Nobody waited for perfect.
They waited for permission.

The question my clients changed

We make creative content for Malaysian brands. Six months ago, AI work came back from client review with one comment: it looks AI. That was a verdict, and it ended the conversation. The work was rejected, and the budget went back to a shoot.

This month the comment is different. Clients now ask is this AI? That is a question, not a verdict. It means they cannot tell at a glance — and, more important, that they are no longer sure it matters.

The models did improve in six months, but not by that much. Viewers of 后西游记 still complain that the faces are stiff and the expressions do not move. It still came first among provincial satellite channels in its slot on opening night. The quality did not cross a line. What changed was the audience’s permission to accept it.

Who gave the permission

In Malaysia, adoption waits for the customer. In China, this time, the state set the pace — and it did so on purpose.

August 2025. The State Council’s “AI+” opinion (国发〔2025〕11号) told the country to use AI “to assist in creating more cultural content carrying Chinese cultural elements.” A year later, the first AI drama in prime time is a sequel to Journey to the West. That is not a coincidence of taste.

The same month. The broadcast regulator’s 21-point reform (广电21条) opened a “produce, review and broadcast in parallel” pilot. 后西游记 is the first series to run under it. Five episodes aired first; the rest are being finished while the show is on air.

17 September 2026. The regulator ruled that AI-made programmes must carry a label and pass human review — and in the same statement said it would “encourage innovation where it should be opened up.”

Label it, don’t ban it. A regulator that writes the labelling rule is telling the market the product is legal. And the market did not wait for the rule: the stock hit its limit the morning the show aired, and the regulator confirmed the direction 17 days later.

What the production line now costs

The director, Li Dongshen, gave the numbers himself. A comparable traditional production runs about 300,000 yuan per minute. His runs at 20,000 to 30,000 yuan per minute — under a tenth. The idea came up in March, it was approved in May, and it went to air on 31 August. The usual timeline is one to two years.

The parallel-review model exists because of that cost. Change a scene in a live-action series after the audience reacts and you pay for another shoot: the crew, the location, the cast’s calendar. Change it here and you re-generate. In his words, with a real shoot “the cost would increase without limit. Only AIGC can do it.”

The crews already know. In Hengdian, China’s largest studio town, the number of short-drama crews shooting fell from 150 in August 2025 to 33 in May 2026. Day rates for extras on modern-dress shoots dropped to about 90 yuan. I wrote in The House Isn’t the Risk. The Thirty Years Are. that roughly 95% of China’s micro-dramas in the first quarter were AI-generated. The long drama was the last format that needed a set. As of 31 August, it doesn’t.

To be fair: the cinema has not died

“The film industry has collapsed” is the easy line, and it is only half true.

China’s 2026 summer box office closed at 12.5 billion yuan, up 4.45% on last year, with 340 million admissions. People still go to the cinema. But the summer of 2023 took 20.6 billion, so this summer was about 40% below that peak — even with a record 38.5 million screenings and the lowest average ticket price since 2022. Audiences have not left. The price of what they are sold keeps falling.

Hollywood’s problem is older than the models. Los Angeles logged 4,711 on-location shoot days in the second quarter of 2026, against 8,632 in the same quarter of 2019. The causes there are mostly cuts to reality TV and productions moving to cheaper states and countries, not AI. What it shows is that the old crew economy was already breaking before the model arrived. AI does not have to beat a healthy industry. It only has to be cheaper than a weak one.

The technology also still has a hard limit. Li calls long-range memory “the biggest bottleneck”: over a 40-minute episode the model forgets, faces drift, and a team checks the shots by hand. That is a job — and it is the next point.

Where production goes next

This is my read of the next three years. It is a prediction, not a forecast from any data set.

The camera stops being the entry ticket. Li said he used to prefer film-school graduates, and now he is not sure: art-school graduates imagine more freely “because there is no camera in their heads.” For a century, knowing how to operate the machine was the qualification. That qualification is being devalued fastest.

The work moves to the two ends. He describes his own job as two things: “imagination at the start and judgment at the end.” The model produces hundreds of versions of a shot, and a person picks one. What sits in the middle — the lighting, the set, the extras — is what the model does.

Television becomes a release cycle. Produce, review and air in parallel is how software ships: release, watch the numbers, patch. A drama that changes in response to its audience is no longer a finished product. It is a live service.

The law just created a job. “Human review is mandatory” means someone is paid to watch every AI frame before it airs. The regulator that removed the set also wrote the job description for the reviewer.

For Malaysia, this lands first in brand content, not cinema. “It looks AI” was the last objection that protected the price of a shoot. Once a client stops raising it, the budget conversation starts from the model’s cost, not the crew’s. A finished video made once and replayed is the case where AI really is cheaper; live content is the opposite case, as I explained in AI Livestream Is Not a Discount. I also showed where the price of recorded content was heading in RM20 a Video, Nobody Left to Bill.

If you run a production team, the move is plain: take your best people off operating equipment and put them on selection. The scarce skill is no longer making the shot. It is knowing which of the hundreds of versions to keep.

The bottom line

Acceptance did not wait for AI to be perfect. The faces in 后西游记 are still stiff, and it went to prime time anyway. It waited for permission. In China the state gave that permission on a provincial channel at 8pm. In Malaysia the client gives it one brief at a time. In my office, the last time “it looks AI” was a reason to say no was about six months ago.

If you are working out what this means for your own production budget or team, I am happy to compare notes over a casual conversation. No pitch.