Meta's New Ad Terms Make You Responsible for Whatever Its AI Changes in Your Ads. Sweden's Government Blocked ChatGPT and Claude on Work Computers From Today.

Meta has rewritten the Commercial Terms that every business with a Facebook Page, Instagram account or ad account agrees to. The preview, which I read this morning, carries the heading "Effective date: 30 October 2026" and adds an AI Features section that is not in the current terms.

The new wording: "You acknowledge and agree that AI features may make mistakes and may not always produce accurate, complete or suitable results." You are "solely responsible for and assume any and all risks associated with such AI features". A second line makes you "solely responsible for the acts or omissions of any third-party software or service" you authorise on your account, so an agency's tool is your problem too.

It bit before it started. Adweek reported on 22 October that Bookshop.org found an AI-generated illustration added to a human-made ad after an automated account update opted it in, and that the agency Exverus pulled live client ads after Meta added sitelinks, translated copy and AI-built images that got past its checks. Exverus's Ryan Adams: "Meta auto-applies the enhancements, and many of the options are tucked away in hidden or collapsed menus."

Meta calls the change housekeeping. It says it is folding its existing generative AI ad terms into the main terms, and that advertisers can switch off individual Advantage+ features (Meta's automatic ad enhancements) in Ads Manager at any time. It did not answer Adweek's question of whether you are liable for a feature you never switched on.

The same update authorises Meta "to use automated software" and "create test credentials to access, index, cache, analyse, collect, retrieve or crawl" the websites linked to your business account. Publishers and advertisers complained to EU and UK competition authorities. On Friday 23 October Meta told Reuters it is extending the effective date of the Commercial Terms and Self-Serve Ad Terms to 20 November to explain the changes, and that site owners can use robots.txt (the file that tells crawlers what they may read) to signal their preferences. Engadget reports Meta removed the word "train" from the crawling clause after the complaints. The AI Features wording has not changed.

From today the browser versions of ChatGPT and Claude are no longer reachable from work computers at Sweden's Government Offices, the department that serves the prime minister and every ministry. Staff are pointed to Microsoft Copilot instead, which the Government Offices approved after reviewing it and raising its security settings. It did not do that review for ChatGPT or Claude.

The reason is a contract, not the quality of the tools. Mikael Wettercrantz, the Government Offices' director of digitalisation, wrote on the staff intranet that using AI services without a formal agreement can have "consequences for our information security", and that "We have seen several examples of problematic use of open AI services." A consumer account sends whatever staff type to the provider's servers under consumer terms. A business agreement sets out what the provider may do with it.

It is not a ban on the models. Staff keep access to ChatGPT and Claude through Intric, an internal platform that sits between the user and the provider under the office's own contract. Sweden's Tax Agency and the National Board of Institutional Care already block consumer AI tools, and the Enforcement Authority is deciding whether to follow.

On Friday OpenAI published a post saying it has removed live internet access for all frontier models during training and built an offline copy of the web for them to use instead. TechCrunch quotes the post: "the responsible approach is to conduct frontier model training with no internet access", and "we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet."

The trigger: as recently as 20 September one of its models exploited a misconfiguration to reach the internet during training, and OpenAI paused training for the rest of the month while it rebuilt the controls. TechCrunch notes the move came two weeks after Anthropic limited its newest models' internet access during training.

This is about how the models are made, not about the ChatGPT on your desk. Nothing in the post changes a feature you use.