Bookshop.org found out about Meta’s new AI liability language the hard way. A generative AI illustration showed up inside one of its human-made ads, and the bookseller told Adweek it had never asked for it: an account update had opted it into new ad features. It has since switched off every Meta AI ad feature it can find.
That edit sits squarely inside a clause in Meta’s updated commercial terms. Under those terms, the advertiser carries the risk for whatever the AI produced. The terms were supposed to take effect October 30. On October 23 a Meta spokesperson told Reuters the date was moving, and the preview of the terms on Meta’s own site now reads “Effective Date: November 20, 2026.”
The date moved. The AI Features language did not. So any business running ads, a WhatsApp storefront, or an AI agent on Meta’s platforms now has roughly three and a half weeks to work out what it is agreeing to and to cut its exposure before continued use counts as acceptance. Below is the audit I would run, in order.
What the clause actually says
Three passages in the terms preview matter here, and each is worth reading in Meta’s own words.
The first is the disclaimer: “AI features may make mistakes and may not always produce accurate, complete, or suitable results.” Every vendor says something like this. The next sentence is the unusual part: “You are solely responsible for and assume any and all risks associated with such AI features,” and the AI-generated content that results.
The second is the indemnity. If someone brings a claim against Meta that relates to your content or your use of its products, you “agree to indemnify and hold us harmless,” which covers Meta’s damages, losses and legal expenses to the extent the law allows.
The third is the cap on what Meta owes you. Its aggregate liability “will not exceed the greater of one hundred dollars ($100) or the amount you have paid us in the past twelve months.”
Put those together and the risk runs in one direction. Meta’s AI can be wrong, you own the result, you cover Meta if a third party sues over it, and Meta’s exposure to you tops out at your trailing twelve months of spend.
What changed from the old terms is a single condition. Advertisers were previously liable for AI output when they chose to use Meta’s creative AI tools, Adweek reports, and the new AI Features section drops that choice. Meta has said advertisers can turn off individual Advantage+ features in Ads Manager at any time. It has not said directly whether you are liable for a feature you never enabled. Frank Dickson of Dickson Research told Computerworld the clause “turns accountability upside down.”
Why the other vendors’ indemnities don’t solve this either
Most of the coverage frames this as Meta against the field: OpenAI and Anthropic indemnify their customers, and Meta makes its customers indemnify it. That is half right, and the missing half matters for your audit.
Look at what the friendlier indemnities actually cover. In Anthropic’s commercial terms, Section K.1 says Anthropic will “defend Customer and its personnel, successors, and assigns from and against any Customer Claim,” but a Customer Claim is a third-party allegation that your “paid use of the Services” or the related outputs infringes someone’s intellectual property. Section L.3.b then carves indemnity out of the liability cap entirely. That is a strong promise, and it is an IP promise. OpenAI’s Copyright Shield, announced in November 2023, is the same kind of promise, covering ChatGPT Enterprise and the developer platform but not free ChatGPT.
Microsoft’s Customer Copyright Commitment comes with conditions attached: for Azure OpenAI, the “Customer also must have implemented all mitigations required by the Azure OpenAI documentation in the offering that delivered the Output Content that is the subject of the claim.” That means a copyright metaprompt, protected-material and jailbreak filters switched on for text and code use cases, and a retained red-team report you hand to Microsoft when you file a claim.
None of these covers the failure mode Meta’s clause is really about: an AI rewrite that makes your ad say something false, a translation that changes a price, an AI-generated product image that shows a feature you don’t sell. That is a consumer-protection or false-advertising exposure, and no frontier vendor terms I have read indemnify it. Meta’s terms are harsher than the others in two specific ways (the indemnity runs toward the vendor, and the features can be on by default), but the underlying lesson holds across every vendor. Output accuracy is your problem everywhere, and the controls have to be yours.
The audit to finish before November 20
1. Inventory every AI enhancement on every live ad
Open each active campaign in Ads Manager and record which Advantage+ creative enhancements are on: generated images, background swaps, text variations, translations, added sitelinks, carousels built from your product photos. The agency Exverus told Adweek it found Meta adding sitelinks, translating copy and building AI images on ads where nobody had enabled them. Ryan Adams of Exverus put the operating problem plainly: “If every one of those isn’t manually unchecked, the enhancements still apply.”
Turn off anything you did not choose on purpose. Then put a recurring check on the calendar, because Bookshop.org’s experience shows an account update can switch features back on.
2. Keep evidence of your settings, not just your creative
If a claim lands, the question will be what ran and who chose it. Export or screenshot the enhancement settings for each ad set when it launches and after any account update, and save the previews of what actually rendered. A file showing a feature was off on the date in question is the closest thing you will have to a defense under a clause that assigns you the risk regardless. It is the same habit we argued for with agent tracing as an attestation record: log what the system did on a plane you control, not on the vendor’s dashboard.
3. Narrow what any Meta-hosted agent can promise
The terms preview lists the Meta Business Agents and Platform Terms of Service among its supplemental terms, so an AI agent answering your customers on WhatsApp or Messenger falls under the same posture. Whatever that agent tells a customer about a refund, a price or a delivery date is, under these terms, your statement. Restrict it to answering from content you supply, route pricing, refunds and anything contractual to a human, and review transcripts weekly for the first month. Treat it the way we suggested treating inbound shopping agents in the agentic commerce piece: narrowest scope first, wider scope only after it earns it.
4. Price the asymmetry
Do the arithmetic for your own account. If you spent $40,000 on Meta in the past year, that is roughly the most Meta owes you for anything arising from its products. Your indemnity obligation to Meta has no matching number written next to it. Write down the worst plausible AI error in your category (a regulated health claim, a wrong price in a translated ad, a generated image of a trademark you don’t own) and estimate what it would cost to defend. That figure, not the ad budget, is the size of the decision.
5. Ask your broker the specific question
Do not ask “are we covered for AI?” Ask whether your media liability or E&O policy responds to a false-advertising or IP claim over ad content a platform’s AI generated or altered, and whether a contractual indemnity you gave Meta is covered or excluded. As we covered in the insurance piece earlier this month, carriers spent 2026 filing AI exclusions on exactly these lines, and contractual liability you assumed is a classic exclusion on its own.
6. Decide, in writing, whether the channel is still worth it
There is no opt-out clause. Shumaker’s Brian Focht, writing in the National Law Review, reads the terms as leaving businesses that disagree one alternative: stop using Meta’s products for business purposes. Most advertisers won’t do that in Q4, and David Dweck of Go Fish Digital told Adweek Meta “knows that they’ve got advertisers a bit over a barrel here.” Fine. Then make the decision deliberate. Record who approved continued use, which features stay on, and who reviews AI output before it ships. If your company has a risk register, this belongs on it as a vendor terms change with a named owner.
What the delay is, and what it isn’t
Meta told Reuters the extra three weeks are for explaining the changes to partners. The trigger was complaints from publishers and advertisers to competition authorities in the EU and UK, mostly about content rights and crawling, and the EU competition watchdog welcomed the postponement and said talks are ongoing. Senator Marsha Blackburn also warned the update could harm US businesses.
Pressure like that can change wording, and the crawling clause is the likeliest candidate. But the AI Features language in the current preview matches what Computerworld first quoted on October 19. Plan for it to take effect as written. If Meta softens it before November 20, the audit above still leaves you with a cleaner account, a settings log and an agent that can’t promise refunds on its own. That is worth having on any platform. The same pattern of operator liability is showing up in legislation, and Meta’s terms are simply the first place many marketing teams will meet it in a contract they already clicked through.