How to tell if an image is AI generated
Four checks, from quickest to most reliable. None is proof on its own.
The short answer
Check four things: a label in the file (Content Credentials), the picture itself (light, text, details), its source (who posted it first, and when) and a detector built for the job. Perspicax reads the label and the pixels together and answers likely AI-generated, likely real or inconclusive.
Check an image now
Drop, paste or link it. The label check and the pixel check run together.
Four checks
Look for a label
Many generators sign their files with Content Credentials (C2PA), and some platforms show an AI label. A signed label is strong evidence, but one re-save or screenshot removes it, so a missing label proves nothing.
- Content Credentials (C2PA)Live
- SynthID watermark checkSupported
Look closely
Newer generators rarely get hands or text wrong, so the old giveaways no longer work. Look instead at things that must agree: light and shadows, reflections, small print and logos, and background details that merge or repeat. Skin and surfaces that look too clean are a weak hint, not proof.
Check the source
Find where the image first appeared, who posted it and when. A reverse image search can turn up the original, other angles of the same scene, or an earlier fact-check.
Use a detector
A detector reads what the eye cannot. Perspicax examines the whole image region by region and answers in three bands; when the evidence is weak, it says inconclusive instead of guessing.
Questions people ask
Can ChatGPT identify an image?
ChatGPT can describe a picture and guess, but it isn't built to detect AI images and can be wrong either way. OpenAI's separate verify tool only checks for OpenAI's own signals. A dedicated detector checks the pixels of any image.
Does ChatGPT use SynthID?
Yes. OpenAI says images from ChatGPT carry a SynthID watermark as well as C2PA Content Credentials. Either signal can point to an OpenAI image, but only the watermark usually survives a re-save.
Can C2PA be removed?
Yes. Content Credentials live in the file's metadata, so a screenshot, a re-save or most social uploads remove them. We saw it ourselves: one re-save erased Google's signed label from our test photo. That is why Perspicax also checks the pixels.
Can you tell if an image is AI just by looking?
Sometimes, but less and less. The newest models get hands, text and lighting right most of the time, so looking is a first filter, not an answer.
How accurate is an AI image detector?
It depends on the detector and the images. Ours is measured on public test sets and the newest models: up to 99% of AI images caught, depending on the model, with fewer than 1 in 100 camera photos wrongly flagged.
What if the detector says inconclusive?
Treat it as "not enough evidence": go back to the source, look for the original file and check it for Content Credentials.
Before you believe it, check it.
Your first scan is free, no account needed. A free account gives you 2 scans every day.