How to Tell if a Photo Is AI-Generated (2026)

The fastest reliable check in 2026 isn't a vibe or a magnifying glass — it's provenance. Look for a C2PA Content Credential or a SynthID watermark that says how the image was made. If a trustworthy credential is present, you have a real answer. If it's absent, you have a hint, not a verdict. Everything else — visual tells, metadata, detector apps — is supporting evidence you layer on top.

Here's how to actually run the check, in order of how much each step is worth.

Start with provenance: Content Credentials and SynthID

The industry has converged on two layers of "birth certificate" for images.

C2PA Content Credentials. This is a signed metadata manifest — a cryptographic record of how an image was created and edited. The C2PA coalition now spans 6,000+ members including Google, Microsoft, Adobe, Meta, OpenAI, Sony, the BBC, and Amazon. Microsoft began attaching C2PA metadata to Microsoft 365 content in early 2026.

• How to check: Look for a small "Cr" credentials icon in supported apps, or drop the image into Content Credentials' Verify tool (verify.contentcredentials.org). It'll show whether the file declares an AI origin and what edits were logged.

SynthID. Google's imperceptible watermark, baked into the pixels themselves so it survives cropping, compression, and screenshots better than metadata does. Google has watermarked over 20 billion images, and by mid-2026 SynthID verification for images, video, and audio was available in Gemini and expanding to Search and Chrome surfaces.

• How to check: Use Google's SynthID verification in supported surfaces to test whether the image carries the watermark.

The key limitation: a present, valid credential is strong proof of AI origin. But absence proves nothing. Most AI tools in 2026 don't add watermarks at all, and even a marked file can lose its credential through screenshots, re-compression, or editing. No credential means "unknown," not "real."

Inspect the metadata (a hint, not proof)

Every camera photo carries EXIF metadata — the make and model of the camera, lens, exposure settings, and often a timestamp and GPS coordinates. Purely AI-generated images usually have none of that camera data.

So when you open a file's metadata and find no camera model, no lens, no exposure info — just a bare file or a generator's name — that's a signal worth noting. You can read exactly what's embedded with a free EXIF viewer; it takes seconds.

But hold this loosely, because metadata is weak evidence in both directions:

• Missing EXIF doesn't mean AI. Screenshots, social-media uploads, and privacy tools all strip EXIF from genuine photos. Every platform that scrubs metadata leaves a real photo looking "camera-less."
• Present EXIF doesn't mean real. Metadata can be faked or copied from a real photo onto a generated one.

So treat metadata as one input. A real photo can have its GPS and camera data removed on purpose — that's exactly what happens when you remove GPS from a photo before sharing. Absence of camera EXIF is a nudge toward "investigate further," not a conclusion.

Look for the visual tells that still work

Detectors miss things; your eyes still catch some. The tells that held up best in 2026:

• Text is the weakest link. License plates, shop signs, book spines, and menus often show garbled, warped, or nonsensical lettering. If the letters don't form real words, that's a fast, high-confidence tell.
• Hands and fingers. Fused, extra, or oddly bent fingers remain common.
• Impossible physics. Depth of field that doesn't add up, shadows falling in conflicting directions, reflections that don't match the scene, jewelry or straps that merge into skin.
• Too-perfect texture. Skin, foliage, and backgrounds that look uncannily smooth or repeat in subtle patterns.
• Melting details in the background. The subject looks sharp, but signage, crowds, and small objects behind them dissolve.

Zoom in to 100% and scan the edges and background — that's where generators cut corners.

Use detector tools — and know their ceiling

AI image detectors are useful, but be honest about what they deliver.

• Leading detectors advertise 95–99% accuracy on clean lab datasets, and the better ones land around 85–95% on standard test sets.
• Real-world numbers are much lower against current models. A 2026 benchmark of 23 detection tools found roughly 75% accuracy against 2020–2021 generators — but only about 18–30% against today's commercial models like Flux and recent Midjourney.
• Detectors also generalize poorly: one trained mostly on Midjourney can underperform on DALL-E-class images.

Practical rule: a detector flagging "AI" is worth weighing, especially if several agree. A detector saying "human" on a modern generation is often just wrong. Never rely on a single tool's score alone.

Worth knowing: automated tools still beat pure human guessing. In one large Microsoft study, human judges hit only about 62% accuracy across 600,000+ images. Your eyes are a complement to the tools, not a replacement.

Put it together: a practical checklist

No single step is decisive. Stack them:

1. Check provenance first. C2PA Content Credentials and SynthID. A valid AI credential ends the investigation.
2. Read the metadata. Missing camera EXIF is a hint to dig deeper — never proof on its own.
3. Reverse-image search. Find the earliest appearance and original context; it often exposes the real source.
4. Scan the visual tells. Text, hands, physics, texture, dissolving backgrounds.
5. Run two or three detectors. Treat agreement as meaningful, a lone "human" verdict as unreliable.
6. Weigh the source and context. Who posted it, when, and does the claim make sense?

The bottom line: in 2026 there's no one-click "is this AI?" button you can fully trust. Provenance gives you certainty when it's present; everything else is probability. Combine provenance, metadata, reverse search, visual inspection, and detectors, and let them vote.

A good place to start is the metadata itself. Drop any image into GeoTag.world to find where a photo was taken and see its full EXIF — camera model, timestamps, GPS, the works. If those fields are simply missing, you've got your first clue worth chasing. It's free and runs entirely in your browser.