You can't prompt your way to the end

"You can't prompt your way to the ends." That is Dave Clark, co-founder and chief creative officer of the AI studio Promise, in Forbes in July 2026. He is not arguing against the tools. He builds with them. He is describing the part of the job that the tools do not touch, and almost everyone shipping paid AI work describes the same part.

What follows is what practitioners with work on air say is still hard. Every quote is named and dated. Where a quote is old enough that it should be read as history rather than current testimony, we say so.

The problem is holding one thing together

Clark's point in that interview is that off-the-shelf models cannot reproduce Promise's layered, 3D-positioned approach. Forbes frames the underlying difficulty as holding one look across a twenty-minute runtime. A single shot is a solved problem. Twenty minutes that feel like they came from the same mind is not.

Clark raises a second thing, and it is the one people miss because it has nothing to do with software. How does a director communicate a vision to a team of a hundred people? That is a language problem, and it was the hard part of directing before any of this existed.

You cannot change your mind late

Tim Katz, co-founder of Knucklehead, puts the production consequence in one line: "changing a shot late is really hard". Which inverts the usual sales pitch. Generation is supposed to make revision cheap, and within a shot it does. Across a film that has already been built, it does not. AI has to be planned in from the start rather than bolted on at the end, and a project that treats it as a rescue in post will cost more than one that decided in pre-production.

Chris Hewitt, creative director at Airhead, names the other failure in the same conversation. Leave details unspecified and the model fills the gap with "the AI slop we see all over social media." Under-specification is not neutral. The model will always answer. A brief with holes in it produces a film with the internet's average taste poured into the holes.

The model is not accountable for anything

Jason Zada of Secret Level, who made Coca-Cola's holiday films, is blunt about where responsibility sits: "If you're prompting 'Star Wars,' it's your fault, not the model's fault." Derivative output is a brief problem, not a tooling problem, and that applies to legal exposure as much as to taste.

He also says something about who is good at this: "When traditional entertainment people move into this space, I've seen the quality of work massively improve." The evidence for that is sitting in his own curriculum. Secret Level Academy's named topics are quality control for AI images, video continuity and consistent characters, synthetic characters, and integrating real actors into AI scenes. A syllabus is an honest document. Those four subjects are the discipline's admitted weak points, written down by someone with no interest in advertising them.

“When traditional entertainment people move into this space, I've seen the quality of work massively improve.”

Performance is the wall

Liang ZHU, VP at iQIYI and head of intelligent production, states the thing most AI film marketing avoids: "An actor's performance is the core of any film or series." The whole design of iQIYI's new Shanghai studio is an answer to that constraint. Shoot the actor live, then use the footage as reference so sets, lighting, costume, makeup and props can change in post while the original performance survives intact. The generative work is pointed at everything except the acting.

The origin of this argument is older than the current models, and should be read as such. In May 2023, Paul Trillo described directing AI voice performance for a short film and said "one wrong intonation can throw the whole thing off". He edited the performance sentence by sentence, sometimes word by word, and noted that the model was making camera movement, blocking and lens choices for him. That is three years old and the tools have moved a long way since. We include it because it is where the observation starts, not as current testimony about current models.

The work that lands is hybrid, and nobody hides it

Nem Perez, head of studio at The Generation Company, frames the brief on Final Twist as a question rather than a capability: "How do you use GenAI to reconstruct a crime scene from only a handful of archival images and make it feel real?" The answer was generative AI plus 3D, Gaussian splats, Unreal and a volume stage. Four disciplines, one of which was generative.

David von Ritter of cavalcade, on the AI films made for Swissquote, gives the cleanest test in the category: "AI is like plastic surgery: it's good if you can't see it." That interview, from June 2026, also records what it took. The bull had to be rebuilt in 3D, with substantial research and development, before it could perform physically complex actions and stay consistent. The invisible version is the expensive version.

Guy Soulsby, chief creative officer at Made by Humans, reduces it to seven words: "Generative AI tools alone don't make great work, people do."

What the Cannes jurors were actually saying

Three Cannes Lions 2026 jurors, speaking in May 2026, converge on craft as structure rather than finish. Rakesh Menon of ULKA: "Craft is not a layer you put on top; it has to be built into the idea itself." He goes further on what the models cannot reach: "AI doesn't understand the nuances of our culture, our humour or our complex emotions." And then the line that should be pinned above every workstation in a studio like ours: "Creativity feeds off instinct. And instinct is earned, not generated."

Gurbaksh Singh of dentsu describes the standard: "The strongest work does not feel AI-made or human-made. It simply feels thoughtful, original, and relevant." Kainaz Karmakar of Ogilvy locates the human advantage precisely, in dissatisfaction: until a machine can feel discontent with what it has created, the advantage holds. Her second remark is the honest one about responsibility. "The tool is innocent. So far. The users come in all shades of grey."

The reframe that is actually useful on a budget

Darren Frankel of Adobe has the warning: "If you don't have that humanity, it'll be all frosting and no cake." But the most practical sentence anyone has offered belongs to Guy Danella at XYZ Films, and it is a question rather than a position: "Is there a way to use AI to save enough so we can add another day of shooting?"

Take that to a production meeting. It converts the whole argument into a line item. The saving is not the point of the saving. Spend it on the day with the actor.