AI menus draw complaints as Instagram's AI tags misfire again

Instagram's visible AI labels are supposed to help users spot synthetic content at a glance, but they have misfired for weeks, according to The Verge. Users reported that Meta automatically labeled original photos as “AI Content” after using assistive tools like Canva's Background Remover or after blemish-fixing retouches. One Threads user said the label appeared “every time there is bg remover involved”; another said removing “a speckle” with Canva caused the entire image to be flagged.

The incident echoes a similar controversy from 2024, when Instagram labeled images whose Adobe metadata indicated any use of generative AI, even for negligible edits. Meta promised to adjust its approach and says it relies on industry-standard indicators such as IPTC and C2PA metadata, but there is no recent public information about what those indicators are or how and when Meta scans for them. Meta did not respond to The Verge's request for clarification.

Canva told content strategist Jess Bruno that some of its assistive AI tools were being tagged as generative and that the issue had been corrected; its help page says using Background Remover does not add AI-generated-content metadata. However, Threads users still reported the label appearing after using the tool, and The Verge's own testing found that only images edited or fully generated with Meta's AI app were tagged. About Face, the cosmetics company founded by singer Halsey, also had recent photos automatically labeled “AI Content”; the brand's social media manager denied using AI, saying the photos were taken on an iPhone and lightly edited in the Photos app.

Separately, generative AI has been producing strange food imagery on restaurant menus and brand promotions. TechCrunch reported that AI-generated menu illustrations often look eerily flawless, precisely symmetrical and oddly smooth, and The Verge catalogued examples including donut shrimp, a burger-like mass that looks more like stone, and a burrito full of unsettling holes.

Researchers who spoke with The Verge explained why diffusion-based image generators make such mistakes. Chris Russell, a professor of AI at the University of Oxford, said the model starts with pure noise and removes it step by step, so coarse structures are decided first and fine textures are added last. If the basic structure is wrong, the model paints vivid details onto a flawed form, like adding a sixth finger to a hand. Giovanbattista Califano, a behavioral scientist at the University of Naples Federico II, said diffusion models are notoriously weak at producing “thin, continuous, terminating structures,” such as noodles, strands and tendrils; repeated textures like bubbles, seeds and holes are similarly hard to contain, which helps explain why AI food is so often full of noodle-like artifacts and clustered holes.

AI models also lack knowledge of the physical world. Roland Meyer of the University of Zurich told The Verge that AI image generation reproduces looks without proper knowledge about the world. Michael Cook of King's College London added that the model has no reason to avoid making food look like non-food materials; the result is ice cream that resembles cracked concrete or burgers that appear carved from stone.

The sameness of many AI menus is linked to training data, TechCrunch reported. Reality Defender CTO Alex Lisle said the outputs often look like “a Chili's menu from 2015” because the models learned from that style of commercial photography. He distinguished the current menus from “model collapse,” saying they show convergence, a less extreme degradation that makes images smoother and more generic. Lee Rainie of Elon University told TechCrunch that AI systems are optimized for pleasingness and end up shaving off the edges of images and language, creating homogenized menus. In an experiment on X, a user named Labtec used ChatGPT to make a menu and then edited it 100 times; the food in the images became progressively less recognizable. TechCrunch said it replicated the experiment and saw the same trend.