What the production actually tested
Gyllenhaal said Flesh Impact’s commissioner asked her to explore artificial intelligence for scenes in which Dakota Johnson plays Marilyn Monroe in her thirties. According to the director, the production did not use a general public model or scrape material without permission: it used what she described as licensable AI and trained it on Monroe films the production had licensed. The team then tried placing a generated Monroe face over Johnson’s performance. That is the verifiable boundary of the account. The available report does not name the model provider, the exact licences, the contractual permissions or the technical method in enough detail for an outside audit.
Why the licensed approach was still rejected
The decisive failure, in Gyllenhaal’s telling, was creative rather than procedural. She said Johnson’s unaided work felt more moving, human and alive, while the overlay killed the qualities that justified the short in the first place. The team therefore threw away the AI version. This does not establish that every licensed synthetic-image process will fail, nor does it make a legal finding about other systems. It shows something narrower and more useful: a production can clear an initial rights route, spend time testing the technology and still decide that its visible effect weakens the actor’s contribution.
The consequence for performers and audiences
For performers, the example makes consent and licensing only the beginning of the production question. A technically authorized transformation can change which details of expression survive on screen and whose creative labour audiences are actually seeing. Gyllenhaal also raised environmental and economic concerns, including the possibility that cheaper synthetic production could make conventional budgets harder to secure, but those are her stated concerns rather than measured outcomes from Flesh Impact. What matters now is whether the finished short and any later production documentation explain the discarded experiment further, and whether other filmmakers publish comparable decisions instead of treating licensed AI as automatically acceptable or automatically successful.