How to spot AI generated photos while reviewing
The amount of AI generated photos has been slowly increasing over the last couple of months, and they get more and more realistic. I thought I’d share a guide that can help reviewers spot AI generated submissions and/or confirm that they are AI generated for reporting purposes.
Step One: How to spot typical red flags
In my experience, the most commonly faked submissions are murals, little free libraries, trail markers/signs, plaques, and in Germany, Stolpersteine and bee hotels.
I generated some samples for educational purposes.
My red flags for murals:
- Main and supporting photo are the same, one is a crop of the other, or supporting doesn’t include the object at all
- Lighting mismatch: colors are more bright than its environment, the sun doesn’t cast shadows or change the color of the mural/object
- Lack of surface texture: AI will smooth over any of the natural lines or cracks of a wall
My red flags for signs:
I completely generated sign #1, while #2 is a real sign that Gemini added to my photo.
The first is more obviously AI
- doesn’t blend into its environment
- text includes many nonsense words
The second is harder to spot but also includes a weird text fragment. I also always check for the texture of the grass: AI models struggle to process the complexity of grass, so the texture is frequently blurry, individual blades blend together and don’t grow in a natural shape. AI will also make use of repeating patterns to fill large patches of grass.
Step Two: How to confirm your suspicions
Your best friend is SynthID. SynthID is an AI-watermarking technology developed by Google. This watermark works by embedding a specific pattern directly into the image, and is invisible to the human eye. Because it is built into the file itself, the watermark survives even if someone crops, resizes, compresses, or edits the file.
Because Google open-sourced the underlying technology, the list of apps and platforms using SynthID has expanded far beyond Google. SynthID is used by many of the common image generators:
Google Gemini (NanoBanana)
The OpenAI Ecosystem (DALL-E 3/ChatGPT)
NVIDIA Cosmos
Kakao (Kanana AI suite/ KakaoTalk)
SynthID for all four systems can be checked via Synthid.com or the GeminiChat by typing @VerifyAI. However, I find that the latter option does not consistently confirm SynthID used by other companies. OpenAI also has their own checker tool: https://openai.com/research/verify/
There is also ImageWhisperer which runs 44 checks at the same time, but also doesn’t always catch SynthID. However, they recently limited free use to two requests per day.
Step Three: How to report
See this wonderful guide for all options: Options for reporting a wayspot remotely
I always recommend to fill out the abuse form for AI generated wayspots. You can include a screenshot of the SynthID.
