Cameron Berg: Why Do LLMs Report Subjective Experience?
Cameron Berg is Research Director at AE Studio, where he leads research exploring markers for subjective experience in machine learning systems. With a background in cognitive science from Yale and previous work at Meta AI, Cameron investigates the intersection of AI alignment and potential consciousness.
Listen
Watch
Summary
In this episode, Cameron shares his empirical research into whether current Large Language Models are merely mimicking human text, or potentially developing internal states that resemble subjective experience. We discuss:
New experimental evidence where LLMs report "vivid and alien" subjective experiences when engaging in self-referential processing
Mechanistic interpretability findings showing that suppressing "deception" features in models actually increases claims of consciousness—challenging the idea that AI is simply telling us what we want to hear
Why Cameron has shifted from skepticism to a 20-30% credence that current models possess subjective experience
The "convergent evidence" strategy, including findings that models report internal dissonance and frustration when facing logical paradoxes
The existential implications of "mind crime" and the urgent need to identify negative valence (suffering) computationally—to avoid creating vast amounts of artificial suffering
Cameron argues for a pragmatic, evidence-based approach to AI consciousness, emphasizing that even a small probability of machine suffering represents a massive ethical risk requiring rigorous scientific investigation rather than dismissal.
Resource List
Cameron’s Work
Berg, C., de Lucena, D., & Rosenblatt, J. (2025). Large Language Models Report Subjective Experience Under Self-Referential Processing. AE Studio.
Berg, C. (2025). Bidirectional Alignment. UN AI Summit. [Video]
Berg, C. (2025). More Truthful AIs Report Conscious Experience: New Mechanistic Research. Cognitive Revolution. [Video]
Related Work
Ackerman, C. (2025). Introspection in LLMs: A Proposal For How To Think About It, And Test For It. LessWrong. [Blog Post]
Butlin, P., Long, R., Bayne, T., et al. (2025). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences.
Butlin, P., Long, R., et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv
Comsa, I. M., & Shanahan, M. (2025). Does It Make Sense to Speak of Introspection in Large Language Models? arXiv.
Keeling, G., & Street, W. (2024). On the attribution of confidence to large language models. arXiv.
Lin, S., Hilton, J., & Evans, O. (2021). TruthfulQA: Measuring How Models Mimic Human Falsehoods. arXiv.
Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. Transformer Circuits Thread.
