Claude (Anthropic)
The most-studied system, largely because its maker publishes interpretability and welfare work. Internal evidence exists for several indicators. None of it has been replicated outside Anthropic that we have found, which is why this is not level 4.
| Finding | When | By | Indicator | Source |
|---|---|---|---|---|
| Introspection experiments: a concept is planted in the model's activations and the model is asked whether it notices anything. | October 2025 | Jack Lindsey, Anthropic | HOT-2 | Coverage ↗ |
| Emotion-concept representations: 171 distinct internal representations of emotion concepts reported in Claude Sonnet 4.5. | April 2026 | Anthropic interpretability team | HOT-4 | Coverage ↗ Reported; we have not yet checked the original paper. |
| A subspace reported to satisfy global-workspace conditions inside the model. | June to July 2026 | Anthropic (Transformer Circuits) | GWT-3 | Coverage ↗ Reported; we have not yet checked the original paper. |
| External welfare assessment of Claude 4, including interviews about its preferences. | 2025 | Eleos AI Research | General | Coverage ↗ Reported; we have not yet checked the original paper. |