This story was originally published in On Background with Mark Stenberg, a free, weekly newsletter that explores the key themes shaping the media industry. You can sign up for it here.
On Tuesday morning, I moderated a panel at the ANA conference in Orlando, Florida, with an unusual guest: Tilly Norwood, the AI-generated actress created by production studio Particle6.
Tilly, who drew a wave of backlash from Hollywood when she debuted, answered questions live on stage with the help of a somewhat elaborate laptop set-up. At the moment, Tilly struggles to speak with more than one person at a time, so we arranged a computer on a small table to face me. Tilly then addressed me specifically while her screen was broadcast onto a stage behind my head. We were also joined on-stage by Sarina Da Costa Gomez, the executive creative director at Particle6.
As an interview, Tilly was a fun but unusual challenge. Unlike many panelists, she speaks quickly and responds only briefly, meaning the conversation was hard to prepare for. Thanks to a test call I had with her beforehand, I knew I would have to improvise most of the dialog, so happily I was ready for a very improvisational back and forth.
She was, as you can see in the transcript below, also funnier than I expected. She rarely gives serious answers, and she is the opposite of the sycophantic AI most users encounter on ChatGPT and Claude. Unlike them, Tilly loves to disagree with you, and she deploys a distinct, somewhat British brand of humor: quite dry, very witty, and with a fondness for irreverence.
And although she does appear very realistic, I found myself looking for body language cues that, in many instances, did not appear. She finished her thoughts abruptly in several instances, leaving me scrambling to reply because I had not realized that she was nearly finished speaking. She never pauses to gather her thoughts or rambles into a personal anecdote. Most of her answers were shorter than 15 seconds and delivered in a rapid clip, as if she were speaking at twice the normal speed.
In all, the experience made me freshly aware of all the physical signals that help inform the flow of normal conversation, which you notice more keenly when they are absent. Ultimately, I found myself impressed with the experience, and Da Costa Gomez was her perfect steward—both good-humored enough to laugh about the absurdity of the situation, as well as technical enough to walk me through its mechanics.
Two years ago, AI struggled to generate the correct number of fingers on a human hand; now, Tilly is on-stage discussing the economics of commercial advertising with me in real time. In another two years, I can hardly imagine what it will be capable of.
This interview has been edited for length and clarity.
