Meta makes advertisers solely responsible for its AI edits: the settings to review before 30 October

On 30 October 2026, Meta's new Commercial Terms come into force for every business that uses Facebook, Instagram, WhatsApp or Ads Manager for professional purposes. Section 5, titled AI Features, has the business acknowledge that Meta's AI features may make mistakes, declares it solely responsible for those features and has it assume any and all associated risks, with no distinction between the features it switched on and those Meta switches on for it. The same text authorises Meta to crawl the websites associated with the business's products, and continuing to use its tools after that date counts as acceptance. On 22 October, Adweek reported that agencies had pulled live ads because Meta's AI had added links, translations or images to them without being asked. Here is what the text says, what it changes for a European advertiser, and the checks to run this week.

The facts

• The preview of the terms published by Meta states an effective date of 30 October 2026. Section 5 says Meta may use or make available features powered by artificial intelligence, machine learning or generative AI to help the business use its products or to optimise, improve or create content; that these features may make mistakes and may not always produce accurate, complete or suitable results; that the business is solely responsible and assumes any and all risks; and that it must not alter or remove any watermark or content-authenticating metadata generated by AI. Section 7 states that continuing to use Meta's products for business purposes after notice of an update means agreeing to be bound by it.
• According to Adweek (22 October), the generative AI ad terms in force since May 2024 made the advertiser responsible only for the tools it chose to use at its sole discretion; the new section removes that condition. The agency Exverus pulled ads in which Meta's AI had added sitelinks, translated copy and generated images or carousels from the brand's photos, sometimes on ads where the feature had not been switched on; its head, Ryan Adams, describes options tucked away in hidden or collapsed menus. AdCellerant turns off Meta's AI automation across all its campaigns; the bookseller Bookshop.org found an AI-generated illustration in an ad made by its own team. A Meta spokesperson replied that each Advantage+ feature can be turned off individually in Ads Manager and that the update consolidates existing terms; they did not say whether the advertiser is responsible for features it never enabled (PPC Land covered the report the same day).
• The website crawling clause was revised during the notice period: the version read by lawyers on 15 October allowed Meta to use the collected content to train its products, and applies to any business using Meta's products, including one that only runs a Page without ads (Search Engine Journal); the word train was removed around 17 October and the purpose narrowed to advertising and business products, a spokesperson stating that Meta respects preferences expressed in robots.txt (Search Engine Journal, Engadget). The authorisation to crawl, index and copy the pages associated with the business's products, with test credentials, remains, and the text does not mention robots.txt. Separately, Meta notified advertisers that from 28 October product browsing inside the app is added to existing ads unless they opt out, citing an internal test run from 6 to 21 August 2026 across about 117,700 advertisers that reportedly delivered 3% more conversions (PPC Land, Relevant Audience, MediaPost).

The contract shifts the risk to the advertiser, including for features it did not choose

Until now, an advertiser who did not switch on Meta's generative AI tools stayed outside their specific terms. The new Section 5 applies to any AI feature that Meta may use or make available, and Meta's own help page states that some Advantage+ enhancements may be turned on by default. The result is a contract in which the platform chooses what applies and the business bears the consequences: an approximate translation, an added link to an outdated page or a generated visual that no longer looks like the product fall under its responsibility, including with regard to Meta's advertising policies and the law applicable to its messages. Meta has not published the list of features concerned; the About Advantage+ creative help page lists around thirty, including translations, image and background generation, text improvements and, recently, product browsing. The table below sets out what the text says and where.

The reported edits affect translation, links and images, and the AI Act adds a layer in Europe

The three cases described to Adweek are the ones we watch most closely in the accounts we manage. An automatic translation changes the register, distorts a commercial promise or alters a legal notice; in Luxembourg, where a single campaign speaks French, German, English, Luxembourgish or Portuguese to residents and cross-border workers, the copy validated by the client, language by language, is what must appear, as we did for the AXA Luxembourg campaign in 20 languages. A sitelink added by the platform can send people to a page that no longer exists or that contradicts the offer. An image generated from the brand's photos can show a product that does not exist in that colour or a place that is not the company's. In finance, insurance, real estate or health, sectors where Meta itself says it restricts access to some generative features, an automatic rewording can be enough to make a message non-compliant.

Europe adds a layer. The transparency obligations of Article 50 of the EU AI Act have applied to deployers since 2 August 2026, as we wrote in June, and the Commission's draft guidelines confirm that they cover advertising: AI-generated or manipulated content that resembles real people, objects, places or events and could be taken as authentic must be disclosed, visibly (analysis by Gala Law, 29 June 2026). The practical guide of the French ARPP dated 3 August 2026 states that translation, optimisation and technical retouching generally do not trigger labelling, while a realistic visual that is generated or deeply altered can; it also recalls that deployer status depends on who decides on the use of AI and on the dissemination of its output. The grace period until 2 December 2026 only concerns the machine-readable marking owed by providers; the disclosure owed by deployers has applied since 2 August (Baker McKenzie, 2 July 2026). When the platform activates and runs the generation, the question of who is the deployer is not settled in the sources we read: an advertiser may be held responsible by Meta for a visual it did not create, and questioned by a regulator about the labelling of that same visual. The ban on removing AI watermarks points in the same direction as this transparency. The above describes a contract and a regulation and is not legal advice: interpretation is a matter for counsel.

Two other settings switch on without the advertiser: website crawling and product browsing on 28 October

The crawling clause targets the company's website, and Meta's crawler documentation distinguishes several agents: meta-externalagent crawls the web for use cases such as training foundation AI models, meta-externalads supports advertising products, meta-webindexer serves Meta AI results, and two other agents (facebookexternalhit, meta-externalfetcher) may bypass robots.txt for integrity checks or user requests. A site can therefore block the training agent while letting through the one that cites it in Meta AI; Meta says a robots.txt change can take up to 24 hours to apply. It is a trade-off between protecting content and being visible in answer engines, and it belongs to the same work as our AI visibility service. Since the terms do not mention robots.txt, the effective protection rests on Meta's public word.

Product browsing on 28 October follows another logic: after tapping the ad, the user first sees a page inside Facebook or Instagram with product information and personalised suggestions before reaching the website. Meta announces 3% more conversions on a two-week test, without publishing the metric measured, the cost per conversion or the comparison method. For an online retailer, the question is the traffic that no longer lands directly on the site: first-party data, average basket, remarketing, audience measurement. Opting out is done through the button in the notification before 28 October or, at any time, ad by ad in the Advantage+ creative and Destination settings; Meta has not said whether ads created after that date will be affected, nor what the button covers for an agency running several accounts.

Six checks to run before 30 October

The checks below are our own practice; Meta publishes none. The aim is to decide knowingly and keep a record of the decision, without switching everything off by reflex.

• Product browsing. Before 28 October, decide account by account whether in-app product browsing is wanted; decline it through the notification or, later, in the Advantage+ creative and Destination settings of each ad, then check after the 28th that new ads do not switch it on by themselves.
• Advantage+ enhancements. For each live ad: Ads tab, tick the ad, Edit, Advantage+ creative section (the number of active enhancements is shown), Edit, then Expand in the preview to open the advanced preview and its Advantage+ creative tab, because Meta's help page states that some enhancements can only be turned off there. Bulk editing exists for ads of the same format. In Advertising settings, Creating ads, Creative features, untick Test new creative features when the box is offered: it only affects features in testing.
• Risk-based policy. Cropping, format adaptation, brightness: can stay on after testing. Translations, text improvements, generated images and backgrounds, sitelinks, summaries, music: off by default, on only after validation. Regulated sectors and messages with legal notices: everything off, human validation of every version.
• Ongoing control. Weekly review of active ads in the advanced preview until the end of the fourth quarter, dated screenshots of what is actually served, a fresh check after every edit, duplication or format change, and one named person for this validation on the advertiser's or the agency's side.
• Website crawling. Decide which content the company agrees to provide to Meta's crawlers; block meta-externalagent in robots.txt if training is refused, while letting meta-webindexer through if you want to be cited by Meta AI; put sensitive content behind real authentication; list the domains associated with the company's Pages, ads and catalogues.
• Governance. Archive the version of the terms in force on 30 October, update the contracts between agency and advertiser (who validates, who answers for an undetected change, which alert path), bring the AI Act labelling question into the creative validation flow, submit the interpretation to a lawyer and train teams to recognise AI enhancements in Ads Manager.

Meta claims performance gains for its AI features, and some deliver them. The contract of 30 October formalises what we already apply: automating a campaign does not transfer responsibility for what it serves. The decision and the risk stay with the business, and Meta now writes it into its contract.

AIxH's view

In our Social Ads team, every Advantage+ enhancement is decided ad by ad before publication, multilingual copy is the copy validated by the client, and the advanced preview is part of the weekly account review; the accounts we manage will be reviewed before 28 October for product browsing and before the 30th for enhancements. A few days ago we wrote that Meta's AI assistant would soon change campaigns on its own, with approvals still to be defined; the terms of 30 October answer the liability question before the feature arrives. If your Meta campaigns run with active Advantage+ enhancements, reviewing these settings is worth doing this week, and writing down.