Start with the claim and the label
Decide what you need to know: whether any AI was used, whether a specific event happened, or whether a real person said something. Then check the platform's disclosure. YouTube labels realistic altered or synthetic content, TikTok labels AI-generated content and reads Content Credentials, and Meta shows AI information on Facebook, Instagram and Threads. Labels help, but unlabeled videos can still contain AI-generated material.
Source: YouTube altered content disclosure (opens in a new tab)
Visible watermarks and metadata
Some video generators mark their output. OpenAI says every Sora video carries both visible and invisible provenance signals, including a visible watermark and C2PA metadata. A visible logo can be cropped or blurred, and metadata can be lost when a clip is downloaded, re-encoded or screen-recorded, so the absence of a mark is inconclusive.
Source: OpenAI: launching Sora responsibly (opens in a new tab)
Check SynthID and Content Credentials
Google says media made with its AI tools, including video, carries an invisible SynthID watermark, and the Gemini app can check an uploaded video, image or audio file when you ask “Is this made with AI?”. Gemini can also read C2PA Content Credentials. Each check only covers its supported signals; a negative result does not rule out other generators.
Source: Google: identifying AI-generated media (opens in a new tab)
Watch for visual and audio clues, carefully
AI video can show warping objects, inconsistent hands or text, unnatural physics, details that flicker between frames, or lip movements that drift from the audio. Real footage can also look odd because of compression, stabilization, low light or editing. Note anything unusual and use it as a reason to find the source rather than as a verdict.
Trace the source
Find the earliest upload you can, the account that posted it and any original publisher. For news events, compare the clip with reporting from the same time and place, and check whether the location, weather and other details match. A genuine recording can be re-captioned to describe a different event, so origin and context are separate questions.
What detectors can and cannot do with video
Video detectors usually sample frames, which can miss short generated segments, and few authenticate audio or speech. Compression on social platforms weakens the signals they rely on. Image Evidence does not identify people, judge whether a real person said something or offer face recognition. Treat any video score as one clue alongside labels, provenance and context.
Is there a free way to check if a video is AI-generated?
Yes, for some sources. You can upload a video to the Gemini app and ask whether it was made with AI, check platform labels, and inspect any Content Credentials with a C2PA validator. These checks only find supported signals, so combine them with source tracing.
How can you tell if a video is a deepfake?
Be cautious with any video that shows a real person saying or doing something surprising. Check whether the person or their organization has published the same clip, whether reputable outlets report it, and whether the platform has labeled it. Lip-sync and voice oddities can be clues but are not proof. Image Evidence does not analyze faces or judge whether a specific person is genuine.
Do AI-generated videos have watermarks?
Some do. According to OpenAI, Sora videos carry a visible watermark and C2PA metadata, and Google says its AI-generated video carries SynthID. Many other tools add nothing, and visible marks can be cropped. A missing watermark tells you little about how a video was made.