AI
The training-data argument has moved from principle to syntax
Two disclosures this week show the AI industry negotiating consent by syntax. Google, OpenAI, Microsoft and Anthropic backed one robots.txt policy token so a publisher can refuse training once for every participating provider, though implementation dates run from months away to February 2027. Anthropic separately described self-assessment, in which Claude grades its own conversations. Both are narrow: the token does not stop crawling, and the monitor does not block misuse. They are still the shape of the coming settlement — machine-readable, auditable, and run by the party being audited.