Several interlocking reasons, none of them accidental: **Technical architecture was designed before consent was a concern** The dominant data pipeline pattern -- crawl the web, deduplicate, filter for quality, train -- was established when the goal was simply "more data." Consent metadata was never part of the schema. Adding it retroactively would require re-crawling sources with consent signals attached, which is a different problem than training itself. **No standard exists for expressing consent** `robots.txt` handles crawling, not training. There is no widely adopted equivalent of a `training-consent.txt` or structured field in HTML metadata that says "this content may/may not be used to train models." The `ai.txt` and `TDMRep` proposals exist but have minimal adoption and no enforcement mechanism. **Consent is legally ambiguous in most jurisdictions** If the legal question of whether consent is even required is unsettled (fair use arguments in the US, ongoing litigation, etc.), there is little commercial incentive to build infrastructure that could be used as evidence of what you did *not* obtain consent for. **Data passes through too many hands** Common Crawl is collected by one party, filtered by another, licensed or released by a third, and trained on by a fourth. Even if the original author expressed a preference, that signal degrades or disappears at each handoff because no format preserves it. **Consent records would create legal exposure** A machine-readable log of "we had no consent for this URL" is a liability document. The absence of such a record is not negligence -- it is the default state, and legally safer. The result is that consent, even when arguably given (e.g. a public Creative Commons license), is not recorded *with the data* in a way a model or auditor can later query. It exists only in external legal instruments that are not linked to the training corpus.