PricingLog in

How to check if an image, video or audio is AI-generated.

Reviewed

Short answer

Look for a declaration before you judge the pixels: read the file’s Content Credentials (C2PA), check for an invisible watermark with its maker’s own detector, look for the platform’s AI label and run a reverse image search for earlier copies. Each check can only confirm what it finds: the C2PA explainer says provenance data can be removed and cannot by itself tell you whether content is true, and Meta writes that “there are ways that people can strip out invisible markers.” A file with no mark is therefore not proven to be made by a person.

Official sources: C2PA Explainer 2.2 · Meta, February 2024

Five checks, strongest first

Ways to check a file for AI, what each shows and where it stops
CheckWhat it can showHow to run itWhere it stops
Content Credentials (C2PA)A signed record in the file: the tool that made it, an AI declaration, who signed it and later editsOur Content Credentials checker reads it in your browser; the Content Authenticity Initiative’s Verify site shows the same recordThe C2PA explainer answers “Can the provenance metadata be removed?” with “Yes it can.”
Invisible watermarkWhether a specific maker’s mark is hidden in the pixels or the soundGoogle: upload the file in Gemini, Search or Chrome and ask if Google AI made it, or use the SynthID Detector; Proofcan marks: VerifyA detector finds only the watermark it was built for (the SynthID Detector scans for a SynthID watermark); Adobe’s open-source TrustMark code can encode, decode and also remove its watermark
Platform labelThat the platform found an AI signal (C2PA or IPTC data, its own systems) or the uploader declared AILook for the AI label on the post or player, or in the expanded description or menuLinkedIn: “it’s not yet possible to identify and label all AI-generated and modified content”
Reverse image searchEarlier copies, the original context and similar imagesGoogle Lens: results include “Websites with the image or a similar image”Finds copies, not how a picture was made; a new image has no earlier copies
Looking and listeningTypical artefacts: seams around a swapped face, blurred teeth or eyes, odd lighting; metallic, monotone or mispronounced voicesWatch at full size and slow speed; listen for emphasis and pronunciationGood fakes show none of them, and automated detectors often work reliably only under certain conditions (BSI)

Step by step

  1. Get the original file where you can. A screenshot is a new image and carries none of the original’s metadata, so download the file itself or ask the sender for it.
  2. Read the Content Credentials in the checker. A valid manifest names the signer and the generating tool and may declare AI generation; a broken or missing one tells you nothing either way.
  3. Check for watermarks with the maker’s tool: Gemini, Search or Chrome for Google’s SynthID, Verify for a Proofcan mark and its public record.
  4. Look where the content was posted: YouTube, Facebook, Instagram, Threads, TikTok, LinkedIn and Pinterest show AI labels, partly added automatically, partly declared by the uploader.
  5. Run a reverse image search with Google Lens and compare dates and sources of the earliest copies you find.
  6. Then look and listen for artefacts, and weigh who published the content, when and why.

What the checks cannot tell you

  • That a file without a mark was made by a person. Meta wrote in February 2024 that “it’s not yet possible to identify all AI-generated content”; marks can be lost along the way or removed on purpose.
  • That content with valid credentials is true. The C2PA explainer: “provenance information alone cannot tell you whether the digital content is true, accurate or factual.”
  • A reliable verdict from an “AI detector”. The BSI reports that in the Deepfake Detection Challenge (2020) even the best model reached an average accuracy of 65.18 per cent, where guessing gives 50; OpenAI withdrew its own AI text classifier on 20 July 2023 “due to its low rate of accuracy”.
  • Anything about other makers’ content from a maker’s detector. ElevenLabs says its speech classifier “cannot detect audio from other providers” and does not reliably classify audio from its Eleven v3 model.

Signals by type of content

Which signals the makers’ pages describe, by type of content
ContentSignals you can checkNote
ImagesC2PA Content Credentials; SynthID (Google and partners); TrustMark (open source); platform labelsReverse image search works on images; for a video, search with a still frame
VideoC2PA Content Credentials; SynthID; platform labels such as YouTube’s, which is added automatically when a video contains C2PA metadataMeta wrote in 2024 that AI video and audio tools did not yet add such signals at the scale of image generators, so it could not yet detect and label that content from other companies
Audio and voiceSynthID in Google’s audio; vendors’ own classifiers, such as ElevenLabs’ for ElevenLabs audioThe BSI lists metallic sound, mispronunciation and monotone speech as typical artefacts
TextSynthID text watermarks in Google’s Gemini app and web experienceOpenAI’s own text classifier found 26 % of AI-written text and flagged 9 % of human-written text as AI in its test set

Questions

Is there a tool that tells me for sure whether an image is AI?
No. Tools can confirm a mark they find, such as a C2PA declaration or a SynthID or TrustMark watermark, but none can prove that an unmarked image was made by a person, and detectors that judge the pixels make mistakes.
Why does an AI image sometimes have no Content Credentials?
Not every tool adds them, and they can be removed, for example by a screenshot, by converting the file or by a platform that strips metadata. The C2PA standard therefore also describes watermarks and fingerprints that can lead back to stored credentials.
Can I check a video or a voice message?
Partly. Videos and audio from Google’s AI can be checked for SynthID in Gemini, and a video with C2PA data can be read like an image; for other voice recordings you depend on the vendor’s own classifier, where one exists, and on listening for artefacts.

Sources

Last reviewed by Volkan Çatak, Founder, Proofcan.

Not legal advice. This page explains official texts in plain words and links each one; check the sources, or a lawyer, before relying on it for a decision.