How can you tell if an image is AI generated?

Visual tells: what still works and what does not

Google's own verification guidance lists the artefacts still worth checking: garbled or inconsistent text inside the image, errors in small details such as hands, teeth and background objects, shadows and lighting that do not agree, and unnaturally repetitive patterns.

What no longer works is treating a clean image as evidence of a real one. Extra fingers and melted jewellery were dependable signs in 2022-era models. Current models render hands and faces correctly most of the time, so the absence of mistakes tells you nothing. The face thread covers how poorly people do when they rely on looking alone.

Reverse image search

Run the image through Google Images or TinEye and look for the earliest appearance. If the same picture predates the generator that supposedly made it, or appears in a photographer's archive or news wire, that is strong evidence it is real. If the only sources are recent social reposts, you have learned nothing either way.

Metadata and Content Credentials

A camera writes EXIF data (model, lens, timestamp). It is easy to strip or forge, and screenshots discard it, so it is weak evidence on its own.

Content Credentials are stronger. The C2PA specification defines a cryptographically signed manifest that records who or what produced the file and how it was edited. You can inspect a file at verify.contentauthenticity.org. A valid manifest naming an AI tool is close to proof. The catch, stated in the specification itself, is that an asset can become separated from its manifest when metadata is removed, so a missing manifest proves nothing. The C2PA thread goes deeper.

SynthID for Google images

SynthID embeds an invisible watermark in images made by Google's generative models, designed to survive cropping, filters and lossy compression. You can check a file by uploading it to the Gemini app and asking whether Google AI made it. Google's help page is explicit about the limit: Gemini currently recognises only Google-generated content, so "no watermark found" means "not Google", not "not AI". It also warns that very simple content or a tiny edit may carry no detectable watermark.

Detector tools

Hive, Illuminarty and AI or Not are classifiers that return a probability. A 2024 study by Ha and colleagues, tested them on art from five generators. Hive reached 98.03% accuracy with no false positives on unperturbed images, but its detection rate fell below 70% when images were protected with Glaze. Illuminarty scored 72.65% accuracy with a 67.40% false positive rate. AI or Not advertises 98.9% accuracy on its own site, which is a vendor claim, not an independent result.

Six ways to check an image, and what each can actually prove

Method	What it can prove	Where it fails
Visual inspection	Suggests AI when text or fine detail is broken	Clean output proves nothing; tells fade with each model
Reverse image search	Real, if the image predates the generator or has a documented source	Silent on new or private images
EXIF metadata	Weak support for a camera origin	Stripped by screenshots, trivially forged
Content Credentials (C2PA)	AI origin and edit history, when a signed manifest is present	Optional; lost when metadata is removed
SynthID via Gemini	Google AI origin, even after moderate edits	Other generators return no result; tiny edits may not carry it
Detector tools (Hive, Illuminarty, AI or Not)	A probability, useful as one signal	Adversarial edits, new generators, false positives on some tools

What that means in practice

• Start with provenance, not pixels. Check for Content Credentials first, then SynthID if a Google model is plausible.
• Treat a detector score as one witness. Combine it with reverse search and the story around the image.
• Ask for the original file. Screenshots and re-uploads destroy the metadata that could have settled the question.