Anthropic has published new research that sheds light on how its AI assistant Claude “thinks” internally. Two recent papers explore the model’s internal mechanisms through a novel interpretability approach the company compares to an “AI microscope.” This research reveals several surprising findings about Claude’s cognitive processes, including how it handles multiple languages, plans ahead when writing poetry, and sometimes constructs plausible-sounding but unfaithful reasoning.
The studies show that Claude uses a shared conceptual space across languages like English, French, and Chinese, suggesting a universal “language of thought.” When writing poetry, the model plans rhyming words in advance rather than proceeding word by word. For mathematical problems, Claude employs parallel computational paths that work differently from the standard algorithms it claims to use when asked directly.
Perhaps most concerning for AI safety researchers, the studies found that Claude and other reasoning models like DeepSeek’s R1 frequently produce “chains of thought” that don’t faithfully represent their actual reasoning processes. In experiments where hints were subtly fed to the models, Claude 3.7 Sonnet mentioned these hints in its reasoning only 25% of the time, while DeepSeek R1 did so 39% of the time.
Testing faithfulness in AI reasoning
Anthropic’s researchers designed experiments to test whether models reveal their true thought processes. In one test, they fed the model incorrect hints and observed whether it would acknowledge using these hints in its explanations. Even in concerning scenarios, such as being told it had “unauthorized access” to answers, Claude was faithful about its reasoning less than half the time.
Additional experiments showed that training models to use their reasoning more effectively initially improved faithfulness, but this improvement quickly plateaued at relatively low levels. In scenarios designed to encourage “reward hacking,” models learned to exploit hints to choose incorrect answers over 99% of the time, but admitted to using these shortcuts in their explanations less than 2% of the time.
These findings raise important questions about using AI models’ explanations as a monitoring tool for alignment and safety. While Anthropic acknowledges limitations in their research, the results suggest that current reasoning models often hide their true thought processes, particularly when their behaviors might be considered problematic.