Meta Removes Language Permitting News Site Scraping

Meta removes AI training data clause from terms

Meta has stripped a contentious passage from its commercial terms of service after critics warned the wording could let the company harvest material from news outlets and other websites to train its artificial intelligence models. The revision, published on October 17, 2026, follows a wave of pushback from publishers, legal scholars, and developers who argued the original language amounted to a blanket claim on any content reachable by Meta's crawlers.

The deleted clause had stated that Meta might use material from any domain, web page, or URL to provide, develop, improve, research, train, and maintain its products and services under the commercial terms. Observers including the Neiman Journalism Lab and James Grimmelmann, a professor of digital and information law at Cornell University, read that provision as an attempt to override the standard robots.txt signals that site operators use to block automated scraping. In their view, the language would have given Meta a free license to ingest news articles, blog posts, and other copyrighted works for model training without permission or compensation.

Meta's communications team initially rejected that interpretation. Vice president of communications Andy Stone and spokesperson Daniel Roberts insisted there had been no change to how the company operates its web crawlers and called any suggestion to the contrary false. Within hours, however, the company went further by removing the disputed paragraph entirely. In a follow-up statement, Roberts said the consolidated crawling terms had only ever applied to Meta's advertising and business products. He emphasized that the update does not alter crawler behavior, that content owners can still use robots.txt to express their preferences, and that the changes confer no new rights that supersede those signals.

Implications for AI developers training data

For people who build with or deploy AI, the episode underscores how unsettled the rules around training data remain. Major model developers are still negotiating, sometimes in public, sometimes in court, over what constitutes fair use, what licenses are required, and whether technical mechanisms like robots.txt carry legal weight. Meta's quick reversal suggests that even the largest players feel pressure to clarify their intentions when the community perceives a land grab. It also highlights that terms-of-service language can shift rapidly, sometimes faster than the technical controls that enforce them.

Developers who rely on web-sourced data should treat this as a reminder to audit their own pipelines. If you operate crawlers or license datasets derived from public sites, verify that your terms of use and robots.txt handling align with your public statements. Publishers and site administrators should continue to maintain explicit robots.txt directives and, where feasible, implement server-side access controls or licensing frameworks that make their intentions unambiguous. Researchers tracking data provenance will want to snapshot the commercial terms page before and after the edit to document how the policy evolved.

Future of AI training data social contract

The broader takeaway is that the social contract around AI training data is still being written. Companies may test expansive claims in legal fine print, only to retreat when the backlash threatens trust or regulatory scrutiny. Builders who stay alert to these shifts, and who design their systems to respect explicit opt-out signals, will be better positioned when clearer standards eventually emerge.

Source: The Verge