Prediction: Meta's AI Training Terms Are About to Change. Here Is Why
On October 15, 2024 Meta quietly edited the publicly posted Commercial Terms for Meta Platforms, Inc., deleting a paragraph that said the company could use content from any domain to train its artificial‑intelligence systems. The change follows a week of criticism from news publishers, civil‑rights groups, and European regulators who warned that the clause could legitimize large‑scale content scraping without compensation. By removing the language, Meta signals a shift in its legal stance, but the practical impact on how its AI models are built remains unclear.
What Happened: Meta Removes Scraping Clause from Terms
The clause in question appeared in the "Commercial Terms for Meta Platforms, Inc." PDF that the company posts on its developer portal. It read, in full: "Meta may use any content from such domains, web pages, or other online locations for the purpose of training, improving, or developing its artificial‑intelligence models." The sentence vanished from the document after a public outcry that began on October 9, when The Verge reported that the language could give Meta a legal shield to harvest articles, videos, and images from any site, including those that have not granted permission. nnMeta confirmed the edit in a brief statement posted to its policy blog on October 16, noting that the change was part of an "ongoing review of our terms to better align with industry best practices." The update was made public at 02:13 GMT, and the revised PDF was downloaded by an estimated 2,300 developers within the first hour, according to data from the analytics firm SimilarWeb. The removal did not affect the broader set of Meta’s data‑use policies, which continue to allow the company to collect user‑generated content for ad targeting and product improvement. nnAccount to The Verge and a follow‑up interview with Meta’s policy team confirm that the specific paragraph was the only reference to unrestricted content scraping. No other sections of the commercial terms were altered at the same time.
Why It Matters: Potential Shift in How AI Consumes News
For news organizations, the original clause represented a legal gray area that could have allowed Meta to train large language models on copyrighted articles without paying licensing fees. Many publishers, including The New York Times and the Associated Press, have already negotiated separate deals with tech firms for AI‑training data, but the blanket permission in Meta’s terms threatened to undermine those agreements. By erasing the language, Meta may be attempting to avoid future lawsuits and regulatory scrutiny, especially in the European Union where the Digital Services Act now requires platforms to be transparent about how they use third‑party content for AI.nnThe change also has implications for the public’s access to information. If Meta’s AI systems—such as the LLaMA series—continue to be trained on news content, the company still benefits from the intellectual labor of journalists, but the removal of the clause could limit the scope of that training to content that publishers explicitly share. This could lead to a more fragmented AI ecosystem where only well‑funded newsrooms can afford to license their archives, potentially biasing AI outputs toward topics covered by larger outlets.nnBeyond the newsroom, the edit touches on broader concerns about data sovereignty. Privacy advocates have argued that allowing a single corporation to ingest massive swaths of online text gives it disproportionate influence over public discourse. By tightening the legal language, Meta may be acknowledging that unchecked scraping could trigger antitrust investigations, as seen in recent filings by the U.S. Federal Trade Commission against several big‑tech firms for monopolistic data practices. The removal, therefore, is not merely a cosmetic tweak; it signals a possible rebalancing of power between content creators and platform owners.
“Meta’s policy spokesperson, Maya Patel, told a briefing on October 17 that ‘the removal of the scraping clause reflects our commitment to transparency and respect for the rights of publishers, while still enabling us to improve our AI responsibly.’”
What We Don’t Know Yet
The announcement leaves several critical questions unanswered. First, it is unclear whether Meta will replace the deleted paragraph with a more restrictive provision or simply rely on existing copyright exemptions. The company has not disclosed any new licensing framework for news content, so it is possible that the same de‑facto scraping will continue under a different legal justification. Second, regulators in the EU and the United States have not yet indicated whether they view the amendment as sufficient to address their concerns. The European Commission’s Digital Services Act task force has scheduled a hearing on AI‑training data in early November, but the agenda has not been made public.nnThird, the impact on smaller publishers remains opaque. Large media conglomerates have the resources to negotiate bespoke data‑use agreements, but independent outlets may lack the bargaining power to secure fair compensation. If Meta continues to train models on publicly available articles without explicit permission, those smaller voices could be marginalized in AI‑generated summaries and search results. Finally, the technical details of how Meta’s models ingest and weight news content are proprietary. Without transparency into the data pipeline, observers cannot assess whether the removal of the clause materially changes the composition of training datasets. These gaps mean that the real effect of the policy change may only become
Despite the headline‑grabbing clause, internal audits show that less than 2% of Meta’s LLaMA training data actually comes from news articles.
