AI video: how it is made and what still gives it away

AI video is made much the same way as an AI image, with one expensive addition: the model has to keep a scene consistent while it moves. A text to video system starts from random static and removes it in stages, but it denoises a whole block of frames together rather than a single picture, so that a face stays the same face and a car keeps traveling in one direction. That extra requirement explains almost everything about these clips: why they are short, why they cost far more than stills, and why what goes wrong now goes wrong in time rather than in space. A bag changes shape after passing behind a pole. A hand gains a finger mid gesture. Feet slide across the pavement instead of pushing against it.

How a video model builds a clip

If you already know how diffusion turns noise into a picture, you have most of this. A compression network squeezes the video down into a much smaller numerical grid, the model repeatedly predicts and subtracts noise from that grid while being steered by your prompt, and a decoder expands the finished result back into pixels.

What changes is what the grid represents. A still image model works on patches of one frame. A video model works on patches that have a duration as well as a position, small cubes of picture covering a handful of frames each, often called spacetime patches. The model looks at all of them together, which is how a decision about the first second reaches the last: when it judges what counts as noise in frame forty, it can see what it has already settled about frames one and eighty.

Three practical consequences follow from that design.

• The clip is resolved all at once, not played forward. The model is not simulating the next moment from the last one, it pulls the whole span out of static together. That is why you cannot usually add one more second without regenerating everything.
• Cost climbs steeply with length. The attention step scales roughly with the square of the number of patches, so doubling the duration more than doubles the work. Hence clips measured in seconds, with longer pieces assembled from several generations.
• The same engine runs from a starting image. Image to video adds noise to a still you supply but stops partway, then denoises forward. Animating a photograph you already have is cheaper and far more controllable than conjuring a scene from words.

Why time is the hard part

There is no scene inside the model. No three dimensional geometry, no list of objects with identities, no physics engine tracking momentum or contact. What the model learned is which pixel patterns tend to follow which, across an enormous quantity of real footage. That carries it a surprising distance, because most motion in real video is smooth and repetitive. It fails in specific, predictable places.

Occlusion is the classic one. Someone walks behind a lamp post. Nothing records that a striped shirt and a shoulder bag went in, so what emerges is whatever looks locally plausible: the stripes shift, the bag thins, a hand turns over. The same weakness appears whenever an object leaves the frame and returns.

Identity drift is the second. Over a long enough clip a face slowly becomes a slightly different face, a logo re-letters itself, the furniture in a room rearranges. Systems fight this with reference conditioning, feeding in a portrait alongside the prompt to anchor the subject, and it is much better than it was. It still degrades the longer the clip runs.

Contact physics is the third. Feet on the ground, fingers closing on a cup, cloth folding against a body, liquid leaving a bottle: all of these involve forces the model never represents, so it approximates them from appearance alone. You get the foot that slides a few centimeters, the fingertip that sinks into the handle, the fabric that moves as though gravity were somewhere else.

The artifacts that still give it away

Watch the clip several times, looking at one thing on each pass, and step through it slowly if your player allows. These failures live between frames, and normal speed playback hides most of them.

The same reasoning underlies checking whether a still image was generated, except that a video hands you dozens of samples of one scene, so a disagreement between two moments is worth more than any single odd frame.

Artifact spotting is a shrinking advantage

Every giveaway that gets publicized becomes a target. Hands and teeth were the standing joke about generated pictures until enough people complained, at which point they were curated for and evaluated during training, and they improved. The same is happening to gait, hair edges and short bursts of on screen text. Any list of tells, including the one above, is a snapshot of where the weak points sit right now.

Compression works against you as well. By the time a clip reaches your phone it has been uploaded, re-encoded, cropped to a vertical frame, perhaps screen recorded and posted again. That pipeline eats exactly the detail you need: edge shimmer becomes blocky compression noise, and a six fingered hand is forty pixels across.

Then there is the awkward fact that a lot of harmful video is not fully generated. A face swap replaces one region of genuine footage: real camera, real room, real body, real physics, real lighting. Everything you were taught to inspect passes, because most of it was filmed. A real recording with a cloned voice laid over it has the same property, and that technique is covered in how a short audio sample becomes a fake voice. Cheapest of all, and still by far the most common, is authentic footage of a real event carrying a false caption. No generator involved.

Provenance is the durable answer

Because the pixels will stop telling you, the long term answer is a record of where a file came from. That is what Content Credentials are for: a signed manifest attached to the file recording what device or tool made it and what was done to it afterwards, which travels with the video if the platforms that handle it are willing to carry it. The details, including how to read one and how easily it can be stripped, are in how provenance labels on media work.

Some providers also embed an invisible watermark, a statistical pattern in the frames that survives moderate editing and is readable by their own checker. It is not something you can run yourself, and an open model can simply omit it. European transparency rules requiring synthetic content to be marked in machine readable form push the large providers this way, but none of it is retroactive and none of it covers everything.

Treat provenance as asymmetric evidence. A valid credential chain back to a camera is strong support. No credential at all means nothing.

How to check a clip that shocks you

1. Wait before you send it on. Nearly all the damage from a fake clip comes from people who reposted within a minute of first seeing it.
2. Find the earliest copy, not the one in front of you. Search a distinctive phrase from the caption, sort by oldest, and see whether the account posting it has any history.
3. Ask who else would have filmed this. A newsworthy event in a public place produces several angles from several phones. One angle, one account, no second source is the single loudest warning sign.
4. Screenshot a clear frame and reverse image search it. That often surfaces the original video, sometimes years old and from another country. Open the file's credential panel too, if the platform shows one.
5. Check details that have nothing to do with the pixels: the language on the signs, the license plates, the weather, whether the trees match the season being claimed.
6. Notice the emotional shape of it. Material engineered to spread is built to make you angry or frightened in the first two seconds, and the wider method for that is in how to check a viral claim before passing it on.

If someone offers a detector score as the answer, treat it as you would one for writing: the reasons those numbers mislead are in why a detector score is not evidence. Sourcing beats scoring, and it will keep beating it as generators improve.