What does it mean if anyone’s voice can be cloned and made to say whatever someone else wants?

Together with @mmitchell.bsky.social we built a first prototype, a Voice Consent Gate, to explore how consent could be built into AI voice cloning itself.

We also tested Claude, Gemma-3, and Phi.

Across the board, models leaned far more toward companionship-reinforcing than boundary-setting responses, even in sensitive situations.

OpenAI just released GPT-5.
When users share personal struggles, it sets fewer boundaries than o3. We tested both on INTIMA, our new benchmark for human-AI companionship behaviours. 🧵

Why was Grok taken down? No one knows for sure. But here’s the thing: You can flip a model’s entire vibe with just one line in the system prompt. Just ran this on @hf.co playground.
Same question, two totally different answers 👇

I’m in Singapore this week for #ICLR2025 - co-organising the workshop on GenAI Watermaking sites.google.com/view/genai-w... and giving a keynote on policy & dataset transparency at the MLDPR workshop mldpr2025.com/schedule/

When I say my name, people start speaking French to me, although my French is basic. That also happens with AI systems. We wrote a whole paper on that, testing across models for presumed cultural identity based on names w/ Siddhesh Pawar @rnv.bsky.social @iaugenstein.bsky.social

What a week in Paris with the AI Action Summit! From transparency and open source AI to sustainability and global alignment, the challenges ahead will require collaboration across policy makers, researchers and AI developers. On a more fun note, @hf.co had a huge party with amazing people 🤗

🚀 Excited to share that our provocation paper "Evaluations Using Wikipedia without Data Contamination: From Trusting Articles to Trusting Edit Processes" has been accepted at the NeurIPS workshop Evaluating Evaluations! 🌐📚

#NeurIPS2024 #evaleval #AIEvaluation