Gyllenhaal Ditches AI in Marilyn Monroe Short, Cites Failure

Maggie Gyllenhaal arrived in Venice this year not just as jury president of the film festival but as a filmmaker with an uncomfortable story to tell about artificial intelligence. Speaking Friday at The Future of Creativity summit, hosted by Finch & Partners and CAA, she revealed that she had tested AI video technology while making her short film Flesh Impact — and then thrown the results away because, in her words, it "fundamentally did not work" 178.

The short imagines Marilyn Monroe at 100 years old, with flashbacks to Monroe in her prime. Dakota Johnson plays the younger Monroe; Ellen Burstyn plays the imagined centenarian version 7810. According to Gyllenhaal, the people who commissioned the project asked her to use AI — not OpenAI's tools, she specified, but "licensable AI" trained on Marilyn Monroe films that had themselves been licensed 7819. The plan was to train a system on Monroe's screen appearances and then place a generated version of her face over Johnson's performance 7920. After what she described as extensive work, Gyllenhaal discarded the AI footage, saying Johnson's unaltered performance was "so much more moving, human, alive" and that the AI version had killed off "all the reasons we decided to make this project to begin with" 7819.

What she actually tested

It's worth being precise about what this episode does and doesn't demonstrate. Gyllenhaal's account describes something closer to a digital face-replacement or likeness-transfer workflow than a from-scratch, text-to-video production. Johnson performed the role on camera; the AI's job was to swap in a generated Monroe likeness over that performance. That is a narrower and arguably harder task than generating an entire scene from a prompt, because the system has to preserve motion, expression and continuity while overlaying a recognizable identity — and Gyllenhaal's verdict was that doing so flattened exactly the qualities that made Johnson's acting work 91020.

She extended that argument beyond the technical result. At the same panel, she said what makes an actor like Jessie Buckley "brilliant" is her interaction with the world around her — with the director, with other performers — something she argued AI cannot replicate 10. George Clooney, at the same event, offered a more qualified view, suggesting some work will be replaced by AI even if performers with true star power won't be entirely supplanted 1017.

Gyllenhaal also raised concerns that went beyond craft. She said AI carries "huge environmental issues," pointed to controversy over data centers in South Memphis, and referenced AI pioneer Geoffrey Hinton's suggestion that AI systems be built with something like a "maternal instinct" toward humans 720. On the economics, she was openly skeptical of the industry's "democratization" pitch — the idea that a $40 million film could eventually be made for $4 million, or even $40,000 — arguing that the companies selling that promise are also the ones positioned to profit from it 720.

The bigger AI-video backdrop

Gyllenhaal's experience lands amid a wave of releases pushing video-generation technology toward exactly the kind of integrated, multimodal filmmaking pipeline her episode complicates. OpenAI's original Sora, introduced in 2024, could generate video from text, animate still images, and extend or fill in existing footage using a diffusion-transformer architecture 1112. Its successor, Sora 2, added synchronized dialogue and sound effects, more accurate physics, and a "characters" feature letting real people insert their verified likeness into generated scenes — with OpenAI stressing that individuals retain control over who uses their digital likeness 16. Notably, both Sora products were later discontinued, a detail the source documentation flags but does not explain in the context of this story 111216.

Google has pushed a parallel track. Veo 3, launched in May 2025, was the first of Google's models to generate video with native audio — ambient sound, dialogue, background noise — alongside the Imagen 4 image model and the Flow filmmaking tool, which combines Veo, Imagen and Gemini into a single creative workflow 1314. By October, Veo 3.1 added more granular editing: inserting or removing elements from a scene, bridging start-and-end frames, and extending clips into minute-long sequences, with Google reporting more than 275 million videos generated in Flow since launch 15. Separately, Google has also rolled out new AI image tools and an agentic Gemini feature for video analysis 35.

Meanwhile, Alibaba launched Wan3.0, its latest video-generation model, timed alongside a $10 billion share placement to fund AI spending — coverage that frames the release as both a technical upgrade and a competitive shot at Google and OpenAI, with one outlet noting Alibaba's Qwen model had already outperformed Gemini and ChatGPT on coding benchmarks 246.

Where the reporting agrees

On the core narrative, the coverage is remarkably consistent. Every account — from AFP's wire report picked up by Yahoo, France 24, RFI and CP24, to Variety, The Hollywood Reporter, The Cut and the Evening Standard — agrees on the essential sequence: Gyllenhaal was asked by the project's commissioners to use AI, agreed to try it out of curiosity, trained a system on licensed Monroe footage, attempted to place Monroe's face over Johnson's performance, and ultimately rejected the result because it stripped away the humanity of Johnson's acting 17891017181920. Multiple outlets also converge on her framing that AI's promised cost savings — the $40 million to $4 million to $40,000 trajectory — mainly serve the companies selling the technology 71720. And several pieces situate her comments within a broader, tense Venice conversation about AI's effect on acting jobs, with Clooney's warnings and festival director Alberto Barbera's more measured optimism cited as counterpoints 1718.

Where it doesn't

The reporting diverges in a few notable ways. Running time is one: AFP-derived stories and Variety describe Flesh Impact as a 17-minute film, while the Evening Standard reports it as 22 minutes — a discrepancy likely tied to different cuts or press materials rather than a factual error by either outlet. Character framing also differs slightly: some accounts say the AI was meant to age Johnson's Monroe up to 100, while The Hollywood Reporter is more precise in noting Johnson plays Monroe at the height of her fame while Burstyn plays the imagined elderly version — a distinction that matters for understanding exactly what the AI was being asked to do. Only The Hollywood Reporter reports that the project had the blessing of Authentic Brands Group, the Monroe estate's rights holder, and names the commercial partners behind the short (Genesis, The Cut, Superprime, INNOCEAN) — details absent from the wire coverage and other outlets, which leave the commissioning party unnamed. Only Tribune and Variety-derived accounts include Gyllenhaal's Geoffrey Hinton anecdote and the South Memphis data-center reference in full; the AFP wire versions compress or omit these secondary points entirely, focusing instead on the labor and job-loss angle.

The evidence best supports reading this as a genuine, if limited, creative test rather than either a wholesale AI failure or a media-friendly stunt. Gyllenhaal's own account is consistent across every outlet that quotes her directly, and the technical premise she describes — face replacement layered onto a live performance — is a plausible, well-defined workflow rather than an exaggerated claim. The variation between outlets looks less like contradiction than like selective emphasis: wire services condensed her remarks for a global audience concerned with job security, while entertainment trades preserved the fuller texture of her environmental and economic critique. Nothing in the sources undermines the central fact: she tried AI, it visibly failed her own artistic standard, and she said so publicly and repeatedly in the same words.

Why it matters

The episode does not prove AI video generation is failing broadly — Sora 2, Veo 3.1 and Wan3.0 all point to rapid gains in realism, audio synchronization and editing control. What it does show is that photorealistic likeness reproduction and compelling performance are not the same achievement. A system can be legally licensed, technically functional and still fail the test that matters most to a director: whether the result still feels human. As studios weigh whether AI augments actors or replaces them, Gyllenhaal's experiment is a data point on the side of caution — not because the technology can't render a face, but because rendering a face isn't the same as making a film.