The baby in the machine: How filmmakers are reimagining cinema with AI

On a film set in Lower Manhattan, actress Audrey Corsa held nothing in her arms. The baby she appeared to cradle was not on set. It would be created later by an artificial intelligence model trained on hundreds of photographs and one woman’s first memory. That memory was the moment of her own birth.

The film is Ancestra, a new short directed by Eliza McNitt, and it features a simulated version of McNitt herself, modeled after photographs taken by her father on the day she was born. These are the same images used to create the baby in the opening scene. When actress Audrey Corsa holds that newborn onscreen in the climactic final scene, she is in fact holding a digitally generated human being, a kind of hallucinated memory projected into being by software and intention. For McNitt, that gesture is more than cinematic sleight of hand. It is an act of emotional excavation.

Noted filmmaker Darren Aronofsky introduced McNitt to generative AI, an experience she says sparked the idea for Ancestra. He invited her to explore what these new tools could do and ended up producing the film. The result is a deeply personal short about a mother’s emergency delivery that fuses live action performance with AI-generated imagery rooted in cosmic and corporeal transformation.

“My mom had to have an emergency C-section when I was born,” McNitt told me during a recent conversation from her home in New York City. “She walked into the hospital for a checkup, and everything changed. That moment became the seed of Ancestra.”

The film tells the story of a mother who, during an emergency delivery, reaches across time and memory to connect with the women who came before her and with the unborn daughter she hopes to save. Ancestra fuses traditional live action and pioneering AI-generated imagery into a narrative that feels at once deeply intimate and dizzyingly cosmic. It also hints at where filmmaking could be going, using AI not to replace human creativity, but to build on it.

McNitt is quick to stress that this isn’t a film about AI. “It’s a film about my mother,” she said. But her use of generative models developed by Google DeepMind, including the video generator Veo and image models like Imagen and LoRA, offers a glimpse of what could very well prove to be a new kind of cinematic language—one built not solely from camera lenses and location scouts, but from prompts, neural networks and deeply personal data.

Veo is an advanced text-to-video system capable of generating short 1080p clips with high temporal coherence, object consistency and stylized rendering. But achieving those results requires carefully written prompts, multiple iterations, and an understanding of how the model interprets visual instructions. McNitt and her collaborators worked within the platform’s constraints—clip length, fidelity, prompt responsiveness—by fusing Veo’s outputs with motion path references and custom 3D scenes built in VR. The team first built animated previsualizations to map out the timing, then used them as a guide for the AI-generated footage.

To keep the film’s look consistent, McNitt’s team built their own companion style transfer model called LoRA. It was not a Google tool but a system they developed specifically for Ancestra to carry the same visual style through every scene.

At one point in Ancestra, we plunge into the character’s brain and spiral down through layers of tissue into the womb, where a fetus floats, struggling for air. Later, we witness a C-section from above, the flesh of the abdomen parting like petals around a burst of light. These moments weren’t rendered with traditional CGI. The team coaxed them into being through a painstaking iterative process of prompt crafting, image selection, motion tracking and AI fine-tuning.

“We generated over a thousand images to get a single usable one,” McNitt said. “It was an intense process, a conversation with the machine.”

For some scenes, the team used Imagen to generate high-resolution stills, which they then turned into short video clips using frame prediction tools like Runway and Gen-2. To keep the imagery emotionally grounded, they avoided generic prompts. Instead, they used poetic language, VR blocking and lighting references based on real hospital environments.

Part of what makes Ancestra unusual is that its use of AI is neither flashy nor gratuitous. There are no synthetic actors, no artificial voiceovers. Instead, the film is grounded in the real: a cast led by human performers, a script rooted in lived experience and a team of artists guiding every pixel of the production.

“That was a line I wouldn’t cross,” she said. “The actors had to be human.” Instead, for McNitt, AI was a way to show what can’t be captured on camera.

That imperative led her team to develop what she calls an “AI unit” on set, a separate creative cohort charged with developing the dreamlike visual sequences. Their work included training a model on her late father’s photographs, many taken while dangling from helicopters as a marine and aerial photographer. His eye, quite literally, helped shape the visual DNA of the film.

“He passed away recently,” McNitt said. “And I had so many questions I never got to ask, about his work, his lens choices, the film stock he used. Training the model on his images became a way of collaborating with him again.”

That idea—using AI not as an anonymous generator but as a vessel for memory and collaboration—runs through Ancestra. It is also part of why the film feels distinct from the often dystopian discourse surrounding AI in the arts.

McNitt is well aware of the controversy. The use of generative tools in film, especially following the Hollywood strikes and ongoing concerns about digital labor, copyright and creative authorship, remains a contentious topic.

“I set very clear boundaries,” she said. “We didn’t use synthetic actors. We didn’t ask the model to mimic other artists’ work. Everything we generated was built from personal material or crafted for the story.”

In discussing her creative process, she drew a line between direct imitation and the more nuanced act of weaving influences into something new. “I think there’s a difference between referencing and copying,” she said. “Every artist is inspired by others, but it matters how we use that inspiration.”

While the ethics of borrowing another artist’s style tend to center on copyright and creative ownership, the discussion around AI in this film also touched on a different concern: whether the technology might displace human labor.

“There were over 200 people involved in making Ancestra,” she said. “The AI unit wasn’t replacing jobs. It was adding to the creative toolkit.”

Still, she acknowledges the limitations. Tools like Veo can only generate a few seconds of video at a time (meaning every shot has to be built in fragments and then painstakingly stitched together). Maintaining visual coherence across a narrative requires enormous effort (involving careful prompt design, style matching and repeated iterations to keep characters, lighting and camera movement consistent) and often tricks the model into breaking its own rules, or at least exposes how unpredictable prompting can be when trying to guide AI toward a specific creative goal.

“When we tried to prompt space imagery, it came back looking like a science textbook,” she said. “Very literal, not cinematic. We had to find indirect ways to prompt it, using more poetic language to guide the aesthetic.”

Sometimes the results were mesmerizing. In one shot, a cosmic abstraction of a C-section seen from above, with the baby emerging into light, came from a model trained on motion references created in virtual reality, with style cues pulled from her father’s ocean photography. The result is something both hyperreal and dreamlike: a moment no camera could ever capture, but that feels emotionally true.

One of the most powerful parts of the film is how McNitt mixes her own story with the technology. The AI wasn’t trained on random images, but on her life. The baby in the final scene is a digital version of her as a newborn.

She used photographs taken by her father in the hospital, along with imagery from his decades of professional work. He had spent much of his career photographing the ocean from above, chasing light and perspective across the sea. The result is that his style, even in abstracted form, shapes much of the film’s visual tone.

“It was like having him as my cinematographer,” she said. “Even though he’s gone.”

The experience also made her rethink what it means to be the creator. If a machine uses your past to make something new, who owns it? The person giving the instructions, the machine or the memory?

McNitt doesn’t claim to have an answer. But she sees the process as deeply human, from the first rough prompts that generate strange, unintended textures to the moments when an AI-rendered figure’s expression shifts in a way no one scripted. “The mistakes we make, the things we don’t plan for—those are often where the art lives,” she said. “And those are human.”

That philosophy guided the production. Rather than use AI as a shortcut, the team treated it like a wild, unpredictable collaborator. Shots were storyboarded by hand, modeled in VR and then reimagined through multiple iterations of generation. Visuals were often constructed from hundreds of discarded attempts, each one closer to a feeling.

“We weren’t looking for perfection,” McNitt said. “We were looking for something that felt alive.”

She hopes this approach can expand beyond cinema. If machines can reflect us—not mimic us but truly reflect us—then perhaps they can become part of a broader human toolkit. Not as a replacement, but as an amplifier. That might mean enterprise uses that don’t feel corporate. A training video with an emotional arc. A company origin story told in images drawn from the founders’ real lives. An explainer that lands more like a short film than a slide deck.

Several companies are already exploring this territory. AI-generated internal corporate videos, safety walkthroughs or even customer onboarding tools are starting to borrow cinematic techniques, layering tone, music and narrative structure onto otherwise dry material. The same tools McNitt used—Veo for visual synthesis, Imagen for still generation, Runway and Gen-2 for animation interpolation—are being tested for enterprise content.

The difference, she said, lies in the intention. It is the distinction between using AI as a passive generator of random output and treating it as a collaborator shaped by a clear creative vision. “AI lets you externalize imagination,” she said. “That’s powerful. But it has to be guided.”

She believes that guidance has to come from storytellers—not from technologists or marketers, but from people who know structure, tone and feeling.

The potential pitfalls are clear. Generative AI can produce content that dazzles with size and novelty yet leaves no lasting emotional impression. Visuals may impress technically but fail to connect with viewers in a meaningful way. McNitt hopes for something quieter, stranger and more personal.

“I don’t want AI to be about spectacle,” she said. “I want it to help us say something we couldn’t otherwise say.”

In the case of Ancestra, the result is a film that feels familiar and uncanny at once. It is a story of mothers and daughters, a love letter written in code, a memory remade.

When the screen goes dark and the synthetic child fades, what remains is not the novelty of the technique. It is the sincerity of the gesture. Someone once stood in a room, cradled the air and believed there was a baby there. Somehow, now so do we.