Check if an image was generated by AI

How AI image detection works

With the spread of tools like Midjourney, DALL-E, Stable Diffusion, and the generators built into ChatGPT and Gemini, telling a real photo from a fully synthetic image has become genuinely difficult — even for a trained eye. Several technical approaches coexist, with very different reliability levels.

Invisible watermarking (SynthID, C2PA)

Major platforms increasingly embed a signal directly at generation time: SynthID, developed by Google DeepMind, inserts an imperceptible pattern into the pixels the moment an image is created. The C2PA standard instead attaches cryptographic metadata tracing the file's provenance. This is the most reliable method — but it only works if the tool that created the image implements one of these standards, and if the metadata hasn't been stripped along the way (a simple screenshot is enough to wipe C2PA).

Forensic analysis (what this tool does)

Without a verifiable official watermark, you can look for indirect traces left by the generation or editing process:

Visual inspection

Certain details remain statistically harder for current models to render correctly: hands and fingers, text embedded in the image, the consistency of reflections and shadows, or repeating background patterns. This is less reliable than it was two years ago — recent models keep closing these gaps — but it's still a useful habit alongside the other signals.

Why verify an image's origin?

• Online marketplaces — spotting AI-generated product photos used to pass off cheap new items as vintage, or fake damage evidence in refund claims.
• Fact-checking — avoiding sharing a misleading image presented as a real event.
• Content moderation — filtering visuals in contexts that require an authentic origin.
• Personal curiosity — simply knowing whether a circulating image is real before sharing it.

Limitations of this tool

To be upfront: this detector relies solely on heuristics computed locally in your browser, without a deep learning model trained on millions of images. It cannot compete with commercial detectors or proprietary systems like SynthID.

• A well-made AI image with no suspicious metadata may trigger no signal at all.
• A real photo that's heavily compressed or cropped can, conversely, trigger false signals.
• This tool does not verify SynthID watermarks or C2PA signatures — for that, you need each platform's official tools.

Use it as a quick first filter, not as definitive proof.

Glossary

Deepfake

Synthetic visual or audio content designed to imitate a real person or scene, often used deceptively. An AI-generated image isn't necessarily a deepfake — the term specifically implies impersonation or falsified reality.

SynthID

Invisible digital watermark developed by Google DeepMind, embedded directly into an image's pixels at generation time. Not detectable by this tool or any unofficial third-party tool — only the Gemini app or Google's SynthID Detector portal can verify it.

C2PA (Content Credentials)

Open cryptographic provenance standard adopted by Google, Adobe, Microsoft, and OpenAI. Relies on metadata attached to the file — a screenshot or recompression is usually enough to erase it.

ELA (Error Level Analysis)

Forensic technique that recompresses an image and compares the result to the original to spot areas where the error level differs from the rest.

EXIF metadata

Technical information embedded by a camera: make, model, date, settings. Its absence is a weak clue, since many real photos lose it when shared via messaging apps.

Frequently asked questions