How to Tell If an Image or Video Is AI-Generated

Can You Tell If an Image or Video Is AI-Generated?

You usually cannot prove that an image or video is AI-generated by looking at it. Visual mistakes and detector scores can support an investigation, but the stronger method traces the earliest credible source, inspects provenance credentials and supported watermarks, searches for earlier versions, checks the surrounding event, and reports what remains uncertain.

The responsible result is not always real or fake. It may be supported, contradicted, altered, synthetic within a named system, or unverifiable from the available record.

Why appearance is weak evidence

Generated media can contain inconsistent text, geometry, reflections, hands, shadows, motion, or physics. Real media can contain the same apparent defects after compression, low light, rolling shutter, stabilization, editing, or a bad camera angle.

Visual inspection is useful for finding questions. It is not a reliable origin test by itself.

Generic AI detectors face a similar problem. They classify patterns learned from examples. Their result can change when a new model, editing method, compression setting, or content category falls outside the evaluation data.

NIST's synthetic content report separates provenance tracking from synthetic-content detection. It describes watermarking, metadata, content authentication, detection, testing, and ongoing maintenance as different parts of a transparency system.

Start with the source

Before studying pixels, ask where the file came from. Find the earliest known post, page, photographer, broadcaster, agency, public record, or witness. Compare the account's history, date, location, motive, access, and whether other credible sources observed the same event.

A repost can remove context while leaving the image unchanged. A real image can be attached to the wrong event. A generated image can be accurately labeled by its creator and then circulated without the label.

For consequential media, save the URL, time, original caption, account, file, and any changes you make during inspection.

Inspect Content Credentials

Content Credentials implement the C2PA standard. A credential can record how an asset was created or edited, identify the software or device that signed assertions, and make later tampering with the signed record detectable.

The current C2PA explainer makes two limits explicit. Metadata can be removed, and provenance does not prove that the depicted event is true.

flowchart TD
    A["Suspicious media"] --> B["Trace earliest credible source"]
    B --> C["Inspect Content Credentials and metadata"]
    C --> D["Check supported watermarks"]
    D --> E["Search for earlier or related versions"]
    E --> F["Corroborate event, date, place, and people"]
    F --> G["Supported, contradicted, altered, or uncertain"]

A valid credential can establish that a trusted camera signed an image and that the signed file has not changed. It cannot establish that the scene was not staged. A credential from an unknown signer is evidence about the record, not automatic evidence that the signer is trustworthy.

An absent credential proves little. Many legitimate cameras and workflows do not create C2PA records, and platforms or editing tools may strip metadata.

Check a supported watermark

Google DeepMind's SynthID page says supported Google systems place imperceptible watermarks in generated images, video, audio, and text. It says users can upload an image, video, or audio file to Gemini and ask whether a SynthID watermark is present.

SynthID is designed for content produced or altered by supported Google AI systems. A negative result cannot establish that a file is human-made or that another model was not used.

The current documentation says image and video watermarks are designed to survive cropping, filters, frame-rate changes, and lossy compression. This corrects the common claim that any crop necessarily removes the signal.

DeepMind's technical announcement also says SynthID is not foolproof. Watermark detection is one piece of evidence with a defined system boundary.

Search for earlier versions

Use reverse-image search, key-frame search for video, exact caption searches, and searches for distinctive objects, locations, or people. Earlier versions may reveal a different crop, original caption, generator label, news report, stock image, or unrelated event.

Video should be inspected as a sequence. Extract key frames, compare cuts and audio, and look for independent recordings from different positions. A single plausible frame does not validate the entire clip.

Metadata can help identify creation time, device, location, software, and editing history. It can also be missing, wrong, copied, or manipulated. Compare it with external evidence rather than treating it as a verdict.

Evaluate the event, not only the file

If the clip claims to show a public event, ask whether the weather, light, buildings, clothing, language, transit, shadows, and public schedule fit the claimed time and place. Look for credible local reporting and independent witnesses.

If it depicts a known person, compare the claim with verified statements and public schedules. Do not contact or expose a private person merely to satisfy online curiosity.

For violence, elections, public safety, health, financial markets, or another high-stakes subject, preserve the file and escalate to a qualified newsroom, forensic laboratory, platform, or relevant authority. Do not publish a confident accusation based on a consumer detector.

Communicate the result precisely

"I found no watermark" is different from "this is real." "A detector scored it at 80 percent" is different from "the file was generated." "The credential validates" is different from "the event happened."

A useful verification note states what was examined, which tools and versions were used, what evidence supported or contradicted the claim, which sources were unavailable, and how confident the conclusion should be.

The final wording can be simple: the earliest credible source supports the file's claimed origin; the file contains a valid credential from a named signer; a supported watermark indicates generation by a named system; the context contradicts the caption; or the available evidence is insufficient.

The practical rule

Treat appearance as a lead. Treat provenance and watermarks as bounded evidence. Treat the event as a separate claim.

When the consequence of being wrong is high, slow down, preserve the record, and say uncertain when the record is uncertain.

AI assisted with research organization, structure, drafting, and validation. Dalton Anderson remains the attributed author and final editorial authority. The transcript and linked public sources control factual claims. Publication remains unauthorized.