Regulation

Meta’s New Business Terms Could Allow AI Training on Publisher Sites

Meta is preparing to change its Commercial Terms on 30 October 2026, with wording that appears to authorise the company to scrape websites associated with businesses using Facebook, Instagram, Messenger and WhatsApp. Meta could use the collected material to develop and train its products, including AI services. Publishers and independent creators may therefore be granting permissions over content hosted outside Meta’s platforms simply by continuing to use their business accounts.

Nieman Lab examined the wording with three legal experts. The key distinction is between claimed permission and observed behaviour: there is no evidence yet of a new crawling programme, affected training datasets or a court ruling that the clause is enforceable.

What Meta’s terms will permit from 30 October?

The updated Meta Commercial Terms preview authorises Meta and parties acting on its behalf to use automated tools to access, index, cache, analyse and collect information from connected web addresses. The clause covers domains and URLs “associated with your products or services“, as well as addresses made available through use of Meta’s products. It also refers to creating “test credentials” and periodically returning for updated information.

The permitted purposes explicitly include training Meta products, including its AI services. This appears to extend beyond photos or captions uploaded to Facebook, potentially reaching material hosted on a publisher’s own website.

This is not a blanket statement that Meta owns everything on the public web. The immediate contract concerns people and organisations using Meta products commercially. What counts as an associated URL, and whether the account holder actually controls all the content found there, remain unresolved.

Which websites could be covered?

A publisher using a Facebook Page to promote its articles is an obvious candidate. So are a retailer linking to a shop and an artist directing Instagram followers to a portfolio. Business Pages, advertising accounts and commercial tools fall within the reported scope.

The harder cases involve external stores, syndicated articles or third-party pages linked to a business’s products. The wording does not clearly define the boundary between a website the account holder owns and material merely associated with its services. A business cannot automatically license a photographer’s image, a wire-service article or a customer’s proprietary file merely because that material appears on a linked page. Those rights depend on separate contracts and copyright law.

One social account may promote years of articles with different authors, photography agreements and syndication restrictions. A publisher cannot safely treat that archive as a single licence.

Meta disputes that its crawling behaviour is changing

Meta told Nieman Lab that the updated wording consolidates permissions already contained in other agreements, and said its crawling of publishers would not change. The company also maintained that site owners can still manage crawler access with robots.txt. The publisher that first identified the provision, The Steampunk Explorer, reported that Meta had not supplied a matching example of the new broad crawling language in the previous terms.

Two questions remain separate: can a site technically stop a bot, and can Meta claim contractual permission despite that block? Robots.txt primarily addresses access behaviour. Whether this agreement overrides other restrictions remains legally uncertain.

There is no separate opt-out while continuing Meta’s commercial services. It instructs businesses that disagree to stop using those services, leaving dependent publishers with a difficult distribution choice.

How to detect crawling without confusing it with AI training

Publishers should preserve origin-server access logs and CDN or firewall event logs before 30 October, then compare activity afterwards. Search for the documented Meta crawler identities, but do not combine them into a single total:

  • Meta-ExternalAgent: the relevant identity to investigate for AI training and product indexing.
  • Meta-WebIndexer: associated with Meta AI search and the possibility of content being discovered or cited in its responses.
  • facebookexternalhit: retrieves information used to generate link previews when URLs are shared on Meta services.
  • Meta-ExternalFetcher: fetches URLs in response to user-driven requests, unlike systematic training-data collection.

Record URLs, timestamps, user agents, client addresses, response codes and transferred bytes. User agents can be spoofed, so validate suspicious traffic. With Cloudflare, check edge security events too: blocked requests may never reach origin logs.

A successful request from Meta-ExternalAgent would demonstrate retrieval, not that a specific article entered a training dataset. We have not independently tested Meta crawling against publisher server logs for this report. Meta has also not confirmed whether the reference to test credentials would be used to access registration-only or paywalled material.

Why blocking every Meta crawler may be the wrong response

It is technically possible to set different robots.txt rules for different crawler identities. A publisher may want to restrict AI training while continuing to let Meta generate Facebook link previews and surface articles in Meta AI search. Indiscriminately blocking every Meta-related bot can remove useful discovery and referral opportunities without answering the contractual question.

Test bot restrictions against shared-link previews and discovery before deployment. Record changes and remember that robots.txt is a request to cooperating crawlers, not a deletion instruction for previously collected data.

What the change means for AI buyers and publishers

AI procurement teams should distinguish model performance from training-data provenance. A system can produce excellent results while leaving unresolved how it obtained source text, photography or other material. Our AI model comparison addresses model selection; licence scope, source traceability and permitted reuse should be additional checks before a commercial deployment.

A private or self-hosted AI model can protect internal prompt and document workflows, but cannot prevent scraping of material already published on the open web.

Before 30 October, save the terms, inventory domains connected to Meta business accounts, review third-party licences and decide which crawlers should retain access. Seek legal advice where archives, syndication agreements or paywalls are involved.

DIY AI assessment: The wording is substantial enough to warrant a rights and crawler audit now. It is not proof that Meta has begun scraping additional sites for model training, nor proof that it can lawfully use every work on an associated website. The unanswered questions are the exact meaning of “associated”, the interaction with technical blocks and how any claimed licence would operate where the account holder does not own the underlying rights.

Reporting status, 16 October 2026: The previewed clause and 30 October date are corroborated by Nieman Lab’s reporting and Meta’s response. During our check, direct access to Meta’s legal preview required a Facebook login. We did not conduct an independent live-crawl test or publisher-specific bot-log analysis.

Written by Steven Jones

AI Tools Reviewer and Technical Analyst

Steven Jones is a technology analyst specialising in artificial intelligence, machine learning workflows, and emerging automation tools.

At DIY AI, he focuses on clear, practical guidance for people comparing AI tools in the real world. His work covers text generation, image generation, video tools, data platforms, developer-focused AI products, and the automation workflows that connect them.

Steven's reviews are built around hands-on testing, practical benchmarks, and transparent scoring rather than vendor claims. He looks closely at where each tool performs well, where it falls short, and what those trade-offs mean for creators, teams, and businesses trying to make sensible AI adoption decisions.

He has a particular interest in safety, reliability, output quality, performance metrics, and dataset quality. When he is not reviewing the latest AI model updates, he experiments with prompt engineering techniques and contributes to DIY AI ongoing work on fair, explainable scoring frameworks for AI tools.

Back to AI News