The sameness problem behind those unappetizing AI-generated menus

The Uncanny Valley Hits the Dinner Table

A growing number of consumers are reporting encounters with unsettlingly perfect food imagery on restaurant menus, a direct result of businesses adopting generative AI tools. These images—from flawlessly symmetrical bagel sandwiches to unnaturally smooth hamburger buns—are creating an "uncanny valley" effect, eliciting a visceral sense of unease. This phenomenon serves as a public-facing example of the current limitations and aesthetic biases inherent in diffusion models from companies like OpenAI, highlighting a disconnect between technically proficient generation and authentic consumer appeal.

Understanding AI's 'Sameness' Problem

The strange aesthetic of AI-generated food stems from how the models are trained and refined. According to Alex Lisle, CTO of AI-detection firm Reality Defender, the models draw from a vast but often stylistically narrow corpus of data, leading to a homogenized look reminiscent of a "Chili's menu from 2015." This issue is exacerbated by a technical process known as convergence, where the model's outputs become increasingly similar as they are fed back into training data, reinforcing a single, bland style.

• Homogenized Training Data: Models learn from existing online images, which often share a similar, highly-produced commercial style.
• Convergence: Unlike total model collapse, convergence degrades output quality by reinforcing a narrow aesthetic, causing the AI to repeatedly generate stylistically similar images.
• Optimization for 'Pleasingness': AI image generators are often tuned to "shave off the edges," removing imperfections and textures that make food look realistic.
• Iterative Degradation: Repeated edits to an AI-generated image, such as changing prices or text, can cause the images to become progressively smoother and less realistic.

Beyond Bad Menus: Eroding Digital Trust

While off-putting menus may seem like a trivial issue, they point to a much larger challenge: the degradation of trust in visual media. Research from the University of Duisburg-Essen confirms that near-realistic AI food images can elicit more disgust than obviously fake ones. This consumer backlash against AI-generated content signals a risk for brands that adopt the technology without considering the psychological impact. As Lisle notes, the foundational belief that "seeing and hearing has always been believing" is no longer stable, a fundamental shift that impacts everything from advertising to evidence presented in court systems.

The widespread deployment of low-effort generative AI in consumer-facing applications presents a clear brand risk. The perceived cost savings can be quickly overshadowed by customer alienation and a loss of brand authenticity, creating a market opportunity for both higher-quality, specialized models and reliable AI detection services.