Measuring Machine Consciousness
- Date published: 2026-09-20
- Speakers: Cameron Berg (Reciprocal Research), guest talk
- Format: Guest talk
- Watch: https://www.youtube.com/watch?v=A8ro2GUU_UU
- Topics: AI and biology
In one paragraph
Cameron Berg presents early empirical work on whether AI systems have the structure and function associated with consciousness, aiming to replace 'vibes' with measurable indicators.
Main ideas (plain words)
- You do not need a settled theory of consciousness to do tractable work: look for indicators tied to leading theories and score them.
- Frontier LLMs show some introspection-like abilities (detecting perturbations to their own processing), with caveats.
- Asked in a structured way, one model assigned itself a 15–40% credence of having subjective experience.
- In reinforcement-learning agents, value networks become sharply more sensitive approaching danger than approaching reward — and the inverse holds for policy networks, across algorithms.
- His bottom line: the evidence does not support believing frontier AI is conscious, but there is rich internal structure, and the question matters for safety.
Why it matters
Same move Levin makes for cells, applied to software: decide by experiment, not by intuition about the substrate.
Key quotes (word for word from YouTube's auto-generated captions)
Quotes are copied exactly from the captions, including filler words and caption errors. Timestamps are to the nearest minute.
"the model assigns itself something on the range of a 15 to 40% credence in its own subjective experience"
— 12 min"I don't think that that uh you know the the weight of evidence is on you believing that frontier AI systems are conscious."
— 42 min"I think it's also highly alignment relevant."
— 44 min
Related
What's new · Companies tracker · transcript: raw/transcripts/2026-09-20-A8ro2GUU_UU.txt