The Specific Problem

Reports this week confirm that Uber Eats, Deliveroo, and Just Eat are being systematically flooded with AI-generated food photographs. Restaurants are uploading images of dishes that bear no resemblance to what actually arrives at the door. The platforms have content policies prohibiting deceptive listings, yet the practice is widespread and, by most accounts, accelerating. This is not a story about bad actors gaming a loophole. It is a story about what happens when one side of a two-sided market develops algorithmic capability faster than the platform can develop detection capacity, and faster than consumers can develop interpretive defenses.

Why This Is Not Just a Consumer Protection Story

The temptation is to frame this as a fraud problem with a regulatory solution. That framing misses the more structurally interesting dynamic. Delivery platforms are algorithmically mediated environments where listing visibility, conversion rates, and restaurant rankings are all downstream of how well a listing performs against platform signals. High-quality food photography has always correlated with click-through rates. What AI-generated imagery does is decouple visual quality from production cost entirely. Any restaurant, regardless of actual food quality, can now present a listing that is visually indistinguishable from a professionally photographed dish. The algorithmic signal - attractive image predicts consumer interest - remains intact, but the informational content of that signal has collapsed.

This is a version of what Kellogg, Valentine, and Christin (2020) describe when they note that algorithmic systems reorganize the incentive structures workers face in ways that are not always visible to platform designers at the time of deployment. The platform's ranking algorithm was not designed to anticipate that the input it was optimizing on - image quality as a proxy for food quality - would become cheap to fabricate at scale.

The Awareness-Capability Gap, Now on the Consumer Side

My dissertation research focuses on the awareness-capability gap as it applies to platform workers: people know algorithms exist but cannot translate that awareness into improved outcomes. The food photography case extends this gap to consumers. There is now substantial public discourse about AI-generated imagery. Consumers are, in the aggregate, aware that what they see online may be artificially generated. But awareness of the phenomenon does not produce reliable detection skill. Gagrain, Naab, and Grub (2024) document a similar pattern in media consumption contexts, where algorithm literacy - knowing that feeds are curated - does not predict whether users can accurately identify why specific content appears to them.

The consumer facing an Uber Eats listing has no reliable schema for distinguishing a photograph of actual food from a generated image. The visual signals that previously indicated quality - lighting, texture, composition - are now producible synthetically at zero marginal cost. The folk theory that "better photos mean better food" has not been updated to accommodate this new production possibility, and there is no obvious feedback mechanism that would force that update. The consumer only receives negative feedback after purchase, and even then, the attribution is often diffuse.

What the Platforms Actually Control, and What They Do Not

Hancock, Naaman, and Levy (2020) draw attention to the ways AI-mediated communication shifts the locus of message production without shifting perceived accountability. When a restaurant uploads an AI-generated image, the platform's interface presents that image under the restaurant's name, preserving the appearance of direct representation. The consumer has no interface cue indicating that the image was generated rather than photographed. The communication feels direct; it is not.

Platforms are in a structurally weak position here. Detection of AI-generated imagery is an active computer vision research problem, and the generative models producing these images are updated faster than detection benchmarks. Sundar (2020) notes that as machine agency in communication increases, human capacity to assign source credibility becomes destabilized. The delivery platforms are experiencing a concrete version of this destabilization: the credibility infrastructure they built on top of user-submitted photography is being hollowed out by the same generative tools that are reshaping every other image-dependent context.

The Structural Point

What this case illustrates is that algorithmic platforms do not just coordinate transactions - they coordinate epistemic expectations. Consumers, restaurants, and platform operators all operate on shared assumptions about what signals mean. When one party can cheaply manipulate a signal that others cannot cheaply verify, the coordination mechanism degrades. Rahman (2021) describes how platform constraints create invisible cages for workers. The food photography case shows that the cage can face inward toward consumers just as readily. The platform's architecture, not any individual actor's intent, is what determines which manipulations become structurally available and which remain costly. That is the level at which these problems need to be analyzed.