Who is responsible when a platform's AI changes your ad? You are, by contract, at two of the biggest ad platforms this week. At both of them the product gives you nothing to approve. A contract that says you approve and a product that gives you nothing to approve is a transfer of loss, not a control.
Two things land within weeks of each other. Meta's revised Commercial Terms were due to take effect on October 30 and now take effect on November 20, after Meta delayed them on October 23. OpenAI turns on "creative enhancements" for ChatGPT ads by default at the end of October, and the exact date depends on which OpenAI source you read, which I will come back to. The companies are different and so are the products. The plot is the same: the platform's AI may rewrite what you made, and the advertiser is the one who answers for it.
Correction, November 1, 2026: on October 23 Meta moved the effective date of its revised terms from October 30 to November 20. The text below has been updated; the OpenAI dates are unchanged. Reporting from the same week says the free-standing AI-training sentence was removed and the advertiser-responsibility section was reportedly left as it was; I have still not read the terms page directly.
I should start with what the calm practitioner will say, because most of it is right. The features were not invented this month. Meta describes the update as housekeeping that folds its generative AI ad terms into the main terms, and says advertisers can switch off individual Advantage+ features. OpenAI's setting is a toggle in Ads Manager that keeps any opt-out you already made. For tools you chose to use, responsibility for the output was always yours, and reporting I read says at least one large agency is not advising clients to pull back in Q4 over any of it. I have no quarrel with those points.
I have a different one. In July I wrote The Default Owner about Google's terms rewrite, and the claim there was that a default is a policy about absence: whoever sets it ends up owning the decisions you never priced. That essay described one company's clause. Two more companies have now put the same sentence in a different font. It also left a question open that matters more this week than it did in July. What do you do when the clause says you approve and the product has nowhere for approval to happen?
What the two documents say
Meta first, with a caveat I would rather state than hide. I tried to read the Commercial Terms directly, and the page is closed to the tool I used. What follows is trade-press reporting of Adweek's reading of the terms, and I will mark it as reported. A new AI Features section reportedly makes the advertiser solely responsible for what Meta's AI features create or change in ads, warns that those features may make mistakes, and drops an earlier condition that tied responsibility to tools you chose to use. Reportedly, the new section reaches features Meta turns on itself. Acceptance is implied: you keep running ads, you have agreed. Meta told Adweek it is consolidating terms and that advertisers can turn off individual features. It did not say whether the clause covers a feature an advertiser never enabled. If Meta's text says something narrower than this, I will correct the essay and say so.
The behaviour behind the clause is reported more consistently than the clause is. Agencies describe sitelinks, translated copy and generated images showing up in live ads that nobody requested. One agency told Adweek that some of the controls sit in collapsed menus and the edits still apply unless each box is unchecked. An online bookseller said an account update opted it into a feature that added an AI illustration to an ad. Separately, Meta is switching on product browsing in existing ads on October 28 unless advertisers opt out, which opens product details inside Meta's apps rather than on the advertiser's site.
OpenAI I could read myself, and I read the help page on October 24. Creative enhancements adapt your existing ad to the conversation or the user's interests, using the ad and content from your landing page. They can change four things: headlines and descriptions, the language of the ad text, image size, and the destination URL. Variations can be prepared in advance or generated in real time. When the setting is on, the page says, enhanced ads can run without your review or approval, and there is no preview of every variation. Reporting is at campaign, ad group and ad level, with impressions, clicks, spend, click-through rate and attributed conversions. The page adds a sensible caution that a change in those numbers over time does not by itself show what the enhancements did. It requires that your ads and landing pages comply with the ad policies. It does not describe a review step for the enhanced versions, and it does not say how labeling applies to each one.
On the date, the help page says the setting turns on for new and existing campaigns starting "Thursday, October 31, 2026." October 31 is a Saturday. The email reported by PPC Land says October 30, and an earlier read of the same page by the same outlet said October 29. OpenAI has not explained the difference, and I would not pick one. Plan for the earliest.
Approval is a property of the record
A signature is only worth something if the signer could have read the document. That sounds like a legal point, and it is partly an engineering one. The contract says the advertiser approves. The product says the advertiser need not. Those two statements can both be true, because they allocate different things. The contract allocates the loss. The product allocates the work. They have been allocated to different parties, and the party that carries the loss is the party that cannot see what it is carrying.
Look at what the report would have to contain for approval to be possible after the fact. It would have to show which version of the ad a given person saw. The OpenAI page lists the metrics, and none of them is the version. I did not find the equivalent in the Meta material I could read, though I did not read Meta's reporting documentation, so that half is untested today. In The Log Records What Happened, Not What It Chose From I argued that a log which records the action but not the set the agent chose from cannot be audited. This is the same gap one level up. You can see that an ad earned clicks. You cannot see which sentence earned them. The version that ran is exactly the information the platform holds and you do not, and it is the by-product of the platform's own optimisation that I called ad exhaust: data you paid for and don't get to keep.
The four changes on OpenAI's list also do not carry the same risk, and a single switch treats them as if they did. Resizing an image changes the form of an ad. Translating a headline changes its meaning in a language the advertiser may not read. Pointing the click to a different page changes what the click promises. A policy owner would want to decide about those separately. Where the control is one toggle for all four, that is a finding in itself, because the only choices left are everything or nothing.
The same design, built the other way
On October 21 Walmart announced a gateway for shopping agents. According to TechCrunch's report, each agent gets its own profile inside the customer's account, which Walmart is rolling out next month. The customer chooses between approving every purchase and letting the agent spend up to a limit the customer sets. Walmart's agent terms say its returns process applies even when an agent does not act as intended. I make no claim about the rest of the gateway. The design fits in one sentence: a scoped identity, a limit set by the person who pays, and a remedy offered by the party that sells.
Set that beside what an advertiser gets. A contract, a toggle, and in the documents I read no limit and no stated remedy. The ad platforms ask more of the advertiser than Walmart asks of a shopper, and give less to work with. I wrote the neighbouring argument two weeks ago in An Agent Has an Identity. Nobody Has Its Decision., which found vendors naming the agent and not the person who signs. Here the person who signs is named in the contract and has no way to read what they are signing.
What to fill in before Black Friday
The table below holds what I could document on October 24. The last two columns are blank on purpose. Those are yours to write, and the second one is a decision, not a note.
| Default-on AI edit | Approval step | Where the opt-out lives | Report shows the version that ran | Who signs / decision |
|---|---|---|---|---|
| OpenAI: headline and description rewrites, translation, image resize, destination URL | None per version; no preview of every variation | One setting under Manage settings in Ads Manager; account default, campaign override | Not documented | Name / leave on, opt out, narrow |
| Meta: generated images, translated copy, added sitelinks (as reported) | Reported as applied without a request | Per Advantage+ feature in Ads Manager; some reportedly in collapsed menus | Not checked | Name / leave on, opt out, narrow |
| Meta: product browsing in existing ads, from October 28 | On by default unless you opt out | Opt-out notice sent to advertisers, per reporting | Not checked | Name / leave on, opt out, narrow |
Three rules for filling it in. The signer is not the media buyer. It is whoever's budget would carry the cost of an ad claim the company never wrote: the refund, the complaint, the retraction. Second, pick a decision for each row, and treat "narrow" as a real answer where the platform allows it. Where it does not, write down that the control is all or nothing. Third, give each row a kill condition while you are calm, such as a policy rejection, a complaint, or a drop in lead quality. One advertiser has said publicly that she paused her ChatGPT ads test over lead quality, so that last one is not hypothetical.
Then ask the question finance will ask in January: which line in the plan carries the loss from an edit nobody reviewed? If the answer is none, you have what Who Owns the Decision? called an unpriced decision, with a contractual owner whose name is on the account.
The best objection
The best objection comes from any good practitioner, and it is short: just opt out. In many accounts that is the right call and it takes ten minutes. I would do it for any account where I could not name the signer. Two things keep me from stopping there. The opt-outs are per feature and per platform, and some are reportedly hard to find, so "just opt out" is a project, not a click. And the documents I read do not say whether opting out changes who is responsible for edits that already ran. I do not know the answer. I am not a lawyer, nothing here is legal advice, and a question about the contract belongs with counsel who has read the actual text.
On January 31, 2027, open OpenAI's Ads Manager reporting documentation and Meta's Ads Manager breakdown options. My claim: neither platform lets an advertiser see, per ad or per impression, which AI-edited version of an ad ran.
On October 24 I found no such report in the OpenAI page I read, and I did not check Meta's breakdowns, so the Meta half is untested today. If either platform has shipped version-level reporting by the date, I was wrong, and I will say so here.
I do not think either company is hiding anything. A platform that can raise relevance by rewriting an ad at the moment it is served has a reason to do it, and a contract that puts the risk on the advertiser is the ordinary way to make that safe for the platform. The result for a business is a loss with a signature on it and no line item. Someone will find that line item after Black Friday, in a refund report or a complaint, and it will be filed under marketing because the account is. The only choice left to the company is whether it names the person who carries it now or learns the name later.