AI Can Now Edit a Finished Video. Here Is What That Changes

A knife comes down on a sushi roll on a slate board. In the second clip the same knife comes down, in the same place, on a steak. The water drops on the blade are still there, and nothing else in the frame has changed.

The pair is an example for FLUX Video Edit, a new tool from Black Forest Labs, published on its page at fal.ai, one of the platforms that sell access to it. The tool does not make a new video. It takes one that already exists and changes part of it on request. That shifts the question people have been learning to ask about online video, and this piece is about how.

What the tool actually does

You upload a clip and describe the change in a sentence. According to the company's product page, the tool can remove, add or replace objects and characters, rebuild the setting, rewrite on-screen text, change colours and materials, restyle the whole clip, and "change what happens."

One item on that list deserves more attention than the rest. The tool can also change or translate dialogue, with lip sync. A person in a clip can be made to say something they did not say, in their own scene, with their mouth moving to match.

The limits are modest. The documentation caps input at 15 seconds and returns video at up to 720p. The price is 3 cents per second of output, which puts the longest possible edit at about 45 cents.

Why an edit is harder to catch

A generated clip is invented from nothing, every pixel of it. That tends to leave traces spread across the whole frame: in the grain, in how light falls, in how motion carries from one frame to the next. Those traces are what most detection tools look for, ours included.

An edit to real footage leaves far less. The camera was real, the sensor noise is real and so is the shake of the hand holding it. Only the replaced object is new, and the tool had a real scene to match it against. The lighting was already there. So was the surface the shadow falls on.

What remains to catch is small: an edge that stays too clean, a texture that holds still while the camera moves. Those details tend to be the first thing lost when a social network recompresses an upload.

The example above shows the effect even without real footage. We ran both clips through our own detector. On the sushi, it leaned towards generated. On the steak, it leaned the other way, towards real. Both answers came with low confidence, but the direction matters: the edit made the clip look more like camera footage to our detector, not less.

What to do before you share a clip

Look for the original. An edited clip has a parent. If the same scene exists somewhere earlier without the change, that settles it. A reverse image search on a single frame is still the quickest tool most people have.

Look for a second angle. Real events with witnesses are usually filmed more than once. Two unrelated people posting the same moment from different positions is hard to fake and quick to check.

Check who posted it first. An account that has filmed the same street and the same kitchen for years is a different source from one created last month.

Be most careful with what people say. A clip of someone speaking can now be altered without touching anything else in the frame. If the words are the news, find the full recording.

For the signs human eyes can still catch in generated video, see how to spot AI-generated video. For how tools mark their output and how well those marks survive, see AI video watermarks explained.

For two years the useful question was whether a video was made by AI. It is turning into a narrower one: which part of this is true, and who is telling you so.