How-to·December 3, 2024, 22:36

Five types of clues can reveal that an image was made by AI

AI-generated from the sources listed, with automated source checks.

Researchers at Northwestern have created a guide to spotting AI images and group the telltale signs into five categories. The authors stress, however, that visible flaws do not always appear.

Images from AI tools such as Midjourney, Stable Diffusion and Firefly can now look so real that ordinary people struggle to tell them apart from real photos. That is the background for a guide from Northwestern researchers, which was posted as a preprint on the research site arXiv on June 12, 2024. A preprint is a research paper that has not yet been reviewed by other researchers.

The authors are Negar Kamali, Karyn Nakamura, Angelos Chatzimparmpas, Jessica Hullman and Matthew Groh. The guide is based on a review of 189 images: 138 made with so-called diffusion models (the type of AI that creates images from text), 9 selected from social media and 42 real photographs.

Five types of signs

The guide divides the possible clues into five categories:

- Anatomical signs: wrong fingers, eyes and bodies, extra limbs, odd bends and too many teeth. - Stylistic signs: the image looks plasticky, glossy or overly cinematic and dramatic. - Functional signs: text that cannot be read, strange clothing and objects that do not work physically. - Physics violations: wrong lighting, shadows or impossible reflections. - Sociocultural signs: scenes that seem implausible or are historically inaccurate.

No foolproof method

The authors acknowledge that the flaws are not always there. That means you cannot reliably determine whether an image is AI-generated just by looking at it. You need context or metadata, that is, information about the image's origin.

The guide dates from 2024, and image models are evolving quickly. Some of the signs may therefore have become less reliable.

More AI literacy

In an interview with Scripps News, Matt Groh of Northwestern says: "Media literacy is super awesome. But it needs to extend to AI literacy." In other words, media and AI skills should go hand in hand, and according to him the guide should be updated every year.

Cole Whitecotton of the National Center for Media Forensics believes everyone should try tools like ChatGPT and Midjourney themselves to understand the technology.

For readers, this means that a critical look at hands, text, lighting and the scene as a whole can raise suspicion. But an image that looks flawless is not thereby proven to be real.