Who trains on your documents in 2026
"Do you train on our documents?" is the question most often asked in a legal software evaluation and least often answered in writing. We read the security pages of the vendors that publish one and quoted what they actually say, in September 2026.
Everything below is the vendor's own statement, not our verification of their practice. Where a page does not address training, we say that rather than inferring an answer from silence.
What each vendor says
| Vendor | States a training position? | What the page says |
|---|---|---|
| Harvey | Yes, explicitly | "We don't use inputs, outputs, or uploaded documents to train underlying models." Adds that it contractually prohibits model providers from training on customer data. |
| GenieAI | Yes, explicitly | "GenieAI does not use customer documents, prompts, or confidential data to train shared or internal AI models." Third-party providers operate under zero-data-retention agreements. |
| LegalOn | Yes, via its provider | States protections "that ensure no customer data is used to train, retrain, or improve their models, and that no customer data is stored for any period of time, even for abuse monitoring and debugging purposes." |
| Spellbook | Not directly | Does not state a training position. Says it has negotiated zero data retention with OpenAI and Anthropic, so data "is not persisted and exists only in memory in order to process a request." |
| Juro | No | The security page does not mention model training. |
| Evisort | Page no longer exists | Its security page now redirects to Workday's contract management product page following the acquisition. |
Zero data retention is not the same as no training
This is the distinction that matters most, and the one the phrasing above hides. Zero data retention is a commitment by the model provider not to keep your data after processing. A no-training commitment is about what the vendor itself does with what it holds.
A vendor can hold your documents indefinitely in its own systems, use them to improve its own retrieval or classification, and still truthfully say it has zero data retention with OpenAI. Both commitments are worth having. They answer different questions, and only one of them is about the vendor.
Harvey, GenieAI and LegalOn each address the vendor-side question directly. Spellbook's page covers the provider side and does not address the vendor side, which may simply mean the page has not been updated to answer a question buyers only recently started asking.
What certifications each page names
Recorded as written on the vendor's own page. We have not inspected any certificate or report, and scope matters more than the name.
| Vendor | Named on their security page |
|---|---|
| Harvey | SOC2 II, ISO 27001, ISO 27701, ISO 42001, GDPR, CCPA |
| LegalOn | SOC 2 Type II, ISO/IEC 27001:2022, ISO/IEC 27017:2015, plus GDPR and CCPA compliance |
| Spellbook | SOC 2 Type II, HIPAA, GDPR, EU AI Act |
| Juro | SOC 2 Type II |
| GenieAI | ISO/IEC 27001:2022, certificate 310012019, valid to 6 February 2029 |
GenieAI is the only one of the five to publish a certificate number and expiry, which is the detail that makes a claim checkable rather than assertable. Harvey names ISO 42001, the AI management standard, which is uncommon and worth asking about.
The acquisition problem, concretely
Evisort is the illustration. Its security page was a standalone disclosure; it now redirects to a Workday product overview. The commitments a buyer relied on before the acquisition are no longer at the address they were read at.
That is not an allegation that anything changed substantively. It is a demonstration that a security position published on a vendor's website is only as durable as the website, which is an argument for getting the commitment into your contract rather than bookmarking a page.
What to ask
- Do you train any model, yours or a third party's, on our documents? Ask about both, because the two are commonly conflated.
- Is that the default, or an opt-out we have to request? Some vendors differ across their own tiers.
- Put the answer in the contract. A statement on a web page can be edited, and after an acquisition it can disappear entirely.
- Which sub-processors see our documents, and what are their commitments?
- What is the retention period, and can we require deletion on demand?
- If you are acquired, what happens to these commitments? Ask before signing; several tools on this site changed owner in the last two years.
Method
Statements were read from each vendor's own security page in September 2026 and are quoted rather than paraphrased. Vendors without a security page at the conventional paths, including Icertis and PandaDoc when we checked, are not listed rather than being marked as failing to disclose. Absence of a page is not evidence of a position.
Corrections are welcome. If a vendor has published a training position we missed, we would rather add it than leave the row as it is.