How to Tell If a Video Is AI

Set yourself up to look properly

Find the best copy of the video you can, ideally the earliest and largest. Watch it once in full for the story, then again on a bigger screen with the sound off to look for visual problems. Use the pause key, the frame-step keys and the playback speed menu where your player offers them. Most of the clues below live in a fraction of a second.

If you want a quicker, more casual version, the page Is This Video AI? condenses the main signs.

Unnatural motion

Bodies carry weight. Footsteps land, shoulders shift, hair lags behind a turn. Generated clips often get the broad motion right but lose the secondary motion: the drag, the follow-through, the small compensations. Look for people who glide, limbs that move without effort, animals whose gait does not match their speed, or a camera that drifts in a way no hand or drone would produce. Also notice rhythm. Movement that repeats in an eerie loop or runs at a perfectly constant pace is worth noting.

Caveat: slow motion, stabilisation and frame interpolation applied by ordinary editing apps make real footage look floaty too.

Object permanence and background transformations

Real objects do not vanish when something passes in front of them. Watch what happens to things that get occluded: a cup behind an arm, a bicycle behind a pillar, a person walking behind another. In generated video the object may reappear changed in colour or shape, or may not return at all. Likewise scan the background across the whole clip. Signs of trouble include windows that rearrange, a crowd whose members merge, a street that adds a lamp post, or a logo that morphs while the camera moves.

Short clips are easier to keep consistent than long ones, so a single continuous shot that runs for many seconds under complex motion tells you more than a rapid montage.

Hands, fingers and fine details

Hands are a classic weak point because they are small, mobile and structurally complicated. Count fingers across several frames and notice joints that bend the wrong way, fingers that fuse or pass through objects, and grips that do not touch what they hold. Extend the same scrutiny to teeth, ears, glasses, watches, earrings and buttons.

Better generators have reduced these errors, so a clean hand tells you little. A bad hand is a stronger clue, though even then, motion blur and compression can mimic it.

Text instability

Writing needs exact shapes that stay the same over time. Pause on signs, shop fronts, shirts, license plates, screens and book covers. Gibberish that resembles letters, words that change spelling from frame to frame, or numbers that shuffle are suggestive. Check text at the start and end of a camera move, since it often holds in one pose and degrades in another. Overlaid captions and watermarks added later are not part of this test, because they were typed on top.

Reflections, shadows and lighting

Light follows rules. A single sun gives parallel shadows, and shadows soften or sharpen with distance in predictable ways. Mirrors, windows, sunglasses and puddles should show what is in front of them, from the right angle. Watch whether highlights on a face move as the head turns, whether a subject cast a shadow at all, and whether indoor and outdoor light mix sensibly. Generated scenes can look attractive but quietly break these relationships, particularly in reflections. Real scenes with colourful stage lighting can also confuse the eye, so judge by consistency rather than mood.

Physics

Ask whether things behave as they would with weight, speed and friction. Liquids should pour, splash and settle. Smoke and fire should evolve. Fabric should respond to movement and air. Objects should collide, bounce and stop rather than blend into one another. Items that hang in the air, change size as they move, or pass through a surface are red flags. Intentional visual effects and CGI also break physics, which is why a physics problem points to synthetic or edited, not necessarily to AI.

Temporal consistency

Pick an anchor, such as a tattoo, a button or a pattern on a shirt, and follow it from one end of the clip to the other. Does it keep its size, shape and place on the body? Does the person's face keep the same proportions? Does the colour grade stay steady unless the lighting changed? Slow flicker, shimmering textures and details that crawl across a surface are all temporal artifacts. For clips featuring someone's face, the deepfake detector page covers face-specific temporal checks.

Lip synchronization

When someone speaks, check that lip shapes match the sounds. Sounds like p, b and m require the lips to close. Watch the teeth and tongue too. Mouth movement that is mushy, that lags a beat behind the audio, or does not change with the pace of the speech can indicate generated or re-voiced footage.

This check is difficult to perform and easy to over-trust. Dubbed videos, bad streaming sync and low frame rates all create mismatches in perfectly real material. Treat it as one clue, and do not take a clean match as evidence of authenticity.

Audio clues

Close your eyes and listen. Natural speech has breaths, small noises, changes in volume and the acoustics of a room. Synthetic speech can sound evenly paced, strangely clean, or emotionally flat, and sometimes the background hum switches abruptly. Check that footsteps, impacts and ambient sound line up with what is on screen and that echo fits the space. Sound can also be added to any video afterwards, so music and a convincing soundtrack say nothing about whether the visuals are real.

Metadata and content credentials

Some files carry camera details, editing history or signed content credentials that record creation and edits. A verified credential can help establish a file's history, but it does not prove that a scene happened as depicted. Ordinary metadata can be altered. Many social platforms strip or rewrite it, and screenshots or screen recordings generally do not preserve the original file's credentials. Absence is common and is not evidence of AI generation. Check the provenance when available, alongside the source and the media itself.

The same limits apply to still images, which are discussed on the AI image detector page.

Why social-media compression complicates everything

Many platforms recompress uploads, and re-sharing can introduce further processing. Compression flattens fine texture, creates blocky patches around motion and softens text. Those same effects hide some generation artifacts and create fake-looking ones. A third-hand repost in low resolution calls for caution rather than a strong conclusion from visual inspection. Look for a better copy from the first poster, the creator's own channel or a report that credits the source.

Why looking is not enough

AI generation is improving rapidly. Flaws that were common last year are less common now, and creators often cherry-pick the best short outputs or clean them up in an editor. Your eyes are also easily influenced by what you want to believe. For all of those reasons, manual inspection is a way to raise or lower your suspicion, not to certify a clip.

An automated estimate adds a different kind of evidence, because it can read statistical patterns that humans do not notice. It has limits of its own, which are explained on the AI video detector page. Together with source checking, though, it gives you a more grounded view than either approach alone.

A simple order of operations

First, find the origin and best copy. Second, inspect motion, details, light and sound as described above. Third, run it through Drubl. Fourth, weigh the pieces together and decide how confident you actually are. If the question matters, wait for confirmation from a reliable source before you share or act. If you do share, describe your uncertainty plainly.