Claude Code's "extended thinking" is a summary- not authentic thinking (patrickmccanna.net)

Claude Code's recent announcement regarding its "extended thinking" feature has raised concerns in the AI/ML community. Although the model records its reasoning in "thinking blocks," users only receive a summary of this reasoning instead of the complete output. This means that while the logs are saved, they are encrypted, and only the company, Anthropic, holds the decryption key. Accessing the full reasoning is restricted to users with an enterprise agreement, which poses challenges for accountability and transparency, particularly for individuals seeking clear audit trails of AI decision-making. The implications of this limitation are significant. Users may be misled into thinking they have full access to their agent's reasoning when in fact, they only have a summarized version, much like editing a converted file format that loses original data. This lack of access can hinder effective oversight and understanding of the model's operations, particularly when users depend on accurate reasoning logs for analysis or debugging. The initial excitement over "extended thinking" might need to be tempered with caution as the actual utility of these features becomes clearer.