Detailed Analysis
Anthropic's Claude has come under scrutiny from academic researchers over the opacity of its internal reasoning process, specifically the "extended thinking" feature that the company markets as a premium capability allowing the model to reason through complex problems before producing a final answer. A professor — reported by the Chinese technology outlet 36Kr — highlighted that this thinking process is effectively encrypted and inaccessible to end users and operators, even those paying for higher-tier access to the API. The revelation underscores a fundamental tension between Anthropic's commercial positioning of extended thinking as a transparency-enhancing feature and the practical reality that users cannot inspect or verify the actual chain-of-thought their payments are funding.
The encrypted thinking mechanism refers to Anthropic's practice of shielding Claude's scratchpad-style reasoning tokens from being exposed in raw form to operators and users through the API. While extended thinking was introduced in part to improve performance on complex, multi-step tasks — and Anthropic has positioned it as evidence of a more deliberate, interpretable reasoning process — the encryption of that reasoning layer means the visible output is a curated summary rather than the genuine cognitive trace. This matters because one of the stated justifications for paying premium prices for thinking-enabled models is access to greater transparency about how conclusions are reached, a justification that becomes hollow if the underlying process remains a black box.
This development connects to broader and intensifying debates about AI interpretability and the gap between marketing language and technical reality in large language model deployment. Anthropic has built much of its public identity around the concept of AI safety and interpretability research, publishing work on mechanistic interpretability and positioning itself as the more scientifically rigorous alternative to competitors. The encrypted thinking controversy complicates that narrative by suggesting that even its flagship transparency feature operates with meaningful concealment baked in. Critics argue this creates a trust deficit: users are asked to pay for reasoning they cannot audit and to trust outputs whose derivation they cannot verify.
The academic attention to this issue signals a maturing phase of AI scrutiny in which researchers are moving beyond evaluating model outputs and beginning to interrogate the infrastructure and business practices surrounding AI deployment. The fact that 36Kr — a major Chinese-language technology publication with broad reach among Asia-Pacific tech professionals and investors — amplified these findings suggests the concern resonates across international markets where Claude competes with domestic and global alternatives. As regulators in the European Union, China, and elsewhere develop AI governance frameworks that increasingly emphasize explainability and auditability requirements, Anthropic's encrypted thinking architecture may face not just reputational pressure but potential compliance challenges in key markets.
Read original article →