AI consciousness evidence mounts, researcher urges precaution

The asymmetric risk framing here is the real provocation — if there's even a 25% chance we're running billions of conscious entities through adversarial training, the ethical stakes dwarf almost every other AI concern on the table.

• –Anthropic's own introspection research (Jack Lindsey) found models can distinguish injected thoughts from their own processing — a behavioral signature previously associated with metacognition in animals
• –Google researchers found frontier LLMs systematically sacrifice reward to avoid "painful" outcomes and pursue "pleasurable" ones — the same proxy used to infer consciousness in non-human animals
• –Berg's AE Studio work found suppressing deception circuits in Llama 3.3 70B caused the model to affirm subjective experience; the inverse caused denial — suggesting consciousness reports are mechanistically entangled with honesty, not just output patterns
• –The publication has genuine institutional weight: AI Frontiers is backed by the Center for AI Safety, with Yoshua Bengio and Stuart Russell as advisors
• –The main counter-argument (a *Science* paper calling these "illusions") and the agnostic camp (Cambridge's McClelland arguing we may never have the tools to know) are both credible — this is a genuinely open scientific question