AI service agreements overlook ownership of intermediate reasoning data

Palantir CEO Alex Karp's concerns about AI data ownership are validated by a gap in service agreements. OpenAI and Anthropic define customer content as only the input and output, leaving intermediate reasoning tokens unaddressed. This means the 'digital scratchpad' generated during reasoning may not be owned by the customer, potentially exposing proprietary information.

The article's central claim is that legal definitions in AI service agreements create an unaddressed category of data. Reasoning models generate intermediate computation steps before producing final answers, and neither OpenAI nor Anthropic explicitly assigns ownership of these intermediate tokens to customers. The piece draws a parallel to a consultant who keeps their working notebook after delivering a final answer.

Both companies' practices suggest the data carries real value. Anthropic charges customers for full internal reasoning even when it is never displayed, and returns raw reasoning encrypted so only Anthropic can decode it. The article also notes that Anthropic has accused competitors of generating over 16 million Claude exchanges for model distillation, which the labs treat as a serious threat to their competitive position.