You can often tell an image is AI-generated by checking its metadata for Content Credentials, looking for generator watermarks, and inspecting a few known visual failure modes — but no single cue is conclusive, so you should combine at least two checks before you decide.
The fastest first move is to open the image's file details and look for a C2PA manifest, which is a signed record that many generators now attach at creation. If that record is missing, move to visual inspection: hands, text, reflections, and background geometry are where image models most often slip. Treat every individual signal as a hint, not a verdict.
The reason detection works at all comes down to how these images are made. An image generator produces pixels by predicting what should come next, which means it has no underlying model of how a hand connects to a wrist or how a reflection should line up with its source. That missing structure is why errors cluster in the same places across different tools.
Metadata is the other half of the story. The C2PA standard, short for Coalition for Content Provenance and Authenticity, lets a generator embed a cryptographically signed note saying which tool made the image and whether it was edited afterward. Google's "About this image" feature surfaces similar context in search results, and the C2PA Verify tool reads the manifest directly.
Neither is foolproof, because a screenshot or a re-upload to most social platforms strips that data clean off.
Here is a worked example. Suppose someone sends you a photo of a crowded street market and claims it was taken last weekend. You open the file properties and find no C2PA manifest, which is common after a messaging app re-compresses the image.
You zoom in on the vendor signs and notice the lettering is shaped like text but reads as nonsense syllables. You check the background and see two people whose shoulders merge into one another, and a window reflection that shows a building not present in the scene. Three independent cues pointing the same direction is a reasonable basis for doubt.
One cue alone is not. A single mangled sign could be a low-resolution photo; a missing manifest could just mean the platform stripped it.
Be honest about where this breaks down. Metadata is fragile — screenshots, social uploads, and many editing apps remove it, so absence proves nothing. Detectors that output a percentage score are useful but produce false positives, especially on heavily edited real photos and on images of paintings or illustrations.
Modern generators have also gotten better at hands and text, so those cues are weakening over time. A practical decision rule: trust a positive C2PA manifest more than any visual cue, because it is signed; treat a missing manifest as neutral; and require at least two independent visual cues before you call something AI-made.
For pricing context on the tools involved, our AI tool database lists Midjourney's Basic plan at $10 per month and Adobe Photoshop's Photography Plan at $9.99 per month, but the detection methods above work regardless of which generator produced the file. If you are checking a piece of writing rather than an image, the same instinct applies — look for structural tells rather than vibes, which is the approach in How can I tell if AI wrote a piece of content before I publish it?.