Nine eras. Twenty-five actors, more than a hundred extras, close to a hundred and fifty crew, and — because somebody had to be on set holding them — a pig, five chickens and dozens of rats. That is what Amplitude Studios says went into the ninety seconds of live action it premiered for Humankind 2 at Gamescom’s Opening Night Live, shot at Nu Boyana Film Studios in Sofia with practical sets and props built for the job.
The internet watched it and said a machine made it.
Not a fringe reaction either. Within hours of the trailer going up the comments had settled on “AI slop”, the phrase doing most of the heavy lifting in arguments about images this year. Amplitude answered by email — real actors, practical sets, custom-made props — and then, on 4 September, put out a behind-the-scenes clip showing the sound stage, the blue screen and the crew standing around in the middle of it.
The tells people used were all real
Here is the uncomfortable part for anyone who makes pictures for a living. The things viewers pointed at were all genuinely in the footage. Every one of them also turns up in work nobody generated.
Shot length, first. A ninety-second sweep through nine eras gives you maybe four seconds a shot, most of them less. Nothing is held long enough to inspect a hand, a face, a piece of set dressing. Generative video clips are short for a different reason — the models were trained and served in short chunks — but the grammar that comes out the other end looks identical. Quick cuts, no lingering. An audience that has learned to distrust footage nobody lets you stare at will distrust this too.
Then the compositing. Nine historical periods in one shoot means blue screen behind almost every plate, environments built in post, and a grade laid over the whole thing to make Bronze Age and orbital mechanics feel like one film. That grade is doing enormous work. It also flattens everyone into the same light, softens the edges where actor meets background, and produces exactly the plasticky evenness that diffusion output has. A comp artist earns their week hiding those seams. The better they hide them, the more the result resembles an image that never had seams to begin with.
Crowd extension is the third one. You shoot a hundred extras and you multiply them, and multiplied crowds have repeating faces and repeating gaits at the back of frame. Tiled duplicates are one of the oldest tells for generative video. They are also just what crowd work has always looked like when you zoom in.
BETC handled the creative, and LiveFX did the CGI on the pieces that could never be practical — a tornado, a nuclear blast, the space sequence. Those are the shots where the seams live.
The montage form is the trap
Strip the argument back and what you get is this: the visual language of “sweep through all of human history in a minute” is the same visual language a text-to-video model produces when you ask it for anything ambitious. Scale jumps. Era jumps. A camera doing impossible moves. Everything gorgeous and nothing sitting still.
Studios have been making that trailer for thirty years. It belonged to them first. But the models learned on it, they are fluent in it now, and audiences have started reading the form itself as a signature. Amplitude walked into a shape that no longer signals what it used to.
I have watched enough reveal reels to know that the ones that survive this scrutiny tend to share a quality: somewhere in them a shot is held long enough that you can see something specific and slightly wrong. A prop with a real scuff. A performer’s weight shifting. The chickens, honestly. Nobody prompts a pig.
What Amplitude did not say
They confirmed the shoot. They did not say whether generative AI touched anything in post-production, and nobody has asked the question in a way that got answered. Those are different claims, and a studio that has just spent a week defending its live-action production has every reason to leave the second one alone.
Which matters, because most of the accusation amounted to “something here was machine-made”, and a behind-the-scenes clip of a sound stage does not settle that question. Compositing tools artists already have on their desks ship generative fill, generative extend, generative cleanup. The line between “we filmed this” and “no model touched this” stopped being the same line about two years ago, and studios have not worked out how to talk about the gap.
The blog covered a version of this in August, when LinkedIn’s million-click “seems like AI slop” button turned out to suppress human renders at the same rate as generated ones. Same failure, different venue. Neither a crowd nor a classifier can tell a heavily-post-processed real image from a synthetic one, because at the pixel level, after the grade, there is often not much left to tell.
The cost is entirely one-sided
An accusation costs one comment. The rebuttal cost Amplitude a press statement, a set of production stills, a behind-the-scenes edit and a news cycle spent talking about their process instead of their game.
That asymmetry is the working condition now, and it lands hardest on the people with the least budget. A studio with SEGA behind it can produce call sheets from Sofia. A freelancer whose ArtStation post gets the same comment has a layered file and a timelapse nobody will watch, which is the same evidence in a form that convinces nobody.
Back in August I wrote about reading Opening Night Live trailers for what they actually show you — in-engine versus target render, the tells that tell you something. Those tells still work. “Looks AI” reads a genre rather than an image, and the genre now belongs to everybody.
Production figures, locations and partners as stated by Amplitude Studios in its behind-the-scenes announcement of 4 September 2026 and in its emailed response to PC Gamer. Trailer premiered during Gamescom Opening Night Live on 25 August 2026. Whether generative tools were used in post-production is unconfirmed either way at the time of writing.