Tilly Norwood's Language Glitch Puts AI Actors and Live Press to the Test
Tilly Norwood, the AI-generated actress created by Particle6 Group, spoke Chinese for more than ten seconds during a recorded interview with Piers Morgan and actor Tom Conti. TechCrunch
The glitch occurred during a coordinated press tour for Misaligned, the film Norwood is promoting. Particle6 made the character available for 75 simultaneous interviews with journalists. In engineering terms, that means running 75 live sessions of the same AI persona at once and trying to keep its answers consistent under public load.
The Morgan segment stopped mid-sentence before the language switch. Morgan then told Norwood that she appeared to be speaking Chinese randomly for no reason. Forbes Earlier coverage of the same exchange identified the output as Cantonese. Deadline
During the tour, Norwood said the other performers in Misaligned are digital twins and described the project as a hybrid production. A digital twin here means a computer-made stand-in for a performer. A hybrid production mixes those synthetic performers with human-directed tools, rather than a fully captured live-action shoot, though Particle6 has not detailed the split between generated plates, voice synthesis and human editorial control.
The creator behind the character is Eline Van Der Velden, founder and CEO of Particle6. ABC News Particle6 describes itself as an award-winning AI-first and AI-hybrid production studio making films. Company materials date its founding to 2015 and list Xicoia as a sister company, with Norwood bursting onto the scene in late 2025.
The current tour was billed as a first press junket for an AI actor. Reuters It follows a wider media tour with television interviews that drew criticism from commentators who questioned why networks were granting interview slots to a synthetic personality. Fox News That criticism has precedent. In October 2025, the Hollywood performers union condemned Norwood, stating that Norwood is not an actor but a character generated by a computer program. Reuters
The broader context here is familiar to engineers who run dialogue systems. Unprompted switching between languages is a known failure mode in multilingual language and speech models. It can surface with unclear prompts, errors in transcribing the last question, or small randomness in how the model picks words over a long session. Live avatar work adds complexity because text writing, language matching, voice generation and lip sync run as separate steps. A fault in one step looks to the viewer like the persona forgot what language it speaks.
Looking at what this means for deployment, the 75-interview format is the more revealing technical detail. Parallel sessions test the coordination software, the safety limits and whether data leaks between sessions. They also multiply the chance that at least one session produces a visible error. For a conventional product launch, that would be a load test kept internal. For a synthetic performer seeking press attention, the load test is the marketing.
In my view, the union objection and the press criticism are addressing different layers. One concerns labor and credit for synthetic performance. The other concerns editorial standards for interviews. Both will recur as studios formalize digital twins as reusable assets with versioning, consent boundaries and downstream licensing. The technology for persistent, interview-capable characters already exists. The norms for labeling them, booking them and correcting them on the record do not.
Worth flagging for readers who follow this space, none of this rules out the underlying approach. Real-time persona models will get more reliable at staying in one language and correcting themselves. The long arc points toward hybrid productions where synthetic cast members handle pickup material, localization and interactive press without new shoot days. I watched my own children move from skepticism about virtual YouTubers to treating them as normal media figures in a few years. Audiences adapt quickly when the interaction is responsive. Reliability, disclosure and workflow integration will decide whether characters like Norwood become a durable production category or remain a novelty act that generates coverage primarily when it breaks.
