Corrections
What we got wrong, and what we changed
Each Terms Watch edition lists corrections to the one before it, apart from vendor changes and never counted as one. They are collected here. If you find a cited fact that does not reproduce, email us; we will check it, list it here and refund your purchase (see refunds).
12 Oct 2026 edition, rebuilt on 15 Oct 2026
Our own review of the first print of this edition found errors. One change was false: the Cohere API’s IP indemnity showed as moving from Yes to Not stated, because a long contract was dropped from the search. Four matrix changes had been accepted on 2 of 3 model votes, when the method requires 3. Three changes found in newly tracked documents read as that week’s news, though they were months old. We fixed the method and rebuilt the edition. The 12 Oct sample on this site is the rebuilt edition.
Listed in the 19 Oct 2026 edition, correcting the 12 Oct 2026 edition
None of the documents behind these values changed, so they are our corrections, not vendor changes.
Anthropic, Claude API, abuse-monitoring retention
Printed Not stated, correct value 30 days. The printed value rested on only 2 of 3 models; at least 3 must agree.
Anthropic, Claude API, human review
Printed In review, correct value Conditional. The models had not agreed on a value.
Cursor, Cursor Teams and Enterprise, trains on your data by default
Printed Not stated, correct value Opt-in. That run did not read the Cursor Terms of Service.
Cursor, Cursor Teams and Enterprise, subprocessor list
Printed Not stated, correct value Yes. That run did not read the Cursor Privacy Policy.
Listed in the 12 Oct 2026 (as corrected) edition, correcting the 5 Oct 2026 edition
Anthropic, Claude API, trains on your data by default
Printed No, correct value Opt-in. The printed value rested on only 2 of 3 models; at least 3 must agree.
Anthropic, Claude API, abuse-monitoring retention
Printed Not stated, correct value 30 days. The printed value rested on only 2 of 3 models; at least 3 must agree.