A video on X can look convincing and still be synthetic, altered, mislabeled, or detached from its original context. Before reposting a surprising clip, inspect the post and account, find the earliest accessible version, look for independent confirmation, review provenance when available, and use an AI video detector only as a screening aid.
This checklist is for ordinary users, creators, researchers, and editors deciding whether a public X video deserves more verification. No label, visual trick, Community Note, or detector score can prove authenticity by itself, and none should be used as the sole basis for accusing a person.
Quick answer
- Preserve the direct post URL, account, caption, time, visible labels, and full clip.
- Identify the exact claim: who, what, where, and when does the post say the video shows?
- Look for an AI disclosure, platform label, correction, or Community Note, without treating its absence as proof.
- Search for the earliest and highest-quality version, not just the most viral repost.
- Compare the claim with reliable reporting, official statements, or independently recorded footage.
- Check several continuous moments while accounting for compression, filters, subtitles, dubbing, and screen recording.
- Run a cautious X / Twitter AI video check, then read the result state, confidence, evidence, and limitations together.
Separate the three questions people mix together
| Question | Evidence that helps | What it cannot settle alone |
|---|---|---|
| Was the picture or audio generated or materially altered? | Source file, provenance, disclosure, frame and audio review, cautious detector output | Who created it, why it was made, or whether the caption is true |
| Is this the original version? | Earliest post, longer cut, intact audio, original account, file history | That the earliest version found is necessarily authentic |
| Is the event described accurately? | Independent reporting, official records, witnesses, other angles | Whether every frame is synthetic or authentic |
An authentic video can be given a false date, location, subtitle, or identity. A generated clip can be clearly disclosed as satire or illustration. Keep media authenticity, source authenticity, and factual accuracy separate throughout the review.
1. Preserve the post before it changes
Record the direct URL, display name and handle, post text, timestamp, quoted post, visible labels, duration, and replies that contain source information. If you are permitted to retain a copy, keep the most direct version available and avoid editing it. Screenshots document context but do not replace the moving video and audio.
Note access limits. A post may be deleted, protected, age-gated, region-restricted, login-only, or visible in the app while its media cannot be fetched by an external checker. A screen recording may introduce a new frame rate, crop, notification, color conversion, or audio path. Those changes can create false clues and hide genuine ones.
For a consequential investigation, write down what was accessible, when you checked it, what you preserved, and what remained unavailable. This prevents a later conclusion from appearing more certain than the input justified.
2. Translate the post into a checkable claim
Viral captions often rely on vagueness: “this happened today,” “they do not want you to see this,” or “a famous person finally admitted it.” Rewrite the claim as a sentence with concrete details. For example: “The post claims that this named person made this statement at this event in this location on this date.”
Then ask what evidence should exist if the claim is true. A public speech may have a full recording, transcript, schedule, press pool, venue feed, or reporting from multiple outlets. A weather event may have local reporting and other footage. A product demonstration may have documentation, release notes, or independent tests. This source-first approach is usually more useful than staring at one odd frame.
Search a distinctive phrase from the caption, a spoken sentence, the claimed event, and the person or organization involved. Search for corrections and fact checks too. Do not count dozens of reposts of the same clip as independent confirmation.
3. Read X labels and context in proportion
Check for creator disclosure, an AI-related label, a Community Note, quoted-post context, and corrections in replies. X’s authenticity policy covers synthetic, manipulated, and out-of-context media that may deceive people or cause serious harm, and describes added, removed, or altered visual and audio information as relevant forms of manipulation. Read X Help: Authenticity.
A label or note is useful context, not a complete forensic opinion. It may address the caption rather than the pixels, or identify an older source without determining how the current file was produced. The absence of a note or label does not authenticate the video. Likewise, a verified profile badge or familiar avatar does not prove that a particular post is accurate.
4. Find the earliest, least-damaged version
Search for a longer upload and the person or organization closest to the event. Compare runtimes, aspect ratios, crops, subtitles, soundtrack, watermarks, and the moments immediately before and after the viral excerpt. A longer version can reveal an edit, joke, translation, correction, reenactment, or synthetic disclosure removed by a repost.
Look beyond X when appropriate. The source may be a livestream, press conference, creator page, news archive, or another platform. Reverse image search on clear keyframes can help locate earlier uses, although a failed match proves nothing. Earlier does not automatically mean authentic; it simply tends to preserve more context and usable media information. The link-check guide explains why a playable post may still provide limited media to an automated check.
5. Check provenance when it exists
Some media carries Content Credentials or other provenance information describing who signed it and which edits were recorded. Read the signer, time, tool, and edit chain rather than looking only for a badge. The C2PA explainer makes an important distinction: provenance can describe an asset’s origin and history, but it does not decide whether its real-world claim is true. See C2PA: Content Credentials explainer.
Missing credentials are not proof of AI generation. Many cameras and workflows do not attach them, and screenshots, crops, downloads, or exports can omit them. Present credentials as one evidence layer, not a universal truth label.
6. Review motion, frames, audio, and editing artifacts
Watch at normal speed first, then inspect several continuous moments around the strongest claim. Look for patterns that persist across time: faces or hands changing structure, reflections or shadows failing to follow a scene, text morphing between frames, accessories merging during movement, audio changing room tone abruptly, or impossible continuity between cuts.
Treat every observation as a lead. Beauty filters, stabilization, rolling shutter, motion blur, low light, sharpening, background replacement, translation, dubbing, frame interpolation, and ordinary editing can produce unusual results. A screen recording or compressed repost can smear detail and make authentic motion look unstable.
NIST recommends considering provenance, capture and editing history, compression or storage, whether the activity can be clearly seen or heard, other available angles, and the source’s purpose. That broader checklist is safer than a single “deepfake giveaway.” See NIST: Examining Digital Media.
7. Use an AI video detector as a screening step
First-pass check
Check an accessible public X post, then read the limits.
If the post is protected, removed, login-only, blocked, or contains no retrievable video, a responsible result should say the check is limited instead of inventing confidence.
Check an X / Twitter video- Result state: did the system obtain enough usable media?
- Confidence: how reliable is this scan given duration, quality, access, and compression?
- Likelihood: how did the checked signals point?
- Checked and unchecked signals: which areas were reviewed, and which were unavailable?
- Limitations and next steps: what should a person verify outside the tool?
Do not share a cropped score without those fields. A high likelihood is not proof of authorship, intent, or policy violation. A low likelihood does not prove the video is genuine or that its caption is accurate. Use the evidence report guide and deepfake screening guide for higher-risk cases.
8. Match the response to the stakes
For a harmless meme, the sensible result may be simply not to repost. For a claim about a private person, election, conflict, emergency, investment, medical advice, or public official, verification needs to be stronger. Look for primary records and independent reporting, contact the relevant organization where appropriate, and involve an experienced fact-checker or qualified forensic examiner before making an accusation.
Avoid amplifying the questionable clip while asking for help. Share the URL privately with a reviewer or describe the claim without reuploading the media. If intimate or exploitative synthetic content may be involved, prioritize the affected person’s safety and use the platform’s reporting route instead of conducting a public guessing game.
Common mistakes
- Calling one distorted frame proof, even though pause artifacts and compression are common.
- Assuming no label means real or that a Community Note is equivalent to media forensics.
- Counting reposts as corroboration when they all repeat one unsupported source.
- Uploading private media without authority.
- Reporting only a detector percentage while hiding its confidence and limits.
- Ignoring a false caption because the underlying video appears authentic.
Frequently asked questions
Can X tell me automatically whether a video was made with AI?
X may display labels or contextual information for some synthetic or manipulated media, but coverage is not universal. Treat a label as useful evidence and the absence of one as unknown, not authentic.
Can I check an X video from its link?
You can try a direct public post. Access can still fail when the post is protected, deleted, restricted, login-only, or the media is not retrievable. A limited result is not a pass.
Does an odd face or lip-sync error prove a deepfake?
No. Generation, dubbing, translation, editing, compression, and playback problems can overlap visually. Look for persistent patterns, source evidence, and corroboration.
Is a low detector score proof that the post is real?
No. It means only that the checked input did not produce strong detected signals under that specific analysis. Short clips, poor quality, inaccessible audio, and heavy recompression can reduce useful evidence.
Bottom line
To check whether an X video may be AI-generated, begin with the post’s exact claim and source—not a detector score. Preserve context, locate the earliest accessible version, seek independent confirmation, review provenance and continuous motion, account for repost damage, and keep uncertainty visible. When the stakes are high, uncertainty is a reason to pause and escalate, not a gap to fill with certainty.
Sources and update notes
Reviewed on September 29, 2026. Platform labels, policies, and media access can change; verify official resources before relying on exact interface behavior.