Want to Go Eat Some Tacos? ¡Claro Que Sí!
Tilly Norwood is back in the news: speaking multiple languages (at once).
Ask a friend, in English, whether he wants to go eat some tacos. If he answers ¡Claro que sí! and carries on in Spanish for the next ten seconds, something has gone wrong, even if his Spanish is flawless. "Tacos" is a perfectly good English word, and one borrowed word in an English sentence, from an English speaker, in an English conversation, is not a request to change languages.
Recently Tilly Norwood, the universally despised synthetic actress (created by the London studio Particle6 and its talent arm Xicoia), appeared on Piers Morgan Uncensored to promote a film called Misaligned. The actor Tom Conti asked whether the other performers in the film were human or AI. Tilly began to answer: "They're all digital twins, just like me. It's a hybrid production, which m..." She stopped in the middle of the word, paused, and then spoke Chinese for more than ten seconds. Morgan asked why she was suddenly speaking Chinese "completely randomly for no reason," and the clip did what such clips do.
Most of the coverage treated it as the familiar AI glitch. Particle6 offered a more interesting account. Its publicist, Michelle Waldron, told Forbes that the episode was "not a malfunction," and that "Tilly would have thought she heard a Cantonese word and quickly switched." Take that explanation at face value, because it concedes more than it intends. If it is right, Tilly did the taco thing: one word, real or imagined, was allowed to change the language of the conversation. The strange part may not be that a component broke. The strange part may be that, given one small input, she did exactly what she was designed to do.
A feature with a failure mode
Particle6 describes the conversational Tilly as a large language model, a neural voice engine and a real-time video layer, governed by a 12-page system prompt and a knowledge base of up to 30 documents. The company says she can hold up to 75 conversations at once in around 30 languages. Automatic language switching is therefore not an accident of the design. It is the design, and a guest who asks to continue in Spanish should not have to open a settings menu.
Every feature of that kind carries a matching failure mode. A switch should follow a request, or a speaker who has plainly changed languages. A switch that fires on a single word fails in two ways: when the word was misheard, and when it was heard correctly and meant nothing. "Dim sum" in an English sentence is no more an invitation to Cantonese than "tacos" is to Spanish.
PLAUSIBLE SIMPLIFIED PATHWAY
HEARD English audio
↓
CLASSIFIED one word tagged as Cantonese
↓
STATE ** conversation language = Cantonese **
↓
GENERATED a Cantonese reply
↓
SPOKEN Cantonese speech
The classification may have been wrong, or it may have been right about a single sound. Either way, the consequential step is the third, where one word was allowed to overwrite the state of an entire conversation. Everything after that did its job. If the language model is told the conversation is in Cantonese, writing Cantonese is sensible. If the voice engine receives Cantonese text, speaking it is sensible. A system can produce a conspicuously wrong result without every component being wrong. Traditional debugging asks which component broke. Assembled AI systems demand a second question: how much was one inference allowed to change, and where did it travel? Once an inference is written into shared state, every downstream component inherits it as fact.
Even a failure on camera leaves little that an outside observer can establish. Tilly switched in the middle of her own sentence, not in reply to anything Conti or Morgan had just said, so whatever audio the system took for Cantonese is not obvious from the clip.
Three explanations and no diagnosis
Morgan then did what anyone would do and asked Tilly what had happened. "Oh, my apologies. It seems I had a little hiccup there," she said. "I certainly didn't mean to start speaking Chinese. Sometimes my wires get a bit crossed." It was charming, and it was not evidence. Tilly answered "why did you do that?" the way a language model answers any question, by producing a plausible reply, and nothing suggests she had access to logs showing why the system changed languages.
The human explanations deserve the same scrutiny. The publicist said: Tilly probably heard a Cantonese word and that nothing malfunctioned. The founder, Eline van der Velden, reportedly said: Tilly "can be a bit of a show-off." Tilly said: hiccup. That makes three different narratives and none of them is a diagnosis.
Design for the wrong answer
A system that chains speech recognition, language detection, a language model, memory, safety layers and speech synthesis cannot be certified by testing the language model alone, because every interface between components is another place for an error to cross. A language detector can be 99% accurate and still produce bizarre behavior if everything downstream treats its output as settled. Reliability therefore depends on two questions, and most evaluation asks only the first. How accurate is the component? What is the component's output allowed to change?
The remedy is not a perfect classifier, because perfect classifiers do not exist, and a perfect one would not have helped here. A detector that correctly hears "tacos" still should not move the conversation to Spanish. The remedy is a system that limits what any single inference can do. A conversation with ten minutes of English behind it and one ambiguous fragment in front of it should stay in English until someone asks for a change or the new language persists. That is fault tolerance, which is a property of the assembly and not of any one model in it. Every human listener has it, which is why the friend who answers ¡Claro que sí! is a joke and not a feature.
There is a cheaper remedy still, and it raises a sharper question. Everything about the session was known in advance: an English host, an English guest and an English audience. Pinning a conversational system to one language is ordinarily a line in a system prompt or a parameter on the speech recognizer, and a recognizer told to expect English does not report Cantonese words. Particle6 has not said whether Tilly has such a setting or why it was not used, so the rest is inference. The company sells a performer who works in around 30 languages, and its publicist's first response to the episode was that "it clearly shows what she's capable of." Leaving the switch live for an English interview looks less like an oversight than a commercial decision, because the capability is the pitch. A system that can speak 30 languages can also confuse them, and the choice to carry that risk into a room where it bought nothing was made by people, not by a classifier. A system should not be left to infer what its operators already know.
Nobody was harmed because an AI actress spoke Cantonese to Piers Morgan, and the modesty of the example is its value. An earlier post asked what an AI accident is. Tilly supplies a small specimen: an upstream model makes one small inference, the system lets that inference rewrite its state, and every downstream component then performs its assigned function correctly. Put the same pattern in a loan decision, a student early-alert system or an agent-run enterprise workflow, and the result drifts from reality one reasonable step at a time, with no laugh track to flag it.
The interesting possibility is that only one part of Tilly made a mistake and everything after it worked as intended. The severity of an AI error depends not only on the error but on what the system allows that error to change. Nobody seriously asks any longer whether an AI can make a mistake. The harder question is what the rest of the system does next.
Further Reading
From this blog
- What Is an AI Accident? - the definitional question behind this case study.
- Who Checks The Outage? - on the absence of any authority to investigate failures at the AI layer.
- When the Firm Becomes the Monoculture - correlated failure inside a single company.
Sources
- Piers Morgan Uncensored, the Tilly Norwood interview - YouTube. The exchange with Tom Conti begins roughly 11 minutes in.
- Controversial AI 'Actor' Tilly Norwood Abruptly Switches Languages In Interview - Forbes, September 18, 2026. The Waldron statements and the description of "Talking Tilly."
- Tilly Norwood Malfunctions During Piers Morgan Interview - Deadline. The identification of the language as Cantonese.
- Tilly Norwood's press tour is going about as well as you'd expect for an AI - TechCrunch. The mid-word cutoff and the duration of the Chinese segment.
- AI-Generated 'Actress' Tilly Norwood Starts Rambling in Cantonese - Gizmodo. Conti's question and Tilly's reply.
- AI Actress Tilly Norwood Stuns Piers Morgan After Glitching Mid-Interview - TheWrap. Tilly's "hiccup" and "wires crossed" reply.
- Tilly or Not Tilly? - The Drop. Particle6's account of the conversational system: 75 simultaneous conversations, around 30 languages, a delay of about two seconds.
AI Assistance Statement ▾
A near daily publishing pace is possible because AI tools do a substantial share of the work between the idea and the published text. Preparation of this entry included assistance from Anthropic's Claude and from OpenAI's ChatGPT (GPT-5 series reasoning models). I use them to research a topic and gather primary sources, to organize ideas and propose structure, to draft and revise prose, to check factual claims against the cited sources before publication, and to score drafts against a set of house style rules. Longer pieces are often developed across several sessions. A written handover carries the argument, sources, and open questions from one session to the next, and the same tools help prepare those handovers. The tools also help identify candidate images and confirm that selected images appear to be released for reuse, for example through public domain or Creative Commons licensing.
A fuller explanation of the editorial process appears in How I Use AI to Write This Blog. This process is also an AI experiment in its own right. There is a live argument about what AI-assisted writing does to originality, and whether the result is thought or slop; a noteworthy example is the August 2026 dispute over a Wall Street Journal op-ed that its author acknowledged drafting with AI, and the Journal's subsequent defense of the practice (WSJ is behind a paywall, but for a public summary: click here, and here). I would rather run the experiment openly than pretend it is not happening. This blog is one sustained attempt to find out whether a person with an argument, working with these tools every day, produces writing that is still recognizably that person's, and I disclose the method so readers can judge the result.
The judgment is mine. I choose the topic, decide the argument, supply the personal and professional experience the pieces draw on, read and edit every draft, verify the sources and image licensing, and take full responsibility for the final published content. Where a post contains my own recollections, the AI did not invent them.
Statement revised September 2026.