Why AI-generated clips look AI-generated (and what actually fixes it)

There's a specific feeling people get watching a clip that was cut by software. Nobody can usually name it. They just scroll. The clip had a good moment in it, the captions were on screen, the aspect ratio was right, and it still felt off in a way that made them leave.

That feeling is almost never the AI picking the wrong moment. It's a handful of small, boring, mechanical defects around the edges of an otherwise fine clip. Every one of them is fixable, and most of them you can check for in about sixty seconds. Here's the full list, what causes each one, and what to do about it.

1. The clip ends mid-thought

This is the single most common one, and the most damaging, because it breaks the last thing the viewer sees. A clip that stops halfway through a sentence reads as broken software, not as an edit.

The cause is mechanical. A language model picks a start and end timestamp from a transcript, and nothing checks that those timestamps land between sentences rather than inside one. Transcript segments are the only reliable record of where sentences actually begin and end, so if the tool doesn't snap its cut points to those boundaries, it will sometimes cut mid-word. We shipped exactly this, and had to add a step that pushes each cut point out to the nearest real segment boundary.

There's a second, subtler version that survives that fix: the clip ends on a grammatically complete sentence that still leaves the viewer hanging. The one that taught us this ended on "...are you?". That's a perfectly valid sentence boundary. It's also a question, and the answer (the entire reason the moment was interesting) was in the next four seconds. A good ending is a resolution: the answer, the conclusion, the punchline. Not the setup for one.

How to check: watch only the last three seconds of every clip before you post it. If the final line is a question, a list that hasn't finished, or ends on a word like "and", "but" or "because", extend it.

2. Captions that drift out of sync

Slightly-wrong captions are worse than no captions. Viewers don't consciously register a 400ms offset, but they do register that something feels cheap.

Drift has one cause almost every time: captions were timed against the original video, and the clip was cut from the middle of it, so every caption is off by however far into the source the clip started. For a single continuous clip this is easy, just subtract the start time. It gets genuinely tricky when a tool stitches several separate pieces of a video together, because then each piece needs its own offset and a single global one silently pushes the captions further out of sync at every cut. If you're using a tool that joins segments into one recap video, this is the first thing to check.

How to check: watch the last ten seconds, not the first. Drift accumulates, so the end of a clip is where it's visible. If the start is fine and the end is off, that's an offset bug, not a transcription problem.

3. Captions burned in from a transcript that was never real

Speech-to-text models do not return "nothing" when there's nothing to transcribe. They hallucinate. Run a video with music, engine noise, crowd noise, or plain silence through Whisper and you'll frequently get back short, confident, completely invented lines. The classics are "Thank you.", "Bye.", and "I don't know."

We got a real example of this recently: a 100-second video whose entire transcript was ten words, spread as "I" at the start, "Thank you." stretched across thirty seconds of silence, two "I don't know."s, and "Bye." at the end. Nothing in that is real. If a tool burns those in as captions, you get a video with confident subtitles for words nobody said.

The fix is a threshold rather than a clever model. Real conversational speech runs roughly 120 to 150 words per minute. Anything under about 30 is not a transcript, it's an artifact, and the honest response is to skip captions entirely rather than print noise over someone's video.

How to check: if your source has long stretches without speech, read the captions in those stretches specifically. That's where invented text hides.

4. The crop that cuts the subject in half

Turning a 16:9 video into 9:16 throws away about two thirds of the frame's width. Most tools centre-crop, because it's fast and it's right often enough. It is badly wrong in three specific cases: two people sitting apart, a speaker standing off to one side, and anything where the interesting thing is spatial (a race, a match, a screen share, a whiteboard).

This one is worth being clear-eyed about, including about our own tool: a centre crop has no idea who is speaking. If your framing puts the subject off-centre, a centre crop will cut them off, and no amount of good moment-selection saves that clip. The realistic answers are to reframe your source before uploading, or export at the original aspect ratio for content where the width genuinely carries the meaning.

How to check: scrub the vertical export, don't just look at the thumbnail. Crop failures are usually intermittent, appearing only when someone moves.

5. It opens on a pronoun

"...and that's exactly why he did it." Who is he? What is it? The viewer has no idea, and unlike the mid-word cut, this one isn't obviously a bug, so it survives review and gets posted.

A clip has to work for someone who heard nothing before it, which is a much harder constraint than "this was an interesting thirty seconds". A moment that depends on context established two minutes earlier is a bad clip even though it was a good moment. Starting one sentence earlier usually fixes it completely, at the cost of three seconds.

How to check: read the first sentence alone, out of context. If it contains "he", "she", "they", "that", "this" or "it" referring to something unexplained, move the start back.

6. Every platform got the same caption

The same text posted to TikTok, Reels and Shorts reads as automated on at least two of them, because the three platforms genuinely have different conventions. Hashtag counts differ. Tone differs. A YouTube Shorts description written like a TikTok caption looks out of place, and a TikTok caption written like a YouTube description looks like a press release.

This one is cheap to get right, since generating three variants costs a fraction of what transcribing the video did. More detail on what actually differs is in our breakdown of the three platforms.

7. The clip is padded to hit a number

A tool told to produce thirty-second clips will produce thirty-second clips, whether or not there were thirty seconds worth keeping. What you get is twelve good seconds and eighteen seconds of throat-clearing, and the throat-clearing is what determines your completion rate.

Length should be an output, not a target. If the moment is fifteen seconds, the clip is fifteen seconds. This is one of the few places where trimming a clip yourself after the fact reliably beats anything automated, and it takes about ten seconds per clip.

8. B-roll that doesn't match the words

Automatic stock footage fails in a recognisable way: someone says "we grew fast" and a generic shot of a rising line chart appears. It's technically related and adds nothing, and viewers read it as filler immediately.

Real editorial b-roll is sparing and literal. It shows a specific thing that was specifically named, and most clips should get zero cutaways rather than one. If a tool inserts b-roll into every clip, turn it off. The version with no cutaways is better than the version with decorative ones.

9. The watermark is doing too much

A large, animated, or centrally-placed watermark tells every viewer that this was made by software before they've decided whether they like the content. It also makes you measurably less likely to post the clip at all, which is worth naming plainly since we put a watermark on our own free tier: a loud one defeats its own purpose, because unposted clips advertise nothing.

If your tool's free tier brands exports, look at where and how big. Small and cornered is a fair trade. Anything you notice before you notice the content is not.

The sixty-second check before you post

You don't need to watch every clip end to end. Four checks catch nearly all of the above:

• Read the first sentence alone. Does it make sense to someone who heard nothing before it?
• Watch the last three seconds. Does it resolve, or does it just stop?
• Watch the last ten seconds with sound on. Are the captions still in sync at the end?
• Scrub the middle. Is anyone's head cut off by the crop?

Clips that pass those four don't read as machine-made, regardless of what made them.

The honest summary

"AI clips look bad" is usually the wrong diagnosis. The moment selection in most current tools is genuinely decent. What's actually wrong is boundary handling, caption timing, and a refusal to output nothing when there's nothing worth outputting. Those are engineering problems with correct answers, not limitations of the approach.

Which is also the useful thing to take into a tool comparison: don't evaluate on whether the AI found good moments, because most of them will. Upload something hard (a quiet section, an off-centre speaker, a source with music over the top) and look at the edges of what comes back. That's where tools actually differ. If you're weighing up options, our criteria-first comparison covers the rest, and how moment detection actually works covers what's going on underneath.