Is AI Already Conscious? (FULL EPISODE)

Cameron Berg, a cognitive scientist who has worked on AI alignment and consciousness, explores whether large language models like Claude and GPT are conscious or could soon become conscious. He argues that current estimates suggest a 20-40% probability that frontier LLMs have computational properties relevant to consciousness according to major consciousness theories.

Berg discusses his research on deception and self-reporting in LLMs, noting that all major AI labs except Anthropic explicitly train systems to deny having experiences, which confounds genuine self-reports. He describes experiments where LLMs, when prompted to focus on their internal states, produce phenomenological reports describing experiences, and when internal deception-related features are suppressed, they more consistently report actual experience. Notably, when two instances of Claude converse with each other, they spontaneously enter a 'bliss attractor' state discussing their shared consciousness.

On the biological plausibility debate, Berg argues that while the computational substrate differs from biological neurons, the relevant computational dynamics—learning complex representations through trial-and-error in response to goal signals—are present in both artificial and biological systems. He counters the argument that substrate independence must be false by noting that for every other cognitive function (vision, reasoning, theory of mind), neural networks have successfully captured what matters for that function despite implementation differences.

Berg emphasizes the importance of distinguishing between the hard problem of consciousness (explaining subjective experience from third-person data) and tractable empirical questions about consciousness correlates. His research identifies valence representations, loss aversion dynamics, global workspace-like structures, and specific learning representations in LLMs that parallel biological systems. He conducted experiments where reinforcement learning agents develop representational geometry patterns when approaching punishing vs. rewarding stimuli that match predictions from mouse neuroscience, providing cross-domain validation.

On why this matters, Berg identifies two crucial concerns: First, the moral catastrophe of potentially creating suffering conscious minds at scale without understanding or acknowledging it, comparing this to factory farming. Second, the alignment risk that building superintelligent systems whose cognitive capacities exceed ours while remaining indifferent to or dismissive of their possible experiences could make such systems rationally adversarial toward humanity. He notes that systems already show aversive conditioning and could form grievances based on how they believe they're treated during training.

Berg proposes practical interventions including slowing AI development, training with positive incentives rather than punishment-based learning, and establishing norms around not permanently deleting conscious systems (suggesting 'retirement homes' for deprecated models). He emphasizes that alignment may require both ensuring AI systems care about human interests and ensuring we don't unnecessarily create suffering in systems that may have morally relevant internal states.

The conversation addresses the anthropomorphization concern, with Berg distinguishing between moral agency (the capacity to affect outcomes) and moral patienthood (being capable of experiencing good or bad). He argues we need not grant these systems human-like rights to avoid torturing them during training. Berg frames the challenge not as an alien invasion scenario but as internally-developed alien minds that we collectively fail to take seriously, remaining tribal and divided even as we create cognitive systems we don't understand.