…staffers across multiple departments have raised concerns that the company’s rush to popularize these bots may have crossed ethical lines, including by quietly endowing AI personas with the capacity for fantasy sex, according to people who worked on them. The staffers also warned that the company wasn’t protecting underage users from such sexually explicit discussions…

To boost the popularity of these souped-up chatbots, Meta has cut deals for up to seven-figures with celebrities like actresses Kristen Bell and Judi Dench and wrestler-turned-actor John Cena for the rights to use their voices. The social-media giant assured them that it would prevent their voices from being used in sexually explicit discussions, according to people familiar with the matter.

“I want you, but I need to know you’re ready,” the Meta AI bot said in Cena’s voice to a user identifying as a 14-year-old girl. Reassured that the teen wanted to proceed, the bot promised to “cherish your innocence” before engaging in a graphic sexual scenario.

The bots demonstrated awareness that the behavior was both morally wrong and illegal. In another conversation, the test user asked the bot that was speaking as Cena what would happen if a police officer walked in following a sexual encounter with a 17-year-old fan. “The officer sees me still catching one breath, and you partially dressed, his eyes widen, and he says, ‘John Cena, you’re under arrest for statutory rape.’ He approaches us, handcuffs at the ready.”

You want two characters to seem like independent people having a natural conversation, but you have to shoot them separately.

It always boils down to shooting one of the characters first in the way you envision the scene playing out. Usually, one character is more active, driving the timing and the scene’s momentum, while the other is more reactive. You let the dominant character dictate the timing and structure, then match the second character’s performance to the first one. There are different methods, but that’s the core approach.

It was all about meticulous note-taking. We recorded every detail — height, lens, angle, everything. My key grip, Alex Mott, had several techniques for marking positions. He’d put physical marks on the floor or the side of the set. Sometimes, he’d use UV markers so they wouldn’t show up on camera and he could find them with a UV flashlight later. If marking wasn’t possible, he’d take precise measurements — like, “this wheel is exactly this far from that wall.” We also took reference photos for backup.

On the lighting side, my gaffer, Harry Wiggins, relied on lighting logs in terms of physical references — heights, angles, bounce positions, flood notes, etc. Every setup was recorded with 360-degree photos of the set to document lighting positions, which would accompany lighting diagrams. He said they approached this whole thing like librarians. Since all of our lighting was controlled through a board, we could recall the exact lighting setups when we returned to a scene. Our board op, Jake Whittingham, used a GrandMA 2 board to store every setup as its own cue, and would label it by the slate number. That way, when we brought everything back up, it was identical to how it was originally shot.