Why LA Comic Con thought making an AI-powered Stan Lee hologram was a good idea

https://arstechnica.com/feed/ Hits: 74
Summary

"The only words that are gonna be in Stan's mouth are Stan's words," DeMoulin said. "Just because I haven't personally seen [the model hallucinate] doesn't mean that it's impossible, but that hasn't been my experience." The living version of Stan Lee appeared at the Wizard World convention in 2018, shortly before his death. Credit: Getty Images The living version of Stan Lee appeared at the Wizard World convention in 2018, shortly before his death. Credit: Getty Images While a moderator at the convention will be on hand to repeat fan questions into a microphone (to avoid ambient crowd noise from the showfloor), DeMoulin said there won’t be any human filtering on what fans are allowed to ask the Lee avatar in the 15- to 20-minute group Q&A sessions. Instead, DeMoulin said the team has set up a system of "content governors" so that, for instance, "if you ask Stan what he thought of the last presidential election he's gonna say 'That's not what we're here to talk about. We're here to talk about the Marvel universe.'" For topics that are Marvel-related, though, the AI avatar won’t shy away from controversy, DeMoulin said. If you ask the avatar about Jack Kirby, for instance, DeMoulin said it will address the "honest disagreements about characters or storylines, which are gonna happen in any creative enterprise," while also saying that "‘I have nothing but respect for him,’ which is I think largely what Stan would have said if he was asked that question." Hyperreal’s Scott said the company's approach to training digital avatars on verified content "ensures responses stay true to Stan's documented perspectives and values." And DeMoulin said the model is perfectly willing to say when it doesn’t know the answer to an appropriate question. In early testing, for instance, the avatar couldn’t answer a question about the Merry Marvel Marching Society, DeMoulin said, because that wasn’t part of its training data. After a subsequent update, the new model provided a relevant answe...

First seen: 2025-09-27 11:22

Last seen: 2025-09-30 12:37