How to tell if a video is real: deepfakes, edits, and what body language can and can't tell you

Last Updated: September 22, 2026

"Is this video real?" usually bundles together three different questions: is the footage itself generated or altered, is it being shown in a misleading context, and is the person in it telling the truth. Each question needs a different kind of check. This guide walks through all three, honestly, with steps you can actually do.

1. Check the source before you check anything else

Most videos that turn out to be "fake" were never manipulated at all — they're real footage that got mislabeled, recut, or stripped of its original context. This is also the fastest check, so do it first.

• Reverse-search a frame. Screenshot a clear frame (a face, a distinctive background, a logo) and run it through Google Lens, TinEye, or a similar reverse image search. This often surfaces the original upload, sometimes years older than the version you're looking at.
• Find the original account or outlet. Search the video's likely keywords plus the event, date, or location. If it's real news, a wire service or major outlet almost always has the same footage, usually with more context attached.
• Check the upload account's history. A account created days ago with no other posts, or one that only reposts inflammatory clips, is a weaker source than an established outlet or a verifiable eyewitness.
• Look for a date and place. Weather, shadows, signage, license plates, and building details can confirm or contradict a claimed date and location — this is the basic method professional fact-checkers use, and it works without any special tools.
• Cross-check with fact-checking organizations. If a clip is going viral, search its claim alongside "fact check" — outlets like Reuters Fact Check, AFP Fact Check, and Snopes frequently get there first.

2. Look for signs the footage itself was generated or altered

If the source checks above don't resolve it and you suspect the video itself — not just its framing — was manufactured, look for these, while knowing that none of them is conclusive on its own:

• Audio-lip mismatch. Watch consonants closely (P, B, M sounds require the lips to close). Generated or dubbed video often drifts out of sync in ways a real recording doesn't.
• Lighting and shadow inconsistency. Does the light on the face match the light in the rest of the scene? Mismatched shadows are one of the harder artifacts to fully erase.
• Edge and detail artifacts. Warping or blurring around hairlines, ears, glasses, or where the face meets the neck — especially in older or lower-effort generations.
• Unnatural texture or blinking. Skin that looks unusually smooth or waxy, or blink patterns that look mechanical — but treat this one skeptically, since it's the first artifact newer generation models fix, and normal video compression can also smooth skin texture.
• Consistency across the whole clip. Does the person look like the same person a minute in as they did at the start? Short, cherry-picked "gotcha" clips are easier to fake convincingly than a full, continuous conversation.

None of these are things the naked eye can apply reliably at scale, and manipulation techniques keep improving. For anything consequential — a video you might act on, republish, or cite — the more trustworthy approach is a dedicated deepfake or synthetic-media detector, a category of forensic tool built specifically to analyze pixel-level and compression artifacts in the footage itself. That is a different job from behavioral analysis, and it's the right tool when the question is "was this footage generated," not "how is this person behaving."

3. Watch for selective editing and out-of-context clips

A clip can be 100% unaltered footage of a real event and still be deeply misleading. This is the most common way people are fooled, and it requires no fakery at all.

• Watch the full, unedited version if you can find it. A 10-second clip cut from a 40-minute press conference or interview can reverse the apparent meaning of what was said.
• Check for jump cuts. A sudden change in clothing, lighting, or background mid-sentence signals that footage from different moments was stitched together.
• Read the caption separately from the footage. Ask whether the caption's claim is actually shown in the video, or just asserted alongside it. A real clip with a false caption is extremely common and needs no editing skill at all.
• Check what's just outside the frame. Tight cropping can remove a sign, a second speaker, a laughing crowd, or anything else that would change how the moment reads.

4. What a person's own behavior can — and can't — tell you

Suppose the footage checks out: it's real, unaltered, and shown in context. A different question sometimes follows: does the way the person is speaking tell you anything? This is where behavioral signal analysis, including RedFlag, fits — and where its limits matter most.

Facial and vocal signals during speech are real and measurable: blink rate, gaze stability, head stability, brow furrow, jaw tension, smile authenticity, voice volume, pitch, speaking rate, filler words, and more. RedFlag tracks 13 such signals from a public YouTube video, entirely in your browser, and computes a binary verdict per moment — 🟢 Full Harmony or 🔴 Red Flag — from three of them: pitch monotony, person-relative blink elevation, and voiced-continuity elevation, compared against that speaker's own normal range over the video (episodes shorter than 3 seconds are dropped as noise). The full method, including every threshold, is published at /methodology/.

What a divergence does not mean: stress, discomfort, or unusual delivery have dozens of innocent causes that have nothing to do with honesty — nervousness about being on camera, a difficult question, illness, fatigue, unfamiliar language, cultural differences in expressiveness, poor lighting, or simply an unusual moment in an otherwise normal conversation. No peer-reviewed body of research supports reading truthfulness from body language or voice — this is also why polygraphs are not admissible as proof of truth in most courts. A red flag from RedFlag, or from your own eyes, is a description of a moment, not a verdict on the person.

So: behavioral analysis cannot tell you whether a video is authentic (that's §1–2), and it cannot tell you whether someone is lying (nothing can). What it can do is show you, precisely and consistently, where a real person's own delivery deviated from their own baseline — a starting point for your own judgment, not a substitute for it.

The checklist

1. Reverse-search a frame and find the original source before doing anything else.
2. Check the uploading account's history and the video's actual upload date.
3. Search the claim alongside "fact check" against known fact-checking outlets.
4. If you suspect the footage was generated, run it through a dedicated deepfake/synthetic-media detector rather than eyeballing it.
5. Watch the full, unedited version — check for jump cuts, mismatched audio, and what the caption claims versus what's actually shown.
6. Only once the footage is confirmed authentic and in context, consider what the person's delivery does or doesn't show — and treat any divergence as a question, never an answer.

If a video is authentic and you're curious how someone delivered their answers — a debate moment, an interview, a statement — RedFlag charts the signals and lets you look for yourself. Related reading: deepfake detection vs. behavioral analysis and the 54% test — on why chance-level lie detection is the honest baseline for any method, including this one.