Piling On The Wrong Bike

Another week, another online pile-on against AI in advertising. Readers would be forgiven for being as weary of this song and dance as the wider world clearly is. But this week’s iteration is both more nuanced and more insidious – and it betrays just how much anti-AI animus is based on outdated assumptions about generative image capabilities, and how much of it can be pointed at the wrong people.

The inciting incident this time around was an article in Business Insider (dated Wednesday this week) which captured a Reddit post about an Instagram ad from multi-brand, US outdoor retail co-op REI. 

(There’s a whole separate circle of weariness reserved for reporters who have to write about stories that begin life on one social platform and then get tail-eaten by another, but that’s a digression for another time.)

In the offending advertisement, REI’s official Instagram account was promoting a bicycle from one of the brands it stocks, Van Rysel. The image in the ad was clearly, and badly, altered or generated with AI. It carries all the early-stage, small model hallmarks of implausible background geometry, self-lit skin, nonsensical bokeh and so on.

More importantly, it mangled the brand name on the bike frame and gave it a second set of handlebars stuck behind the saddle.

There are several layers of reputational harm at work here. As a multi-brand stockist, REI is promoting another company’s products, so presumably Van Rysel would have been none too pleased had it seen the ad before it reached the public. But REI itself also runs one of the industry’s most rigorous policing regimes for the products it stocks, and for the inclusivity of the marketing assets used to promote them. So the anger had an additional layer of righteous betrayal to it, for customers who shop at the co-op because of its exacting intake credentials. 

(You can take a look at REI’s current Product Impact Standards, which are a precondition of “doing business with REI” to better understand the high bar the retailer sets for the brands it engages with – and how that bar extends to “Marketing Diversity and On-Model Photography” to get a feel for how REI runs one of retail’s most granular imagery-and-marketing-asset vetting regimes as part of the same standards.)

So, as much as the internet loves to point out the obvious, and obviously-terrible, uses of AI in marketing images, how did this happen? The Interline has no horse in this race, and nobody here is an active REI member (the company doesn’t even sell in the UK), but something doesn’t quite add up here.

Why would a company with such high standards for the products it will list and the assets that need to be provided as part of the intake package, look at an ethereal image of a bike with two handlebars and effectively say “it’s fine, ship it”?

The answer should be obvious: they almost certainly didn’t. For REI specifically, AI alteration or generation is a direct breach of published rules that govern “accuracy of product claims,” and would undermine not just the promise made between the retailer and its potential members and shoppers, but the co-op and the companies it stocks.

It seems very likely, on the balance of fact, that nobody at REI or Van Rysel saw this content before it was aired on Instagram. Because the content was generated on-the-fly as part of a programmatic advertising stack, from Meta, that has now extended its tendrils into asset generation. And that stack is set to become a much bigger problem for fashion over time.

What’s more, people’s readiness to fixate on these most egregious instances is muddying what should really be the targets of their ire: the state of the art content they don’t notice, and the platform holders that continue to mine their social graphs and profiles to serve up ‘personalised’ advertising without the parties on either end of the equation consenting.

It’s The Interline’s firm opinion that getting angry about prime slop is missing the wood for the trees. As our upcoming AI Report will set out, our policy is that hating AI is as valid as loving it, but that neither aim is well-served by focusing on the crappiest outcomes and refusing to look at the fast-moving frontier of undetectable synthetic media.

Or, to put it more bluntly: nobody’s cause is being advanced by railing against the one visibly awful AI ad you spotted, and letting the other 90% of visual PDP content that’s going unremarked and, for the time being, unlabelled, pass you by. Unless the explicit awfulness is the objection, in which case you have a problem with AI quality, not AI usage itself.

So with that in mind, how, specifically, does this happen? How do ads make it in front of users without the brand or retailer’s sign-off?

To grapple with that, we have to acknowledge the scale of Meta’s advertising business, and to draw attention to the part it doesn’t already own – in that order.

First up: expert predictions, cited in Reuters, are that Meta will become the world’s biggest seller of digital advertising by the end of this year. Its net revenues from advertising are expected to hit $243.46 billion for the 2026 calendar year, putting it a shade ahead of Google, which has held the podium for basically forever.

In a very literal sense, this means that the way Meta sees the advertising market is the way the ad market ends up being manifested. If you hold the most inventory (even inside a walled garden) and make the most money, you also get to make the rules.

Second: Meta owns, within its own ringfenced platforms, most of what’s traditionally been called the programmatic advertising stack. Until the last year or so, programmatic advertising referred to the selling and automated buying of media inventory through DSPs, SSPs, and auctions that complete in fractions of a second.

The Menlo Park giant has also been very clear that, as well as pairing user intent and target demographics with relevant content from advertisers, it also wants to programmatically create that content on the brands’ behalf. And it intends to use AI to do so.

This is the process that led to the REI bike-with-two-sets-of-handlebars. The image was created by Meta’s AI models and automatically slotted into media inventory that the company’s algorithms decided would deliver results.

And to some extent: it worked. The Reddit post that spurred the pushback was clearly written by someone who understood REI’s brand promise enough to be put off by the idea that AI was contravening it. Or, in other words, the ad did end up in front of a set of eyeballs that might have been attached to a hand that would click the link… if the quality of the ad hadn’t been so poor.

And therein lies the problem. The ad was bad because, we have to assume, Meta is using the cheapest and smallest models possible to fulfil these programmatic briefs. And because the ad is bad, the default workflow purposefully does not expose it to the client in advance.

Part of this will be down to speed: Nano Banana Pro and GPT Image 2 both take minutes to return a generation, while it seems likely that Meta’s programmatic placements need content putting in them in seconds, especially if they’re being personalised to appeal to an individual or a granular demographic. 

The other part will be down to cost and margin: why should Meta, if it owns every layer of this stack, let other model providers like Google and OpenAI take a cut of its programmatic placement-and-content pairing?

The final part is the kind of brute force tactics that only the biggest platform holders can get away with. In this instance, Meta apparently generated ad creative that was never approved by the brand or the retailer. It can’t be coincidence that this creative would never have been approved by either party, because it sucked. 

There’s evidence to that effect: another Business Insider analysis, from October of last year, spotlighted brands whose products were being misrepresented (models swapped from middle-aged men to elderly women, or materials substituted in generated images so they no longer matched the real products) as well as including some damning quotes about the opt-out business practices Meta applies to its programmatic stack.

Speaking to Business Insider at the time, the CEO of an agency that spends around a hundred million USD on Meta ads per year revealed that his company was spending “two to three mornings per week” turning off generative advertising for accounts that have previously opted-out.

And beyond this dark pattern practice, there’s the important caveat: the images just aren’t any good! The Interline does not know what image models Meta is using, but the programmatic content we have seen suggests they are a long way behind the state of the art.

As stated above, this creates and perpetuates the impression that generative AI cannot create good campaign photography, when it very clearly can. There are numerous companies – many of them profiled in that upcoming AI Report 2026 – that are doing exceptional business precisely because they can replace product photography and campaign creation at speed and reduced cost, with visual quality that passes most people by.

But we would not even need to interact with those bespoke solutions to obtain better results than what Meta is producing on its ad clients’ behalf here. Thirty seconds in a generative workspace that gives you access to a variety of models would get you not just better first results, but also reusable workflows and techniques to iterate and create alternates for different channels and purposes.

So much of the pushback to this particular story focuses on the misconceived notion that AI cannot create effective or invisible advertising. This is demonstrably not true. The real story, and the one The Interline believes everyone in fashion should pay attention to, is that the biggest advertising platform holder decided a year ago that nobody on either side of the aisle would care.