How to Tell If a Photo Is AI Generated

Look at the hands first, then check the text, then stop trusting your eyes entirely. That's the honest three-step answer. Visual glitches like six-fingered hands or garbled signage still show up in AI photos, but the newest models from OpenAI, Google, and Midjourney close those gaps fast. The more durable check is provenance: does the file carry a C2PA Content Credential or a SynthID watermark that survives editing? Even that isn't foolproof, since screenshots and re-uploads strip most metadata. And automated AI image detection tools, tested independently, still misclassify real photos as fake often enough that no single tool should be your final word.

That's the short version. Here's what actually holds up when you're staring at a photo someone sent you and need an answer in the next five minutes.

The visual tells that still work, sometimes

These are the classic signs of an ai generated photo, and they still catch a meaningful share of images, especially ones made with older or free tools. They are getting less reliable every model generation, so treat each one as a single data point, not a verdict.

Hands, teeth, and repeated features

Extra or fused fingers, teeth that blur into a single ridge, or jewelry that doesn't match between a person's left and right ear are still common failure points. Look for earrings that differ in style or aren't mirrored, or a wristwatch that appears on the wrong wrist relative to how the person is holding something. Crowd scenes are a good hunting ground: check whether faces in the back row repeat or blend into each other.

Text and fine detail

Signage, license plates, and text on clothing remain one of the more reliable tells. Diffusion and transformer image models still struggle to render coherent small text, so look closely at store signs, book spines, and keyboard keys. A caption baked into the image is also worth zooming in on, since garbled or nonsensical characters are a strong signal.

Reflections and physics

Water, glass, mirrors, and polished metal have to reproduce a second, geometrically correct version of the scene. Check whether a reflection in a window matches what's actually visible, or whether a shadow falls in a direction that makes sense given the light source. Reflections in eyeglasses or sunglasses are a good place to look because they're small and easy for a model to get subtly wrong.

Background geometry and noise

Look at doorframes, tiled floors, and fences for lines that warp or fail to stay parallel. Skin and hair can look slightly too smooth, especially in portraits, and backgrounds sometimes have a soft, painterly quality compared to the sharper subject in front of them.

Why these tells are getting less reliable

None of this is a permanent trick, and it's worth being honest about that upfront. Fact-checkers who do this professionally are blunt about it: the older signs, warped hands, nonsense text, six fingers, are disappearing as generators improve, and images from current-generation tools increasingly fool careful, experienced observers. A photo that passes every visual check above can still be fully synthetic if it came from a strong, recent model. Treat visual inspection as a fast first pass that can catch obviously fake images, not as proof that a clean image is real.

The more durable signal: provenance metadata

Because visual inspection degrades as models improve, the industry has shifted toward attaching a verifiable record to the file itself rather than trying to spot artifacts after the fact. This is the closest thing available to reliable ai image provenance metadata.

What C2PA and Content Credentials actually record

The Coalition for Content Provenance and Authenticity (C2PA) is the standards body behind Content Credentials, a specification for attaching a tamper-evident, cryptographically signed record to a media file at the point it's created or edited. According to C2PA's own FAQ, the credential can capture what tool created or modified the file, when, and what other assets went into it, and any change made after signing breaks the cryptographic chain in a detectable way.

The catch, which C2PA itself acknowledges, is that the credential lives in the file's metadata. Re-saving through an app that doesn't preserve it, uploading to a platform that strips metadata on ingestion, or simply taking a screenshot produces a new file with no link back to the original manifest. C2PA's answer to this is what it calls "soft bindings," invisible watermarking or fingerprinting meant to help recover the association even after the visible metadata is gone, which is effectively an admission that metadata by itself isn't durable enough.

What SynthID does differently

Google's SynthID takes the soft-binding approach directly. Instead of, or alongside, file metadata, it makes imperceptible adjustments to pixel values across the image that are designed to survive common transformations like cropping, resizing, recompression, and screenshotting. Google's own description of its SynthID Detector explains that the tool scans an uploaded image for that pattern and returns a confidence result rather than a flat yes or no, because real-world edits can partially degrade the signal.

In May 2026, OpenAI and Google aligned on a dual-layer approach: OpenAI became C2PA-conformant and began adding SynthID watermarks to images generated through ChatGPT, the API, and Codex, on top of its existing C2PA metadata. The logic is straightforward. A screenshot or re-encode strips the C2PA manifest almost immediately, but a watermark embedded in the pixels themselves has a better chance of surviving, which is why pairing the two is more resilient than relying on either one alone.

Where both still fall short

Neither approach covers images that were never watermarked in the first place. Open-source and self-hosted generators that don't participate in either scheme produce images with no embedded signal to check for, so a clean provenance check tells you nothing about those. Provenance also only proves origin, not truthfulness: a real photo with a fabricated caption, or an AI-generated background composited under a real subject, can carry misleading signals on the parts that genuinely came from a camera. For a related mechanism applied to a different medium, see how SynthID's audio watermarking works, and for the regulatory side of why platforms are being pushed to attach this kind of metadata at all, see the EU's 2026 AI transparency requirements.

Should you trust an automated AI image detector?

Be skeptical of any ai image detection tool that hands you a single percentage and calls it a verdict. NewsGuard ran an audit in 2026 testing five widely used detection tools, Hive, AI or Not, ZeroGPT, Sightengine, and ScamAI, against 45 images: 15 authentic news photos, 15 lightly edited versions, and 15 heavily manipulated ones. The tools collectively misclassified genuine, unaltered photos as AI-generated 13.33 percent of the time. One tool, ScamAI, flagged 40 percent of authentic images as fake. The five tools disagreed with each other on 35 of the 45 images tested. Industry sources cited in the NewsGuard report pointed to resizing, compression, unusual lighting, and high contrast, exactly the conditions common in real news photography from breaking events, as reliable ways to trigger a false positive.

That matters beyond curiosity. A detector confidently telling someone a real photo is fake, or the reverse, has been used to dismiss genuine evidence and to lend false credibility to fabricated images. If you're building or recommending detection tooling for clients, treat any single detector's output as one weak vote, not a finding, the same caution worth applying when someone cites a tool's output as proof inside an AI-generated scam pitch.

A practical checklist

When you actually need to assess a specific photo, work through it in this order.

• Check for a Content Credential or SynthID result first, since it's the closest thing to a durable signal, but note whether the image has clearly been re-saved, screenshotted, or downloaded from a platform known to strip metadata.

• Scan hands, teeth, jewelry, text, and reflections for the classic tells, understanding this catches older or lower-effort fakes more than current frontier models.

• Run the image through more than one detector tool if you use them at all, and treat agreement across tools as more meaningful than any single score.

• Consider the source and context: where the image first appeared, who posted it, and whether the surrounding claim makes sense independent of the pixels. Questions about who's accountable for AI output apply just as much to images as they do to AI-generated code.

• When the stakes are high, legal, financial, or reputational, don't rely on a single automated answer at all. Combine a metadata check, a reverse image search, and, where possible, contact with the original source.

None of these methods is a silver bullet by itself. Used together, they get you closer to a defensible answer than any one check alone.

Frequently asked questions

What is the most reliable way to tell if a photo is AI generated?

No single check is fully reliable. The closest thing to durable evidence is provenance data, a C2PA Content Credential or a SynthID watermark, checked alongside classic visual tells like hands, text, and reflections. Provenance survives editing better than visual artifacts do, but both can fail, so treat them as complementary checks rather than one final answer.

Do AI image detector tools actually work?

Inconsistently. A 2026 NewsGuard audit of five widely used detectors (Hive, AI or Not, ZeroGPT, Sightengine, and ScamAI) found they collectively misclassified real, unedited news photos as AI-generated 13.33 percent of the time, with one tool wrong 40 percent of the time, and the tools disagreed with each other on 35 of 45 test images. Use them as one input, not a verdict.

Can you tell if a photo is AI generated after it's been screenshotted or re-uploaded?

It gets much harder. Screenshotting or re-encoding a file typically strips embedded C2PA metadata entirely. A SynthID-style pixel watermark is designed to survive that kind of transformation better, since Google reports testing it against cropping, resizing, and recompression, but no method is guaranteed to survive every edit.

What are the classic signs of an AI generated photo?

Look at hands and teeth for extra or fused features, check whether earrings or jewelry match between left and right, zoom into any text or signage for garbled characters, and check reflections in glass, water, or eyeglasses for physical inconsistencies. These signs are still useful but are becoming less reliable as image models improve.

What is C2PA and does every AI image have it?

C2PA (Coalition for Content Provenance and Authenticity) is the standards body behind Content Credentials, a cryptographically signed record of a file's origin and edit history. Not every AI image carries one. Coverage depends on whether the generator or platform that made the image participates in the standard, and open-source or self-hosted models often don't attach any credential at all.