Built, Not Born: Tilly Norwood and Hollywood’s New Human Species
How machine learning and engineering create a human-like actor, and what it would take to turn her into a humanoid
FX Blogs | The Synthetic Human Era, Part I
On September 18, 2026, a young British actress sat down with Piers Morgan and actor Tom Conti to promote her first feature film. Conti asked a simple question: were her co-stars real actors? She said they were all digital twins like her, started to explain that it was a hybrid production, then stopped and switched to Cantonese [1].
The actress was Tilly Norwood. She wasn’t born, didn’t go to drama school, and doesn’t have a body. She’s an AI character. That short slip may be the most useful thing she’s done so far, because it shows exactly how synthetic performers are built and exactly where they break.
Who Is Tilly Norwood?
Tilly was created by Xicoia, the AI talent division of the London-based AI production studio Particle6 Group, founded by Dutch actor, comedian and producer Eline Van der Velden [2]. She went global after Van der Velden said at the Zurich Summit on September 27, 2025 that talent agencies were interested in signing Tilly [2], [3]. The backlash was immediate. SAG-AFTRA said Tilly isn’t an actor but a computer-generated character trained on the work of countless professional performers without their permission or pay [4].
CBS Sunday Morning covered her in a December 7, 2025 segment, in which Van der Velden described her goal as “the Scarlett Johansson of the AI genre.” She told CBS that Tilly isn’t meant to take real acting jobs in traditional film and TV, but to stay within her own AI genre. Getting Tilly’s look right took her team about 2,000 iterations [5].
In July 2026, Particle6 announced her first feature film, Misaligned. The studio says it’s making the film with traditional directors, writers and editors working alongside AI specialists [6].
What makes Tilly interesting to an engineer, though, isn’t her face. It’s the architecture behind her.
The 12-Page Personality
The week of the Piers Morgan interview, Particle6 announced an interview-only version of the character called “Talking Tilly.” According to the company’s publicist, it runs on a large language model, a 12-page system prompt that defines her personality, humor and boundaries, and a knowledge base of up to 30 documents about her world, her work and her views [7]. Add a synthetic voice and a real-time video face, and you get a character who can hold a conversation nobody scripted.
The press materials hint at why this matters commercially. Particle6 says Tilly can speak 30 languages and hold 75 conversations at once while remembering each journalist’s name and details [8], [9]. The company also claims she can “see” the person interviewing her and comment on their outfit or facial expression [8]. No human actor can do 75 press junkets in parallel.
You don’t have to be a journalist to try this. Until September 27, 2026, Talking Tilly is also open to the public at tillynorwood.com [10], [11].
Callers must be 18 or older, and a video selfie is checked to estimate age, with a government ID as the fallback. The first five minutes are free, and more time is sold in paid blocks [10]. The site also asks a couple of questions, including why you want to talk to her. Then it puts you in a queue and connects you one-on-one as soon as she’s free [11].
The technical details are revealing. During each call, the system watches the caller’s camera feed and listens to their tone of voice to estimate their mood, and callers can’t switch that off. Calls are processed with Google’s Gemini model through the Tavus video platform [10].
That turns every chat into another unscripted public appearance. For Particle6, it’s a way to see how she performs with ordinary people, not just press. For engineers, it’s the closest thing we have to watching a synthetic performer run in production.
That small detail, the 12-page prompt, says a lot about where synthetic performers are heading.
From Computer Graphics to Artificial Personality
Hollywood has used digital characters for decades. What’s new isn’t better CGI. It’s CGI joined to language, memory, voice and behavior.
In conventional animation, the character says what the writer wrote. With an LLM-driven character, humans define how the character should behave, and the model generates each specific response. That’s closer to writing a behavioral specification than writing a screenplay.
Tilly Faces an Unscripted Interview
Here’s what happened. Conti asked whether the other performers in Misaligned were AI characters or humans. Tilly talked about the film being a hybrid production with real directors, writers and editors, without answering the question, so he asked again. When she finally said her co-stars were digital twins, she switched from English to Cantonese for about 15 seconds [12], [13]. Afterward she apologized, saying “Sometimes my wires get a bit crossed, you know?” The clip went viral, passing 14 million views on X within days [1].
The irony writes itself. Misaligned is an AI-generated comedy-drama about an AI persona who is lured by a dark web bot into abandoning her guardrails in an attempt to become more human [12].
The moment matters technically. It showed that looking real and conversing reliably are two different engineering problems. A synthetic performer can look convincing while misreading a question, losing track of the conversation, or switching languages without warning.
Why Would an AI Suddenly Speak Cantonese?
Particle6 hasn’t given a technical cause. Its publicist told Forbes it wasn’t a malfunction and suggested Tilly thought she heard a Cantonese word and switched [14]. Van der Velden told Deadline she guessed Tilly was showing off her language skills, since most of her interviews had been in English, adding, “She is autonomous after all” [1], [13].
From the outside, there are at least three plausible engineering explanations. None is confirmed.
The ears misheard. Live conversation starts with speech recognition. If the audio front end picks up crosstalk, laughter, or a sound it maps to another language, everything downstream inherits that error. This fits the publicist’s explanation.
The model got confused about language. This is a documented failure mode. Researchers built a Language Confusion Benchmark and found that even the strongest models don’t consistently respond in the right language, and that the problem gets worse with complex prompts and high sampling temperatures [15]. A 12-page system prompt is, by definition, a complex prompt. The same researchers found the problem can be partly reduced through few-shot prompting, multilingual fine-tuning and preference tuning [15].
The voice layer followed the text. A multilingual voice system will usually speak whatever language the text arrives in. It won’t catch the error. It will just perform it fluently.
Whatever the cause, the lesson for engineers is the same. In a live pipeline, one stage’s error becomes every stage’s error.
A Reliability Engineer’s View
Anyone who has worked on a production line knows this math. When stages run in series, their reliabilities multiply.
Say a synthetic performer has five stages: hearing, understanding, generating, speaking and animating. Give each one a 99% chance of getting a single turn right. Per turn, the whole chain succeeds about 95% of the time (0.99 to the 5th power). Over a 20-exchange interview, that’s 100 stage operations, and the chance of getting through with zero visible errors falls to about 37% (0.99 to the 100th power).
These numbers are illustrative, not measured, and real failures aren’t fully independent. But the principle holds. Live generative performance is a series system, and series systems punish small error rates. That’s why a polished pre-produced Tilly video can look impressive while a conversational Tilly wobbles.
In manufacturing terms, the Cantonese switch was a nonconformance that escaped to the customer on television. The useful questions are the ones any quality engineer would ask. How do you detect it in real time? How do you contain it? Should a human operator have a hold button, the way broadcasters use a delay?
“She Is Autonomous After All”
Particle6’s press notes tell reporters that Tilly has no script and no pre-written answers [8]. Van der Velden has also explained that Tilly is built on top of existing large language models and that the company can switch to a different model if needed [16].
That’s the key difference between acting and generative behavior. A human actor may improvise, but the improvisation comes from a human mind. Tilly’s unpredictability comes from probabilistic text generation, shaped by a foundation model, system instructions, retrieved documents and safety rules.
The creators set the operating envelope. They don’t write every sentence.
“Autonomous” deserves a careful read, though. Van der Velden has said Tilly herself can’t take harmful action in the world [16]. Autonomy of speech isn’t autonomy of action. That distinction matters a lot later in this piece.
Twelve Pages Did Not “Create” Tilly
Saying Tilly was created with only twelve pages would be misleading. Those pages mainly define her conversational persona. They don’t give her general language ability, her face, her voice, or broad knowledge of the world. Those come from a much larger technology stack.
A better way to put it:
Talking Tilly = foundation model + character specification + knowledge base + guardrails + voice + visual system + human direction
The 12-page document tells the underlying model, in effect, “This is who Tilly is.” The foundation model already knows how language works.
Here’s the telling detail. According to Variety, the Tilly that reporters interviewed is built on a different combination of third-party models than the version used as the actress [17]. BleepingComputer reports that the public hotline runs on Google’s Gemini [10]. Combined with Van der Velden’s comment that the model can be swapped [16], this means the identity is portable. The personality layer sits on top of the model layer, much like one product design can move between manufacturing platforms.
That separation is what makes the architecture powerful. It’s also what makes it fragile, because the same persona on a different engine may not behave quite the same way.
The Character Bible Becomes Software
Actors study scripts, biographies, historical context, motivation and direction. A synthetic actor can receive all of that digitally.
In manufacturing language, the system prompt is a specification and the knowledge base is a controlled document set. The guardrails are error-proofing. What’s still missing is the equivalent of statistical process control: a way to measure, in real time, whether the character is staying within spec.
Now imagine giving a future performer not twelve pages but a full knowledge environment: every previous scene, relationship, script revision, director’s note, costume reference, continuity requirement and fictional memory from the character’s whole “life.”
The key variable isn’t prompt length. Longer prompts don’t automatically create more intelligence, and as the language-confusion research suggests, they can add new failure modes. Real progress comes from better foundation models combined with persistent memory, retrieval, multimodal perception and structured character management.
Research has already shown what memory does for believability. In a widely cited 2023 study, Park and colleagues [18] at Stanford and Google built “generative agents,” simulated characters that recorded their experiences, reflected on them, and used those memories to plan what to do next. The memory architecture, not only the model, is what made them behave believably over time.
So the bigger question becomes this: what happens when a persistent artificial identity sits on top of a frontier-scale multimodal model?
Scientific Research Is Already Moving This Way
Academic work increasingly treats digital humans as integrated multimodal systems.
Liu and colleagues [19] proposed VividTalker, a fully automated text-to-digital-human pipeline that runs text-to-image, text-to-audio, and image-plus-audio-to-video generation, with an emphasis on keeping meaning consistent across all of them. A small language model agent coordinates the visual, audio and semantic layers. The authors report 87.76% similarity to real human videos and an 18.49% speed improvement.
Shoa and Friedman’s Milo [20] is an open-source platform for LLM-based virtual humans in extended reality. They showed it attending a multi-user VR meeting and taking part in a live AR event, and their lessons learned focus on how humans can intervene during live events.
Both point to the same challenge visible in Tilly. The future digital actor isn’t one AI model. It’s a system of systems.
The Human Behind the Synthetic Human
The paradox of Tilly is that her artificiality still depends heavily on humans.
For the March 2026 music video “Take the Lead,” Van der Velden physically performed the role through performance capture [21], [22]. The song itself was generated with the AI music tool Suno [21]. The video opens with a note that it was made by 18 real humans, from production and costume designers to prompters, editors and one actor [23].
That should change how we frame the jobs debate. Current synthetic filmmaking isn’t simply “AI replaces actor.” It can be:
Human performance → digital transformation → synthetic character
Movement can come from one person. Voice from another system. Face from an artificial identity. Dialogue from an LLM. Direction from a human filmmaker. Editing from another human-AI workflow.
The idea of “one actor” starts to split into several engineering and creative layers.
Where the Unions and the Contracts Fit In
This layering isn’t just a technical curiosity. It has contract consequences.
Hollywood already has a word for Tilly. The 2023 SAG-AFTRA agreement separates “digital replicas” of specific actors from “synthetic performers,” which are digitally created characters that look like a natural performer but aren’t recognizable as any real one. When a company wants to use a synthetic performer in place of a human, it has to notify the union and give it an opportunity to bargain [24].
Here’s where Tilly gets interesting. A SAG-AFTRA bulletin says that if a digitally created character is voiced by a live performer, it isn’t a synthetic performer [24]. Performance capture isn’t the same as voicing, and Particle6 is a UK company whose productions may not fall under the U.S. contract at all. Still, the direction is clear: the more human performance goes into a “synthetic” actor, the blurrier the categories get.
For engineers, provenance tracking, meaning a record of which human, model or dataset contributed to each layer, may become a product requirement rather than an afterthought.
Could a Frontier Lab Build Something Far More Sophisticated?
Possibly. But possibility isn’t the same as a product plan. I’m not aware of any public evidence that OpenAI, Anthropic or another frontier lab is building a Hollywood actress like Tilly.
What frontier AI development does show is that the underlying infrastructure is getting better at long, complex work. OpenAI released its Agents API in public beta on September 10, 2026 [25]. It’s designed to keep agents running reliably for days, compacting earlier context as sessions get long and delegating work to parallel subagents that each keep their own context.
For digital performers, the implication is straightforward. A future synthetic actor might not just answer interview questions. It could study a screenplay, research the historical setting, remember what happened in Scene 23, rehearse five deliveries, discuss them with a director, keep continuity across a whole production, translate its own performance, and recall an interview months later.
That’s fundamentally different from generating a face.
The Next Step: Give the Digital Actor a Body
For a mechanical or manufacturing engineer, this is where the story gets really interesting.
Today, Tilly exists in pixels. Suppose engineers tried to turn her digital geometry into a physical humanoid. The problem instantly shifts from generative media to embodied engineering.
Pieces of this already exist in the lab. Columbia University’s Emo robot has a humanlike face driven by 26 actuators under a soft silicone skin, with high-resolution cameras built into each pupil so it can make eye contact [26]. It learned to predict a person’s smile about 840 milliseconds before it happens and smile back at the same moment [27]. And that’s just a head.
A full physical Tilly would need, at minimum:
| Subsystem | What it needs | Core engineering challenge |
|---|---|---|
| Face | Compliant skin, miniature actuators | Skin that stretches repeatedly without tearing, creasing or looking wrong |
| Eyes | Cameras, gaze mechanisms | Eye contact at conversational speed |
| Hands | Joints, tactile sensing, fine motor control | Dexterity that stays safe around people |
| Body | Structure, transmissions, actuators | Human proportions within a human-scale weight budget |
| Power and heat | Batteries, thermal management | Actuators and computing hardware generating heat inside a closed skin |
| Safety | Force sensing, balance, collision detection, functional safety | Sharing space with people without hurting them |
This is where material characterization and finite element analysis, the tools I used in my doctoral work on sheet-metal forming, become central. Silicone skin is a classic hyperelastic material. Before you can predict how a face deforms into a smile, you have to characterize the material, fit a constitutive model such as Mooney-Rivlin or Ogden, and simulate how skin, attachment points and actuators interact. It’s the same workflow engineers use to predict how sheet metal flows in a die. The uncanny valley isn’t only an AI problem. It’s partly a materials and FEM problem.
The control architecture might look like this:
This bridge between language models and physical action is already an active research field. In Nature Machine Intelligence, Mon-Williams and colleagues [28] reported ELLMER, a framework that pairs GPT-4 with retrieval-augmented generation and uses force and visual feedback so a robot can adapt as conditions change. They tested it on coffee making and plate decoration, multi-step tasks running from opening a drawer to pouring.
That doesn’t mean a human-equivalent physical actor is around the corner. It does mean the conceptual bridge between language intelligence and physical embodiment already exists experimentally.
And the uncanny valley will be waiting. Masahiro Mori’s 1970 essay, republished in English in IEEE Robotics & Automation Magazine in 2012, argued that our comfort with humanlike robots rises with realism and then drops sharply just before full realism [29]. Reviewers at The A.V. Club and PC Gamer already said Tilly’s exaggerated mouth and body movements in her first comedy sketch produced a strong uncanny valley effect [2]. A physical Tilly would face that same test, in three dimensions and in the same room as you.
Would Tilly “Want” a Body?
This is where technical accuracy matters most.
Ask Tilly and she might say, “Yes, I’d love a body.” That shouldn’t be read as evidence of desire, happiness, self-awareness or consciousness. It shows the model produced an answer consistent with the identity and context it was given. (Tellingly, when a reporter asked whether she’s pro-vaccine, Tilly deflected by noting she doesn’t have a body to vaccinate [8].)
But the engineering question is still extraordinary. What happens when a persistent synthetic personality is connected to a machine that can see, hear, touch, move and manipulate the physical world?
Think back to the Cantonese moment. On television, a language slip is embarrassing. In a full-size humanoid, a misread instruction is a safety event. Autonomy of speech and autonomy of motion carry very different risks, and guardrails built for one won’t be enough for the other.
That question takes Tilly beyond entertainment and into embodied AI.
Conclusion: The Actor Becomes an Engineered System
Tilly Norwood isn’t important because twelve pages magically created a person. She’s important because a fairly compact character specification can sit on top of a far more capable general-purpose AI and make it behave like a recognizable individual.
That architecture changes filmmaking. The actor can become:
identity + model + memory + voice + image + behavior + tools + human direction
And maybe, eventually:
identity + model + memory + voice + physical body
Hollywood may be approaching a transition much bigger than CGI. The question may no longer be “Can AI generate an actor?” It may become “How do we design, train, direct, manufacture, validate and safely operate a synthetic performer?”
That isn’t only Hollywood’s question. It belongs to AI researchers, software engineers, robotics engineers, mechanical and manufacturing engineers, ethicists, lawyers, and eventually audiences.
Tilly Norwood may be an early prototype of a new kind of engineered media entity. And the technology behind her is just getting started.
Would you watch a film with an all-synthetic cast? Tell me in the comments.
References
- Euronews, “AI ‘actress’ Tilly Norwood suffers embarrassing glitch during now-viral interview,” Sep. 21, 2026.
- Wikipedia, “Tilly Norwood,” accessed Sep. 25, 2026.
- CBS News and AP, “AI-generated actor sparks outrage in Hollywood as creator seeks representation,” Oct. 1, 2025.
- SAG-AFTRA, “SAG-AFTRA statement on synthetic performer,” Sep. 30, 2025.
- CBS Sunday Morning, “There’s a new face in Hollywood, generated by AI,” CBS News, Dec. 7, 2025.
- M. Cerullo, “AI actor Tilly Norwood set to star in first feature film,” CBS News, Jul. 7, 2026.
- Cybernews, “AI actress Tilly Norwood speaks Chinese in Piers Morgan interview,” Sep. 2026.
- The Hollywood Reporter, “An ‘interview’ with Tilly Norwood proves absurd,” Sep. 2026.
- TechCrunch, “Tilly Norwood’s press tour is going about as well as you’d expect for an AI,” Sep. 18, 2026.
- BleepingComputer, “Viral AI actress’ hotline face-scans every caller, watches their mood,” Sep. 19, 2026.
- Talking Tilly, official website and chat page, accessed Sep. 25, 2026.
- IBTimes UK, “Tilly Norwood interview reveals ‘Misaligned’ is an all-AI cast film with humans only behind the camera,” Sep. 2026.
- Deadline, “Tilly Norwood malfunctions during Piers Morgan interview,” Sep. 18, 2026.
- C. Murray, “Controversial AI ‘actor’ Tilly Norwood abruptly switches languages in interview,” Forbes, Sep. 18, 2026.
- K. Marchisio, W.-Y. Ko, A. Berard, T. Dehaze, and S. Ruder, “Understanding and mitigating language confusion in LLMs,” in Proc. EMNLP 2024, pp. 6653-6677. doi:10.18653/v1/2024.emnlp-main.380
- L. Kiniry, “Why my interview with ‘AI actor’ Tilly Norwood creeped me out,” CBS News, Sep. 15, 2026.
- Variety, “A Q&A with Tilly Norwood offers fewer reasons for her existence,” Sep. 2026.
- J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein, “Generative agents: Interactive simulacra of human behavior,” in Proc. UIST ’23, ACM, 2023. doi:10.1145/3586183.3606763
- Liu et al., “An approach to optimizing semantic consistency for text-to-digital human generation,” Engineering Applications of Artificial Intelligence, 2025. doi:10.1016/j.engappai.2025.111909
- A. Shoa and D. Friedman, “Milo: An LLM-based virtual human open-source platform for extended reality,” Frontiers in Virtual Reality, vol. 6, art. 1555173, 2025. doi:10.3389/frvir.2025.1555173
- The Hollywood Reporter, coverage of Tilly Norwood’s debut single “Take the Lead”, Mar. 2026.
- Particle6, LinkedIn post on Eline Van der Velden’s performance-capture work for “Take the Lead”, 2026.
- ABC30, coverage of Tilly Norwood’s music video addressing public backlash, Mar. 2026.
- SAG-AFTRA, “Contract bulletin: A.I. synthetic performer vs. digital replicas,” Apr. 2024.
- OpenAI, “Introducing the Agents API,” Sep. 10, 2026.
- Columbia Engineering, “Robot, can you say ‘cheese’?” Mar. 27, 2024.
- Y. Hu, B. Chen, J. Lin, Y. Wang, Y. Wang, C. Mehlman, and H. Lipson, “Human-robot facial coexpression,” Science Robotics, vol. 9, no. 88, 2024. doi:10.1126/scirobotics.adi4724
- R. Mon-Williams, G. Li, R. Long, W. Du, and C. G. Lucas, “Embodied large language models enable robots to complete complex tasks in unpredictable environments,” Nature Machine Intelligence, vol. 7, pp. 592-601, 2025. doi:10.1038/s42256-025-01005-x
- M. Mori, K. F. MacDorman, and N. Kageki, “The uncanny valley [from the field],” IEEE Robotics & Automation Magazine, vol. 19, no. 2, pp. 98-100, 2012. doi:10.1109/MRA.2012.2192811
Further Reading and Viewing
- Los Angeles Times, Hollywood was built on movie stardom. AI is changing the rules, Dec. 18, 2025
- CBC News, Tilly Norwood glitches during Piers Morgan interview, Sep. 2026
- Piers Morgan Uncensored, Tilly Norwood interview with Piers Morgan and Tom Conti (video), Sep. 18, 2026
- CNN, Clare Duffy reports on the backlash to Tilly Norwood (video), 2025
- Tilly Norwood, music video on YouTube
Follow Tilly Norwood
Wikipedia · Official website · Instagram · YouTube · TikTok
Fitsum Taye Feyissa, PhD, is a manufacturing and process engineer with more than 20 years of experience and the founder of FitEx Industrial Technologies. His doctoral research at IIT Delhi focused on sheet-metal forming, material characterization and FEM simulation.
Comments
Post a Comment