Podcast Atlas
concept Updated 2026-08-24 Topics: Technology

AI Public Likeness Generation

EP277 对话贾樟柯（下）：我没有背叛真实世界，我只是在寻找电影的新可能 adds a film-production example through 贾樟柯. Jia describes AI generating multiple versions of a “李老师” figure for a team to choose, then notes that the generated person is not the real teacher and that portrait-right issues immediately appear.

AI public likeness generation is the risk pattern where an AI product can use a public profile, username, or visible identity surface to generate images in a person’s likeness. OpenAI’s GPT-5.6 release raises questions about White House control over new models adds the concept through Meta’s Muse Image, which the episode says can create photorealistic images on Instagram and WhatsApp from prompts involving public usernames unless the user opts out.

The source’s key governance problem is default direction. Opt-out controls put discovery, settings navigation, and burden on users whose likeness may be used, while opt-in controls would require affirmative consent before generation. The episode also notes a child-safety boundary: Meta would say the feature is automatically off for users under 18.

E234｜未来实拍电影还存在吗？与导演陆川聊聊AI给影视人的恐惧与自由 extends the concept from public profile images into entertainment faces and voices. The episode’s Hollywood AI-video opening and 黄英 / Huang Ying’s account of voice contracts show that likeness governance includes actors, dubbing performers, voice clones, blended voices, and the technical proof needed to show unauthorized use.

Key Claims

  • Film production raises likeness issues before release when generated candidate faces, bodies, or performances are used during development.
  • Public-profile availability does not automatically mean people expect their likeness to become a generative prompt target.
  • Opt-out privacy settings can be hard to find, especially if users must reinstall or update an app before seeing the relevant toggle.
  • Likeness generation inside social and messaging apps can spread faster than generation inside a standalone creative tool.
  • Age-based defaults help, but they do not settle consent, adult privacy, impersonation, or harassment concerns.
  • Likeness controls connect synthetic-media provenance to product defaults, not only watermarking or labeling after creation.
  • Voice likeness makes consent and proof harder because a generated voice can be close, blended, or denied by the party using it.

Connections