Predictions in this field age badly, so this is not a list of forecasts. It is a description of what has actually changed in the last year and what those changes imply for someone selling products online, which is a more useful thing to have.
The quality problem is mostly solved, and it was the wrong problem
Two years ago the honest objection to AI video was that it looked wrong. Hands, physics, objects that changed shape between frames. That objection has largely dissolved: current models produce footage that a viewer scrolling a feed does not question.
What became obvious once quality stopped being the constraint is that quality was never what made an advertisement work. A beautiful clip of your product doing nothing in particular is still a beautiful clip of your product doing nothing in particular. The bottleneck moved from rendering to deciding what to render, which is a writing problem wearing a technical costume.
This is why the tools that produce results now are the ones that plan before they render, and why prompt-to-clip tools plateau for advertising even as their output gets more beautiful.
Audio stopped being a separate step
Models that generate sound with the picture changed the workflow more than the output. There is no longer a stage where somebody picks a music bed, which removes both a cost and a characteristic failure mode: the library track that everyone recognises from other ads.
Length became a structural question
Short models produce clips of a few seconds. Getting to twenty or thirty seconds means assembling chapters, and assembling them badly is obvious: a visible cut, a change of room tone, footage that repeats itself at half speed. The tools that handle this well plan each chapter to add something the previous one did not, and stitch audio across the joins.
For advertising this matters because a hook, a benefit and a call to action do not fit in eight seconds, and a video that only has room for the hook is a video that never asks for anything.
Volume beat polish, and it keeps beating it
The single most reliable pattern in paid social over the last two years is that the advertisers who win are the ones who can produce new creative faster than the platform burns it. Not the ones with the best single asset. Creative fatigue is the actual enemy, and it is fought with throughput.
This is a strategic point, not a technical one, and it is the one that should change what you do. If producing an ad takes a shoot day, three variants a month is ambitious. If it takes minutes, testing is just a habit.
What has not changed
Video still cannot make people want a product they do not want. It removes a specific friction, the friction of not being able to see a thing properly, and where that friction is not what is stopping your buyers, better video changes nothing. Every year of improvement in these models leaves that fact exactly where it was.
Related reading: Working out whether product video pays and Why multi-chapter video outperforms single-clip ads.