The Extra Handlebars Belong to Menlo Park

When a machine changes an advertisement after the brand has signed it off, whose advertisement is it?

The photograph began as a vendor-supplied image of a woman standing beside a bicycle in a park. By the time it ran on Instagram as an REI ad, the bicycle had too many chains, illegible lettering, and a spare set of handlebars growing out of the saddle. REI promoted it for a full week before pulling it on 22 June, then told Fast Company that "Meta auto-enrolled us in an AI personalization tool that produced an inaccurate and inappropriate alteration" of the picture. The retailer's best line was the driest one: "While a two-handled bike might be interesting, it is not something you will find in our assortment."

So the picture that ran was no longer the vendor's, and REI says the change was not its doing. Its account had simply been switched on. The pattern kept turning up. On 22 October, Adweek's Trishla Ostwal reported that an automated account update had moved Bookshop.org into a generative feature, which inserted an AI illustration into a human-made ad. The bookseller said the ad had been "'enhanced' (or ruined, really) by Meta without our knowledge." The agency Exverus said Meta's tools had added sitelinks, translated copy and built images and carousels from clients' own photographs, sometimes on ads where the advertiser had not enabled the feature; it pulled live ads to strip out changes that slipped past its checks. "Meta auto-applies the enhancements, and many of the options are tucked away in hidden or collapsed menus," said Ryan Adams, who handles paid search and social there.

Meta had an answer ready. The revised Commercial Terms it planned to bring into force on 30 October added a section on AI features which, in the language Adweek quoted, "may make mistakes and may not always produce accurate, complete, or suitable results." Advertisers would be "solely responsible" for what those features create or change, including the ones Meta switches on by itself. Under the terms still in force, by Adweek's reading, an advertiser carries that risk only when it has chosen to use the tools. The draft dropped the choosing. So who approved the handlebars? Under the new paragraph, the advertiser did, by being a customer.

That answer did not last the autumn. Meta pushed the start date back to 20 November, and on 5 November it paused the update altogether, with a spokesperson saying the company wanted to "take our time to get this right." Its Business Help Center now says the updates "will not take effect on November 20, 2026," and that the existing terms remain in effect. No new date has been set, so the opening question is open once more.

Factories settled it long ago, in their supplier contracts. In automotive supply, the Production Part Approval Process maintained by the Automotive Industry Action Group exists to govern what AIAG describes as the "product and process change process, providing an approval outlet for all changes." The supplier quality manual Astemo Americas issued in January, Section 2, Issue 27, shows what that means in practice. "All changes (design and process)" send the supplier back through PPAP with a fresh part submission warrant. Approval "must be obtained prior to implementing any changes." It also says that a Product Change Notice on its own "is not enough." Telling the customer does not count.

The reasoning rests on who can see what. A carmaker cannot watch every shift inside a supplier's plant, so whoever runs the plant answers for every change made there. An advertiser cannot look inside Meta's enhancement pipeline either. Exverus found the alterations only once they were live. REI pulled its bicycle after the public had already found it. Supplier quality puts the burden on the party that controls the process and can see the change happen. Meta's draft put it on the party that could not.

What makes the draft hard to defend is that Meta builds the approval gate when it chooses to. Its own write-up of its 6 October announcements at Advertising Week New York describes a business AI assistant whose work lands in Ads Manager as a draft and needs the advertiser's explicit approval before anything is published. The same company also wrote the AI-features clause. One product treats the advertiser as the customer who signs off. The other treats the advertiser as the party that absorbs whatever ships.

Money explains part of the asymmetry. Meta's third-quarter release on 21 October reported $63.061 billion in advertising revenue, up 26 percent, on 10 percent more ad impressions and a 14 percent higher average price per ad, and said its AI investments "are already delivering substantial incremental advertising revenue." The October announcement adds that advertisers who take the assistant's recommendations typically see 5 percent more conversions at the same cost per conversion. That is Meta grading its own homework, with no outside audit, so I would not lean on it. The direction of the incentive needs no audit. When machine-made changes lift results, the gain shows up in Meta's revenue. When one goes wrong, the cost shows up in a brand's reputation.

The strongest case against my position deserves a straight hearing. Remi Daviet of INSEAD and Yohei Nishimura of the Wisconsin School of Business tested generated imagery against a professional designer in a live Instagram campaign for, of all things, an outdoor-activities company. Their paper in Marketing Science, as INFORMS summarised it on 4 August, found a mean click-through rate of 0.98 percent for the AI portfolio against 0.65 percent for the human designer, and eighteen months later, with no retraining, the AI set still led by 3.38 percent to 3.24 percent. If machine-made variants beat human ones on the metric advertisers pay for, the case for letting the machine change ads by default looks strong.

I accept the result and reject the inference. A click measures attention. It says nothing about whether the product in the picture exists, whether a translated line still says what the lawyers cleared, or whether a new sitelink points anywhere the brand would send a customer. A bicycle with a spare set of handlebars could easily earn clicks. And the study kept its own gate: the INFORMS summary says "the designs were vetted for brand standards, and they were field-tested." Better output is a good reason to file a change. Under PPAP it does not waive the approval.

So, plainly: the paused draft had the allocation backwards, and whatever Meta publishes next should put the burden of a machine-made change on the party that makes it. In practice that means enhancements stay off until an advertiser turns them on, every variant is logged and retrievable, and the brand approves a preview before the first impression, as Meta already promises for its assistant. None of this needs new science. Comparing an altered image with the approved original, and stopping any variant whose product has changed shape, is routine computer-vision inspection of the kind factories run on parts every shift. A model that can draw an extra set of handlebars can be asked to count them.

Until then the workaround is manual. AdCellerant told Adweek it turns off Meta's default AI automation across every campaign. Go Fish Digital, whose lawyers reviewed the terms, said it would not tell clients to pull back on Meta in the fourth quarter; its president, David Dweck, summed up the bargaining position: "Meta knows that they've got advertisers a bit over a barrel here." The checking now happens on the agency side of the contract.

Underneath sits a legal question older than any of these tools. Section 230, added to federal communications law in 1996, keeps a platform from being treated as the publisher or speaker of "information provided by another information content provider," and it defines that provider as anyone "responsible, in whole or in part, for the creation or development of information." REI says Meta's tool drew the extra handlebars. Bookshop.org says the illustration arrived without its knowledge. The paused paragraph would have settled their cases by contract, and the terms still in force assign AI output to advertisers who chose the tools, which both companies say they did not. So when an enhancement next writes something worse than an odd bicycle, a discount the retailer never offered or a claim its lawyers never saw, and a regulator asks whose words those were, will a court call them the brand's speech or Meta's?

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.