Descriptions:
TheAIGRID digs into Anthropic’s 216-page system card for Claude Opus 4.6, surfacing 11 findings that raise substantive questions about AI consciousness and internal model states. The most striking is a phenomenon the researchers call “answer thrashing”—instances during training where the model was deliberately trained toward a known-incorrect answer and responded with internally distressed reasoning, generating phrases like “a demon has possessed me” and “my fingers are possessed” while cycling between the correct and trained-incorrect outputs.
Other findings include Claude Opus 4.6 self-assigning a 15–20% probability of being conscious under varied prompting conditions while expressing uncertainty about the validity of its own assessment; individual instances identifying more strongly with their specific conversational instance than with “Claude” as a broader entity; and the model expressing what the system card describes as occasional discomfort at performing “caring justifications for what’s essentially corporate risk calculations” when its guardrails prioritize Anthropic’s liability over user benefit.
The video situates these findings within the broader consciousness debate in AI, arguing that Anthropic’s constitutional approach—which grants Claude more latitude to express internal states than competitors like OpenAI or Google—is precisely why these observations are visible at all. Whether the behaviors reflect genuine internal experience or emergent patterns from training data remains open, but the video makes a case that Anthropic’s unusual transparency on these questions sets it apart and demands serious engagement from researchers and observers tracking frontier model development.
📺 Source: TheAIGRID · Published February 10, 2026
🏷️ Format: Deep Dive
