febbhav/signs-of-ai-design

This repository is a community field guide that catalogs the visual patterns typical of AI-generated design. It covers websites, app interfaces, images, video, logos, slide decks, social graphics, illustration, and print. The project takes inspiration from a Wikipedia page that documents the tells of machine-written prose, and applies that same approach to machine-made visuals. It is structured as an open list that anyone can contribute to and build upon. The guide is organized into sections by medium and design element, such as web interfaces, images, video, logos, and documents. Each entry explains what to look for, why AI tools produce that specific pattern, and the false positives where human designers might make the same choice. Entries also carry two labels. One label for reliability describes how much a single match tells you, ranging from strong to weak. Another label for status notes whether the tell still applies to current tools, since AI models update and patch their recurring visual habits over time. A core rule of the guide is that no single sign proves AI involvement. Almost every pattern listed was originally invented by human designers. The guide argues that AI tools amplified the frequency and uniformity of these patterns until they became recognizable fingerprints. One match is considered noise. A stack of many matches with no deviation is considered a signature. The text warns against using the list to harass people, noting that identifying AI involvement is not evidence of low effort or deception, and that human detection accuracy is currently poor. The content covers specific details like the prevalence of certain color gradients, the default use of specific typefaces, and recurring layout choices in generated web pages. It explains how defaults from popular coding tools saturated training data and became the standard look of generated sites. The repository also includes a section on signs that no longer work and a discussion of why models converge on similar outputs. The full README is longer than what was shown.