” According to Ars Technica, the social media giant revised the wording of its AI scraping policy while making clear that it would not add a clause committing to respect the robots.txt protocol. In other words, Meta is willing to make a concession on paper but unwilling to impose constraints on its actual conduct.

Deleting One Word, Adjusting a Posture

The controversy began with wording in Meta's commercial terms about how user content may be used. Previously, the terms allowed Meta to use content posted by users to “train” its AI models, wording that outsiders interpreted as the direct expropriation of creators' labor. As the boundary between AI-generated content and original content becomes increasingly blurred, more users, publishers, and creator organizations have voiced dissatisfaction, arguing that platforms are turning public content into “fuel” for model capabilities without explicit consent.

On paper, Meta's deletion of the word “training” appears to make the terms less aggressive. But critics were quick to point out that deleting a verb does not mean changing the underlying logic of its behavior. As long as the platform retains the ability to scrape public content at scale, softening the wording is more of a public-relations strategy than a policy shift.

“Changing the wording is not the same as changing the practice. The real question is whether platforms are willing to give content creators the technical means to say ‘no.’” — This has been the core question raised repeatedly in the controversy.

robots.txt: The Moral Weight of a Technical Protocol

Another focal point of the dispute is Meta's refusal to commit in its terms to respecting robots.txt. Born in 1994, this text file is the oldest “gentlemen's agreement” between websites and crawlers: website administrators use it to tell automated programs which pages may be crawled and which should be avoided. It has no legal force, yet it has long been regarded as one of the cornerstones of internet self-governance.

In the AI era, the significance of robots.txt has been magnified anew. For news sites, independent blogs, and content platforms, it is one of the few technical tools through which they can express a wish that their content not be crawled. Yet when AI companies conduct large-scale data collection to train models, the constraints of robots.txt appear quite fragile. Meta's reluctance to write it into its terms means the company wants to retain discretion over scraping—even if that means overriding website operators' explicit wishes.

It is worth noting that Meta is not alone in doing this. In recent years, several AI companies have faced similar questions about their data sources, and the common industry pattern is: technical capability comes first, rule-based constraints follow, and limited adjustments are often made only under public pressure.

Industry Context: Data Hunger and Copyright Anxiety

To understand why Meta is so cautious in its wording, one must look at the data logic behind the large-model race. High-quality corpora are a scarce resource for training high-performance models, and public webpages, forum discussions, news reports, and user-generated content constitute the most important data sources. As model scale expands, demand for fresh, diverse, and authentic text only grows.

At the same time, content creators and publishing institutions are waking up to copyright. From news groups to independent writers, more and more parties are demanding that AI companies pay for the use of their content or at least provide a clear opt-out mechanism. Legal battles are also advancing, with lawsuits over the boundaries of “fair use” continuing to emerge, leaving AI companies' data collection strategies in a state of high uncertainty.

Against this backdrop, Meta's revision of its terms looks more like careful risk management: delete sensitive words to reduce public friction while preserving room to scrape in order to maintain model competitiveness. This strategy of “soft words, hard actions” may ease some pressure in the short term, but it is unlikely to fundamentally repair the trust relationship between the platform and the content ecosystem.

Editor's Note: The End of the Wording Game Is a Trust Deficit

From a public-relations perspective, deleting the word “training” costs very little and may have an immediate effect—media headlines will change, users' immediate impressions may soften, and regulators' attention may be diverted. But from a governance perspective, this adjustment evades the real question: in what manner and under what authorization should AI companies use other people's content?

robots.txt is not a universal solution. It cannot resolve the retrospective problem of data already scraped, nor can it cover ordinary users who lack the technical ability to configure the protocol. But it at least represents a basic attitude: acknowledging that content producers have the right to express their wishes. Refusing to write it into the terms amounts to abandoning that expression of attitude.

For Meta, the real test is not what words appear in its terms, but whether its crawlers will stop when a website clearly says “no.” In the contest between AI and the content ecosystem, wording can be revised and technology can be iterated, but once trust is overdrawn, the cost of rebuilding it will be far higher than any single model training run.

This article was compiled from Ars Technica.