Sanxingdui: Future Past is, by almost every measure, a bad film. Its dialogue was written by an AI tool called Bowen, and it shows — characters announce that a 'superbrain superintelligence costs $10 million dollars a month to subscribe to' with the conviction of a chatbot completing a prompt. The plot involves astronauts, ancient Egyptian ruins, bronze age artefacts, and space-time vortices, arranged in an order that defies cause and effect. The Guardian's correspondent watched it in a nearly empty Beijing cinema. The sole other audience member agreed it exceeded their very low expectations. Technically, Sanxingdui represents a genuine first. All of its visuals and some of its sound effects were generated by AI; humans handled the edit. The uncanny-valley problem is uneven: Chinese characters land closer to photorealistic, white characters drift further into the wrong territory. Everyone runs with the smooth, frictionless gait of a video-game NPC, which is disorienting when they're supposed to be fleeing catastrophe. One protagonist looks creepily like Tilly Norwood, the 'AI actor' billed in an upcoming Hollywood film called Misaligned — an unintentional meta-joke the film cannot have planned. The box office tells its own story: 1.7 million yuan (roughly £191,000) against a production cost the producer describes as 'tens of millions.' That is a catastrophic return, but Chinese industry insiders are not treating it as a verdict. A screenplay developer in Beijing reports that the majority of projects she's been approached about in recent months have been AI films. The logic is blunt: what once required 100 people can now be done by five. The rest, she says, won't have jobs. The comparison that matters most isn't to prestige cinema — it's to the AI microdrama ecosystem that already exists. Tens of thousands of one-to-three minute AI-generated dramas are uploaded to Chinese streaming platforms monthly. That industry, which did not exist a few years ago, is now estimated at $3 billion. Sanxingdui may be struggling partly because audiences won't pay 50 yuan (£5.60) for a cinema ticket when comparable content is free on their phone. The experience premium of the cinema — the reason people dress up and leave the house — doesn't yet apply to AI-generated work. The deeper question is whether Sanxingdui is a proof-of-concept or a warning. State media called it 'riddled with flaws and imperfections' while suggesting it 'may yet go down in cinematic history as the starting point of what came next.' That framing — first draft of history, not finished product — is exactly how investor logic works in China's tech-forward industrial policy. The film doesn't need to be good. It needs to demonstrate that the infrastructure exists and that audiences can be brought along eventually. What Sanxingdui actually demonstrates is the gap between technical viability and emotional craft. AI can generate faces, landscapes, and action sequences. It cannot yet generate stakes. The characters don't feel like they'll be missed if the algorithm wins. Audiences feel that absence even when they can't name it — which is why 'not entertaining' is the correct critical verdict, and also an incomplete one. The question isn't whether AI can replicate cinema. It's whether cinema is what AI-generated moving images are even trying to become.
