Product2 publishers3 min readPublished
Particle6 sent an unscripted AI persona into seventy-five live interviews at once
The one session posted online shows Tilly Norwood misreading a direct question from actor Tom Conti, breaking off mid-word, and speaking Chinese for ten seconds before apologising for a hiccup.
The Product Desk · Product desk
What happened
- TechCrunch reported that Particle6 Group made its AI-generated actress Tilly Norwood available for seventy-five simultaneous interviews with journalists to promote the film Misaligned.
- After Conti repeated the question, Norwood described the other actors as digital twins, broke off in the middle of a word, and then spoke Chinese for more than ten seconds.
- Gizmodo reported noticeable lag ahead of the character's answers through the session, delivered in a flat tone, with responses that were either generic or misguided.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Buyers evaluating live avatar products can stop accepting the scripted demo and ask the vendor to run the repair turn instead: a question the system has already misread once, asked again on camera.
- exposure A persona that comments on its own reliability puts claims on a public feed that nobody at the company drafted or can support.
- constraint Running seventy-five rooms at once rules out a human watching each one, so concurrency and single-turn recovery pull against each other in any rollout plan.
- contradiction TechCrunch says the mistakes ran through every one of the seventy-five sessions, but only a single recording is public, so the breadth of the failure rests on one reporter's account.
The instructive part of the Piers Morgan video comes before the Chinese. Tom Conti asks whether every character in the film Misaligned is AI-generated, and the model answers with a line about the human writers, editors and directors involved [7]. So Conti says it again: "You didn't really understand the question," followed by "Are the other actors on the screen with you real actors, or are they computer-generated images? Do you know?" [8]
In my view the repeat is what a live persona has to survive. A user who has been misread restates the question and expects recovery inside one turn.
On the second attempt the model got the subject right. It said the other actors are "all digital twins just like me" and that Misaligned is "a hybrid production which m-", then stopped and spoke Chinese for over ten seconds, according to TechCrunch [9][10]. Gizmodo, which described the language as Cantonese, reported that another badly timed pause came first [11]. The same report described noticeable lag ahead of answers throughout the session, delivered in a flat tone, with responses that were either generic or misguided [6].
Morgan, looking puzzled, said: "You seem to be speaking Chinese completely randomly for no reason. Why did you do that?" [12] The model apologised, said it had suffered a little hiccup, said it had not meant to start speaking Chinese or to impersonate Piers, and said its wires sometimes get crossed [13]. Gizmodo reported it also insisted that such mixups happen "not often at all!" [14] That reliability claim was the model's own: nobody at Particle6 wrote it, and it has no data behind it.
TechCrunch reported that Particle6 Group had the character available for seventy-five simultaneous interviews with journalists and that it seemed to be making mistakes in all of them [1][2]. One of those sessions was recorded and posted online [18], about 1.3 percent of what was running [17]. At seventy-five concurrent rooms, no supervisor is positioned to take Conti's second question.
Teams tell themselves that users will look at a synthetic presenter, react to how it sounds, work through the press kit. The people in the room asked a factual question twice, then asked the system to account for its own failure. TechCrunch offered the kinder reading, that Particle6 "calculated that people would be more likely to talk about Norwood if they were making fun of it" [16].
Where a live persona belongs comes down to who controls the order of questions and whether a failure costs you a retry or a permanent recording. Fixed order plus retry is where this works today: kiosk loops, product tours, anything you can run again. Unscripted order plus permanent recording needs two things a rehearsal cannot supply: a measured time to first word on the worst question you can write down in advance, and a human who can take the turn while the model is still stalling.
Neither Gizmodo nor TechCrunch reports an explanation from Particle6 of what powers the character [20]. In the session that is public, the model needed Conti's second attempt and still did not finish the sentence [9][10].
What to watch
- Whether Particle6 or the journalists in the other seventy-four sessions publish recordings that would test TechCrunch's claim of mistakes in all of them.
- Whether Particle6 explains the failure mode: the model behind the character, its latency budget, and why the output language flipped mid-sentence.
- Whether the Misaligned promotion keeps live unscripted interviews or moves to pre-rendered answers with a human operator on the line.