Don't ask, don't tell

Joel Kuwahara produced animation on early seasons of "The Simpsons." This week he told the Los Angeles Times that Hollywood studios are using ByteDance's Seedance with "a wink and a nod." His exact phrase: "It's kind of like a 'don't ask, don't tell' kind of a thing."

ByteDance is the former majority owner of TikTok. Earlier this year, a fifteen-second Seedance clip of Brad Pitt fighting Tom Cruise on a rooftop went viral and the Motion Picture Association demanded the company halt its "infringing activity." The copyright dispute is unresolved. The geopolitical tensions are real. Disney's JioStar joint venture executive calls the model's quality a "wake-up call" for American competitors. And studios that have not officially approved Seedance are using it anyway, because it costs nine dollars a minute while Google's Veo costs twenty-four.

Fifteen dollars separates the approved model from the one nobody admits to using.

The gradient and the invoice

The industry spent the spring building institutional frameworks. Copyright law requires human authorship. The Academy requires "human-authored" screenplays. The DGA's new contract, effective four days ago, says AI-generated footage is "treated like footage." The EU's Article 50 arrives in twenty-eight days. The Golden Globes require human creative direction to remain "primary." Twelve institutional responses now sit on the gradient, each measuring a different dimension of human involvement.

Every single one asks who made the creative decisions.

Not one asks which model executed them.

The gradient was built to answer authorship. It was not built to answer provenance. A filmmaker who exercises structured vocabulary, iterates through takes, reviews every frame, and takes editorial responsibility satisfies every institutional test regardless of whether the model underneath is Veo, Kling, Runway, or a ByteDance product that the MPA sued five months ago. The frameworks do not distinguish. The invoice does.

The price did the talking

Seedance costs $9 per minute for video with audio generation on Artificial Analysis benchmarks. Veo costs $24. Runway sits between them. The math is simple. A filmmaker generating a hundred minutes of raw material to select from spends $900 on Seedance or $2,400 on Veo. At Hell Grind's 63:1 generation-to-selection ratio, the cost difference across a feature becomes tens of thousands of dollars.

The producer of "Paranormal Activity" announced "Terrarium," a horror film to be generated entirely with Seedance. The director estimates fifteen seconds of high-definition video costs five dollars. An independent filmmaker making a fantasy series on YouTube averages three million views per episode using Seedance for character consistency. Jon Erwin's "Young Washington" opened in theaters Friday with about a hundred AI-augmented shots across a hybrid pipeline. ByteDance has been quoting major studios two million dollars for unrestricted special access.

Luma's CEO argues the ceiling is limited. Disney will never use ByteDance for the next Snow White. That is probably true. It is also probably irrelevant, because the conversation has already moved past Disney. Independent filmmakers, horror producers, YouTube creators, and the vast middle of the industry that sits below tentpole budgets are making the decision every day. They are not loyal. The director of "Terrarium" said it plainly: "Whatever is the best, we're going to use it."

The supply chain nobody certified

The Human Made Mark certifies whether a human was present. The Academy certifies authorship. Copyright certifies decision-making. The EU certifies editorial oversight. Every framework on the gradient addresses the creator's relationship to the output. None of them address the creator's relationship to the tool's origin.

Seedance was launched in the U.S. at a Santa Monica event hosted by a group linked to the Chinese government. ByteDance hired for a hundred open roles, signed independent filmmakers, held private financing conversations about AI films, and threw a caviar party at Cannes. The company's model was trained on data that included copyrighted characters with enough fidelity to generate a Brad Pitt fight scene that a screenwriter called "terrifying." The MPA sent a cease-and-desist. ByteDance declined to comment on its U.S. expansion.

The institutional gradient has no column for any of this. A filmmaker using Seedance satisfies the Academy's authorship test, the DGA's jurisdiction provision, the EU's editorial exemption, and the copyright framework's human-decision requirement exactly as cleanly as a filmmaker using any other model. The frameworks were designed to measure the human in the loop. They were not designed to measure the company behind the pipe.

The Human Made Mark asks: was a person in the room? Nobody built a certification that asks: who built the room?

The vocabulary does not care

A structured cinematographic prompt works identically on Seedance, Veo, Kling, Runway, or any other model. Forty words specifying lens behavior, lighting direction, compositional placement, and atmospheric texture carry the same creative intent regardless of which company's servers process them. The vocabulary is portable. The series has documented this across a hundred and twenty-six prior articles. The BYOK architecture was designed for exactly this kind of market reality: the filmmaker brings the knowledge, the provider brings the compute, and neither depends on the other.

But portability is not neutrality. Using a tool and endorsing its provenance are different acts. The filmmaker who exercises vocabulary on Seedance produces better output than one who types four words into Veo. The frameworks on the gradient would reward the first filmmaker and penalize the second. The frameworks are correct about the craft question. They are silent about the supply chain question.

"Don't ask, don't tell" is the sound of a market that answered the cost question before the institutions finished writing the authorship question. The gradient spent three months measuring who made the work. The invoice took three seconds to measure which model made it cheaper. Both measurements are real. Only one has rules.

Twenty-eight days

The EU AI Act's Article 50 arrives on August 2. It requires machine-readable watermarks and mandatory disclosure. It exempts content that underwent human editorial review. It does not ask which model produced the content or which country trained it. A filmmaker exercising editorial control over Seedance output is exempt from the label. A family generating moonwalking grandfathers on Google TV is not. The regulation measures editorial judgment, not geopolitical origin.

The DGA contract says AI-generated footage is footage. It does not say which AI. The Academy says performances must be demonstrably performed by humans. It does not say which model rendered the environment around them. Copyright requires human creative decisions. It does not require an approved vendor list.

Every framework measures the filmmaker. None of them measure the factory. And the factory with the lowest prices and the highest capability just walked into Hollywood with a wink, a nod, and a two-million-dollar access fee for anyone who wants to stop pretending.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never winked at a vendor and suspects the gesture loses something over API.