Real or AI? The 3-Layer Check That Actually Settles It

Check if an image is real or AI in three layers: scan the pixels, look for provenance, then weigh the context. Each one still works when the others come up empty.

To check if an image is real or AI, run three layers in order: scan the pixels, look for provenance, then weigh the context. No single layer settles it, and the order matters because each one still works when the layer before it comes up empty. Start with GPTOne's free image detector, which gives you a confidence score and a heatmap in seconds.

Most guides hand you a list of visual tells and stop. That worked in 2023. Generators fixed the hands.

Key Takeaways

• 3 layers, in this order: pixels, provenance, context. Each survives the failure of the others.
• Provenance is the strongest evidence when present, and absent on the large majority of images you'll encounter.
• Visual tells are now the weakest layer on their own, since current models render hands, teeth and text far better than they did 2 years ago.
• Any 2 layers agreeing is a solid conclusion. All 3 disagreeing means you don't have an answer yet, and saying so is the correct outcome.
• The whole check takes about 60 seconds once you've done it a few times.

Why one check is never enough

Every individual method has a clean way to fail.

Visual inspection fails because generators improved. Metadata fails because platforms strip it. Reverse image search fails because a freshly generated image has no earlier copy to find. Detectors fail on heavily processed files. Provenance fails because most generators never applied any.

Stack them and the failures stop overlapping. That's the entire idea, and it's why a single tool, including ours, is not a complete answer.

Layer 1: read the pixels

Start here, because it's the only layer that survives distribution. Screenshots, re-uploads and compression destroy metadata but leave the pixel evidence largely intact.

Upload the image to a detector and read two outputs, not one. The score tells you how strong the overall signal is. The heatmap tells you where it sits, which is the part that actually guides your next move. Our post on why an explanation beats a bare verdict covers reading them together, and how detectors work explains the signals underneath.

Two things to know before you trust the number. Heavily processed real photos get flagged, and the causes are predictable enough that we wrote up the false-positive patterns in full. And there is no honest fixed accuracy percentage for image detection, which is why we published our 60-image benchmark with its weak categories included.

Once the heatmap points somewhere, check that spot against the 12 visual signs. Two independent methods agreeing on the same patch of pixels is much stronger than either alone.

Layer 2: look for provenance

Provenance is a claim about the file's history, and when it's intact it outranks everything else here.

C2PA Content Credentials are a signed manifest describing how a file was made and edited. The C2PA standard is backed by Adobe, Microsoft and OpenAI, and ChatGPT images carry these credentials. Our guide to verifying them walks through reading one.

SynthID is an imperceptible watermark Google DeepMind embeds in the pixels of its models' output. Because it lives in the pixels, it survives cropping and compression better than any metadata field. Our SynthID explainer covers its reach and its limits.

EXIF is the camera's own record. A full, plausible camera block is a decent signal of real capture, and our walkthrough of reading it shows what the values should look like.

The catch across all three: absence proves nothing. Platforms strip metadata on upload, screenshots replace it wholesale, and most generators never marked the file at all, which we cover in why metadata and watermarks can't be relied on. Treat this layer as a bonus that sometimes settles the question outright and usually tells you nothing.

Layer 3: weigh the context

The layer people skip, and often the one that resolves the case.

Who posted it, and do they have a history? When did it first appear, and does the timing fit the event it claims to show? Does anything corroborate it, meaning a second photo of the same scene from a different angle, or a report from an unrelated source? Do the details in frame match the claim, including weather, signage, language and season?

Reverse image search belongs here rather than in layer 1. It's excellent for finding an earlier or original version of a real photo and close to useless on a fresh generation, a difference we cover in reverse image search versus AI detection.

For a viral news image the context layer usually beats both others, which is why our viral image verification guide leans on it heavily.

How to weigh the three layers when they disagree

That last row is the one worth internalising. "I can't tell" is a real finding, and it's a much better outcome than a confident wrong call, particularly when a person is on the other end of it.

The 60-second version

Scan it and look at where the heatmap sits. Check whether any provenance survived. Ask who posted it and whether anything corroborates it. If two of the three point the same way, you have your answer. If they don't, you don't, and the honest move is to keep it unresolved.

Where each layer breaks

Pixels break on screenshots, heavy compression and AI upscales. Provenance breaks on stripped files and unmarked generators, which is most of them. Context breaks on a plausible-looking account with no history you can check.

Knowing which layer is weak for the image in front of you tells you how much weight to put on the others. A screenshot from an anonymous account means all three layers are degraded, and that's exactly when people reach hardest for a single number.

Adjust the order for what you're checking

The three layers stay the same. Which one carries the most weight changes with the situation.

A profile photo. Context first. An account with no history and a single photo tells you more than any score. Our platform-by-platform guide covers where to look.

A marketplace listing. Pixels first, since sellers reuse and generate product shots constantly, and our fake product photo checker guide covers the patterns.

A news image. Context first, and hard. Timing and corroboration decide most viral cases before a detector is involved.

Digital art. Pixels, with the art checker approach, because provenance and context are usually unavailable.

A claim or an assessment. Provenance first, because it's the only layer that produces evidence you can hand to someone else.

FAQ

How can I tell if an image is real or AI for free?

Scan it with a free detector for the score and heatmap, check for C2PA credentials or camera EXIF, then look at the source and whether anything corroborates it. GPTOne's image detector is free to try on signup credits, one scan a day at no cost, and includes the region heatmap and provenance reading that most tools charge for.

What is the most reliable way to detect an AI image?

Intact provenance, when you can get it, because a signed history beats any inference from pixels. Failing that, pixel-level detection plus context. Visual inspection alone is now the least reliable of the three.

Can AI images pass as real in all three layers?

A generated image that's been downscaled, grain-added and posted by an established account can survive all three. That's rare and it takes deliberate effort, but a check that returns "uncertain" is the correct response to it.

Do I need to check the metadata if the detector already gave me a score?

Only if you need a firm answer. Provenance can turn a probable into a certain, and takes about ten seconds when it's there. It just isn't there most of the time.

Is a 70% score high enough to call something AI?

Not on its own. Read where the heat sits, then get a second layer to agree before you conclude anything you'd have to defend.

The short version

Pixels, provenance, context. Run them in that order, accept "uncertain" when the layers disagree, and never hang a decision that affects a person on one percentage.

Start with a free scan at GPTOne, then work down the layers. Checking text as well? The free AI text detector covers every major model at 99.99% accuracy. Both at gptone.me.