Anthropic finds AI's step-by-step reasoning is often unfaithful to its real thinking

Original: Measuring Faithfulness In Chain Of Thought Reasoning

Source: anthropic.com ↗