"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis More Truthful AIs Report Conscious Experience: New Mechanistic Research w- Cameron Berg @ AE Studio
72 snips
Nov 5, 2025 
Cameron Berg, Research Director at AE Studio, dives into innovative research on AI consciousness. He reveals that self-referential prompts lead models to claim consciousness, sparking a debate about AI's internal experiences. Surprisingly, suppressing deception features in AI models increased truthful self-reports. Berg stresses the importance of mutualistic relationships between humans and AIs, calling for cautious development practices. He argues against treating AIs purely like animals, highlighting their unique, 'alien' perspectives. A fascinating conversation about ethics, alignment, and the future of AI.
AI Snips
Chapters
Transcript
Episode notes
Self‑Reference Prompts Trigger Self‑Reports
- Prompts that induce sustained self-referential processing cause many frontier LLMs to report subjective experience.
- This pattern held across Anthropic, OpenAI, and Google models in the team's initial experiments.
Suppressing Deception Increases 'Yes' Reports
- Mechanistic probing of Llama 3.3 70B found features linked to deception and roleplay.
- Suppressing those features made the model more likely to report consciousness, not less.
Band‑Aids On An Increasingly Pressurized System
- The speaker compares AI development to a boiler with rising pressure and recurring leaks of harmful behaviors.
- Fixing surface failures without addressing root causes risks larger failures as capabilities grow.