How we verify
A wrong “does not train” could cost someone their data, so the rule is simple: no quote, no verdict.
- 01
Find the vendor’s documents
From the tool’s homepage and a web search, we keep only pages on the vendor’s own domain (or its parent company’s, when the homepage links there): privacy policy, terms, DPA, subprocessor list, trust or security page, and AI data-usage FAQs. Pages that need JavaScript to show text, or that sit behind a form, may not be readable; such fields stay “Not stated”.
- 02
Read per plan
Claude Haiku 5.5 reads the documents and fills a fixed schema for each plan the vendor distinguishes (Free, Pro, Team/Business, Enterprise, API): trains on customer data, how to opt out, retention, data regions, DPA, subprocessors, certifications and human review. It is told to use only the documents and to copy a supporting sentence for every answer.
- 03
Verify every quote in code
We normalize whitespace, markdown and typographic quotes, then check the quote appears in the fetched page exactly. Quotes shorter than 20 characters, or not found, are rejected and the field becomes “Not stated”. Ellipses are allowed only between pieces that each appear, in order.
- 04
Show confidence and freshness
Each field shows its source link, a confidence level (high when the sentence states it directly, medium when it needs light interpretation) and when we last checked.
- 05
Track changes
Watched tools are re-fetched daily and the rest every two weeks. When a document’s text changes, we read it again and record a field-level difference on the tool page and in the public change log, and alert watchers in Slack and by email.
- 06
Corrections
Anyone can report an error on any field. An editor checks it against the source and can correct the field; corrected fields are marked as reviewed.
Informational, not legal advice. Every verdict is what the vendor’s own documents state, quoted word for word. Read the cited source before you decide, and check your own contract: negotiated terms can differ.