How to Tell If an Image Is AI-Generated: 7 Reliable Checks

AI image generators now produce pictures that fool most people at a glance. There is no single test that proves an image is synthetic, but combining a few checks gives you a much more reliable answer. Here are seven, ordered from quickest to most thorough.

1. Read the file’s metadata

Many AI tools announce themselves inside the file. Drop the image into our AI metadata checker (it runs locally and changes nothing) and look for:

• A C2PA manifest — a signed Content Credentials record. ChatGPT images, Adobe Firefly output and Google’s newer image models include one.
• IPTC Digital Source Type set to trainedAlgorithmicMedia (created by AI) or compositeWithTrainedAlgorithmicMedia (a real photo edited with AI).
• Software or creator fields mentioning Midjourney, DALL·E, Firefly, Stable Diffusion, ComfyUI and others.
• PNG text chunks containing a prompt, seed and sampler — a strong sign of Stable Diffusion.

Positive metadata is strong evidence. Missing metadata proves nothing, because most social networks strip it on upload and anyone can remove it.

2. Verify Content Credentials

If a C2PA manifest is present, the Content Credentials verification site from the C2PA/Content Authenticity Initiative can display who signed it and what actions were recorded. A valid signature from a known AI provider is about as conclusive as metadata gets.

3. Check for invisible watermarks

Some providers hide a watermark in the pixels themselves. Google’s SynthID is the best-known example; images made with Google’s tools can be checked by uploading them to the Gemini app and asking whether they were created with Google AI. These watermarks survive screenshots and moderate edits, but each one can only be read by its own vendor’s tool.

4. Look for visual tells

Modern generators have fixed many classic mistakes, yet these still appear regularly:

• Text and logos that are slightly garbled, especially small print in the background.
• Hands, jewellery and glasses with odd joins, extra parts or asymmetric details.
• Repeating textures in hair, foliage, crowds or brickwork.
• Lighting that does not add up — shadows pointing different ways or reflections that show something else.
• “Too perfect” skin with a waxy, airbrushed finish and identical pores across the face.
• Background objects melting into each other at the edges.

Treat these as hints, not proof: real photos edited with beauty filters can look similar.

5. Run a reverse image search

Search the image with Google Lens, Bing Visual Search or TinEye. If it first appeared on an AI art community, a prompt-sharing site or a stock library’s AI section, you have your answer. If it appears in older news coverage with a named photographer, it is likely real. Google’s “About this image” panel can also show when Google first indexed it and, where available, AI labels from metadata.

6. Check the context and source

Ask the journalist’s questions: who posted it first, when, and do other photos from the same event exist from different angles? Real events usually produce many independent images. A single dramatic picture with no other coverage deserves suspicion.

7. Use AI detectors — carefully

Online AI-image detectors analyse pixel statistics and give a probability. They can help as one signal, but they produce false positives on heavily edited real photos and false negatives on new models they were not trained on. Never rely on a detector alone for a decision that affects someone.

Putting the evidence together

Why this matters

Being able to judge an image protects you from scams, fake reviews and misleading news. It also helps creators: if your own real photo is wrongly flagged as AI, checking its metadata is the first step to understanding why — editing apps often add AI tags after a small generative fill.

Worked example: a viral “photo” of a storm

Imagine a dramatic picture of a lightning storm over a famous bridge spreads on social media. Here is how the checks play out:

1. Metadata: the copy you downloaded from a social network has no metadata — expected, because networks strip it. This tells you nothing.
2. Reverse search: the earliest copy appears on an AI art forum two days before it went viral, posted with a prompt. That is strong evidence.
3. Visual check: the bridge has one cable too many and the lights on the water reflect a different pattern from the sky.
4. Context: no news outlet or weather service reported a storm there that night, and no one else posted photos.

Together these point clearly to AI. None of them alone would be conclusive.

Signals that commonly mislead

• “It looks too good.” Professional photographers and modern phones produce very polished images.
• Missing EXIF. As explained above, most shared images have none.
• A single detector score. Treat it as a hint only.
• Smooth skin. Beauty filters in camera apps create the same effect.

Tools worth bookmarking

• Our AI metadata checker for C2PA, IPTC tags, generator names and prompts.
• The official Content Credentials verify site for signed manifests.
• Google Lens, Bing Visual Search and TinEye for reverse image search.
• The Gemini app for checking Google’s SynthID watermark.

Frequently asked questions

Can metadata prove an image is real?

No. Metadata can be added, edited or removed. Camera-signed Content Credentials from devices such as some Leica, Sony and Nikon models are the strongest evidence of authenticity available today.

Are AI image detectors accurate?

They are improving but still make errors in both directions. Use them as one signal among several, never as proof.

Why does my real photo show AI signals?

Editing tools such as generative fill, AI noise reduction or AI upscaling can add C2PA or IPTC “composite with trained algorithmic media” tags to a photo that is mostly real.

Can I tell if an image is AI from a screenshot?

Screenshots lose metadata, so you must rely on visual clues, reverse search and watermarks that survive screenshots, such as SynthID.

What does “composite with trained algorithmic media” mean?

It means a real image was partly edited with generative AI, for example with generative fill or outpainting.