How to tell if a photo is AI-generated, and what the checks can prove

Do visual clues still work on faces?

Sometimes, but they are weaker every year and should only make you look closer. The FBI's tip list names distorted hands or feet, unrealistic teeth or eyes, indistinct or irregular faces, unrealistic accessories such as glasses or jewelry, and inaccurate shadows. Technology writers add garbled text, extra fingers, plasticky skin, overlapping teeth and glassy eyes, while noting that image models keep improving. National Geographic reports that the tells that worked a year earlier can already be out of date, and that the eye alone is not reliable. IC3 ZDNET National Geographic

• Hands and feet: extra or merged fingers, missing knuckles, or extra limbs and misproportioned necks.
• Faces: unrealistic teeth or eyes, a glassy stare, or skin without pores or wrinkles.
• Accessories and light: glasses, jewelry and shadows that do not sit where they should.
• Text and background: garbled lettering, repeated faces, or reflections that defy physics.

What do Content Credentials actually prove?

They can prove that a file was signed by a listed tool and has not changed since, but not that the scene is real. Content Credentials, built on the C2PA standard, are a cryptographically signed record that can travel with a file and list how it was created, which tools were used, when and where it was made, and how it has changed. Any modification, intentional or accidental, breaks the cryptographic linkage, which is what makes the record tamper-evident. That is narrower than it sounds: a signature cannot verify that the metadata is accurate, and NIST describes it as a notarization that the signer attested to the claim. Credentials can also be stripped, which is why C2PA supports invisible watermarks and fingerprints that may help recover them. C2PA NIST

What does a SynthID watermark prove?

A detected SynthID watermark can show that a file was made with AI by Google or a participating partner; it cannot show that an unmarked file is real. Google says it has watermarked more than 180 billion images and videos, and that its SynthID Detector, available globally in English since October 2026, can check whether an image, video or audio file carries the signal. OpenAI's checker supports SynthID and C2PA, and calls a detected watermark evidence of a specific signal rather than a complete history. A not-detected result means the signal was not found, not that the file is human-made. Google OpenAI

Why can a detector miss an AI-generated photo?

A detector can only report the signals it supports, and those signals do not survive every edit. OpenAI's documentation says a not-detected result can happen when metadata was stripped or tampered with, a watermark degraded, the file came from an older model, or the content predates provenance signals; it also says the tool does not detect other companies' models. Compression, cropping, screenshots, metadata removal and format conversion can erase or weaken a signal. NIST adds that any covert watermark detector has some nonzero chance of false positives and false negatives, and ZDNET reports that even free detectors can be fooled by sophisticated fakes. OpenAI NIST ZDNET

Why does an AI-generated face have no search history?

Because the face belongs to no one, there are no other photographs of that person to find. A reverse image search looks for the picture itself, and a generated picture can be reposted many times, so it may have a trail of uses and claims. A face search asks a different question: whether the same face appears in other photos on public pages. A generated face has no separate person behind it, so there is usually nothing else to match. The opposite pattern matters: if the same face appears under another name, the photos may have been taken from a real person, which the FTC lists as a sign of a scam. National Geographic FTC

Can a face search help you check a photo?

It can, as one piece of evidence, but it cannot tell you by itself whether a photo is AI-generated. FaceSearch compares the face in a photo with images on the public web. If the same face appears under another name, that fits the FTC's warning that a profile picture linked to another name is a sign of a scam. If nothing appears, the result is consistent with a generated face, but also with a real person who has no public photos. FaceSearch cannot see profiles inside dating apps, private accounts or messages, so a blank result says nothing about what is hidden there. Run a face search and read what a match and a blank result mean. FTC

What should you do before you trust a photo?

Check the file and the face separately, and do not let a single result decide for you. The steps below combine the checks in this guide; none of them proves a photo is real on its own. If the person asks for money, gift cards or cryptocurrency, stop: the FBI says not to send money or assets to someone you have only met online, and the FTC's bottom line is the same. Report the profile to the platform, and if you lost money, report it to the FBI's IC3 at ic3.gov or the FTC at ReportFraud.ftc.gov. IC3 FTC