How AI Actually Changed Online Video
Machine learning moved from the edge of video production to its centre. A look at what genuinely changed — and what was oversold.
For most of the history of online video, the constraint on publishing more of it was human attention. Someone had to watch the footage, decide what mattered, and cut it.
Machine learning loosened that constraint. Models can assess footage for the things that correlate with attention — motion, faces, speech, audio energy, scene changes — and act on that assessment without a person in the loop.
What genuinely changed
- Packaging got cheap. Cutting a long asset into short variants used to cost editor time. Now it costs compute.
- Selection became per-viewer. Which clip to show can be decided at request time rather than fixed when the page was built.
- Archives became addressable. Once footage is analysed, moments inside it can be found, which makes old material usable again.
What was oversold
A great deal of the claimed improvement never showed up in anyone's analytics. The pattern is consistent: automation helps where a judgement is mechanical and repeated thousands of times, and disappoints where the judgement is genuinely editorial.
Choosing the strongest frame from ten thousand candidates is mechanical. Deciding what a story means, what to cover, or what angle to take is not — and systems marketed as doing the latter tend to produce output nobody finishes watching.
The useful framing
Treat these models as a way to apply an existing editorial standard more widely, not as a replacement for having one. The teams getting real results from AI video are the ones who knew what a good preview looked like before they automated making them.