74% of AI videos land between 14 and 16 seconds. That's not a choice — it's a default

By Alpha Lay · 2026-09-26

We previously published this line: "86% of AI videos run between 8 and 16 seconds."

After writing it we felt the claim was too loose. 8 and 16 seconds differ by a factor of two; lumping them together says almost nothing. So we re-bucketed all 1,274 cases at 2-second intervals.

The result is far more extreme than the original sentence.

Where the data comes from

Duration is read from each video file's container metadata, not from the publication page — so it is an objective measurement. Buckets are 2 seconds wide.

Using integer intervals instead of buckets makes the shape clearer:

832 cases fall inside a window 0.2 seconds wide.

Finding 1: This is an output default, not a human choice

People do not collectively choose the number 15.1 seconds.

If duration were set by creators according to content needs, the distribution would spread with genre — comedy needs setup, ads need brevity, dance needs a full musical phrase. Instead the distribution is a needle: 65% of samples inside a 0.2-second band.

The only available conclusion is that the model has a fixed default output length per generation, and almost nobody changes it. Those 74% in "14–16s" are not a creative range; they are one factory setting recorded 945 times.

That explains two other things:

1. Why the loose phrase "8 to 16 seconds" circulates at all — it is a rough description of a needle.
2. Why only 130 cases (10.2%) land in 8–12s — producing an 8-second clip requires deliberately trimming, which is more work than accepting the default, so few people do it.

Finding 2: Short drama's mean is pulled up 1 second by a long tail

Short drama's median matches the overall median (15.2s), but its mean runs a full second higher, and its longest case is 160.7 seconds — more than ten times the median.

This is a textbook needle-plus-tail shape: the vast majority are the default 15 seconds, a handful are stitched long-form. Reporting the mean hides this structure; only the median reveals it.

Finding 3: The 81 long cases are almost certainly stitched

There are 81 cases over 16 seconds, and their profile is completely different:

The longest is a music & dance clip at 186.8 seconds. That length cannot come from a single generation — it is far past any model's output ceiling.

>

What you can do with this

If you are planning anything over 30 seconds: do not count on prompting your way there. By the data, you have to cut your content into 2–3 beats of about 15 seconds each, generate each at default length, then assemble. This is also why the techniques "multi-shot switching" and "cut timing" each appear in over 200 cases — they are the infrastructure of long-form.

If you are evaluating models: the 15-second default is a hard constraint worth testing on its own. The question is not "can it produce 30 seconds" but "how much frame consistency and motion continuity degrade past the default."

If you buy media: there are only 92 cases in the 0–8s range, and just 11 of them are product ads. That is to say, there is almost no directly usable public reference for ultra-short ads. That range is currently empty; you'll have to test it yourself.

Method and limitations

• Duration comes from container metadata, with error limited to encoder precision (typically 0.1s).
• Bucket boundaries are arbitrary. A 2-second interval affects how the shape reads; we also give integer-interval figures that do not depend on bucketing.
• "A default output" is an inference, not a documented fact: we reverse-engineered it from the extreme concentration, and found no official text stating this default. If official documentation says otherwise, follow the documentation.
• All figures use the 1,274 index records. After de-duplication (924 cases) the shape is the same: [14,16) covers 71.9%.

Reproduce it

The duration filter lets you check case by case, and re-bucket however you like: https://cases.aishifu.shop/