OCR That Does Not Train on Your Data: How to Verify

To find OCR that does not train on your data, look for a written promise in the vendor's terms that uploads are never used to train or improve any model. Then test the promise: check whether it covers your plan, every company in the processing chain, and feedback features. You cannot observe training from outside, so the wording matters.

A lawyer can do this in an afternoon. This post defines the term, explains why it matters for client files, and gives a verification method with the usual exceptions marked. It describes how to handle documents and is not legal advice. Your jurisdiction's professional rules decide what you owe a client.

What does training on your data mean?

Training on your data means a provider uses what you upload, or the text it extracts, as material to build or adjust a machine learning model. Your documents stop being only the input to your own job and become part of how the system behaves for other people later.

Many OCR products now use AI models to read pages. That fact alone does not mean they train on uploads. Reading a page with a model and teaching a model with that page are two separate acts, and the post on how AI OCR differs from classic OCR explains the reading side.

Vendors use several words for the teaching side. Know them all, because a policy can avoid the word "train" and still describe it.

Processing is none of these. A vendor has to handle your file to convert it. That may mean holding it briefly, running it through a model and storing the result for you to download. A no-training promise does not say the vendor never touches the document. It says the document is not kept as teaching material.

Two neighboring questions get mixed up with training. Retention is how long the file exists on the vendor's systems, and the post on whether OCR software stores your documents covers that. Access is which people can open it. A vendor can score well on one and badly on another, so ask about each.

Why does model training matter for client files?

Model training matters because it sends a client's information somewhere you cannot follow or recall. Bar guidance treats tools that learn from inputs as a confidentiality risk in themselves, and a federal court has pointed to a provider's training terms when deciding that a user had no reasonable expectation of confidentiality.

ABA Formal Opinion 512 describes the mechanism. With self-learning generative AI tools, information from one client's matter can be "later revealed in response to prompts by lawyers working on other matters", or disclosed "to persons outside the firm who are using the same GAI tool." The opinion concludes that a client's informed consent is required before a lawyer inputs information relating to the representation into such a tool. It notes that this view reflects the technology as of July 2024.

The State Bar of California's practical guidance on generative AI names the same risk: that "information provided as an input may be used for learning purposes by the generative AI tool and, as a result, become part of, or be used to generate, a response to another user."

Three features make training worse than ordinary storage.

• You cannot take it back. A stored file can be deleted. Nobody can point to where a document sits inside a trained model and remove it.
• You cannot see it. No log shows you that page 14 of an exhibit was used.
• It widens the audience. Training pipelines can involve people who label or review samples, and systems outside the product you signed up for.

There is also a privilege angle. In United States v. Heppner, a 2026 memorandum from the Southern District of New York, the court relied in part on a consumer AI platform's privacy policy, which it read as saying the provider used inputs and outputs for training and could disclose data to third parties, in finding that a user's exchanges with the platform were not confidential. The post on attorney-client privilege and AI tools covers that ruling and its limits. Bar opinions are advisory and vary by state, and one district court does not bind another.

When you write up that reading, separate client confidentiality from whatever a court may later say about privilege. The duty of competence is met by filing the terms of service you actually opened, with the date. Watch for a deidentification clause that lets the vendor keep stripped text. Ask whether document content can still sit inside that exception.

How can you verify a no-training claim?

You verify a no-training claim by reading the binding terms, asking the vendor direct questions in writing, and checking that the answer covers your plan and every company involved. You cannot verify it by testing the product. What you end up with is a documented, enforceable promise. It is not technical proof.

Start by being honest about the limits. No experiment you can run from a browser shows what a vendor does with a file after conversion. Short of an audit you are unlikely to get, everything rests on what the vendor has committed to and whether that commitment can be enforced. California's guidance says reasonable efforts "require more than reliance on generalized marketing assurances." So the goal is to move the claim from the marketing page into something that binds.

1. Find the sentence. Open the terms of service, privacy policy and any data processing agreement. Search for "train", "improve", "develop", "machine learning" and "research". Copy the exact sentence that makes the promise and the name of the document it sits in.
2. Check whose models. "We do not train our models" leaves open training by another company whose model does the reading. Look for wording that covers third parties, or ask.
3. Check which plan. Promises are often tied to paid or business tiers. Confirm the sentence applies to the plan you are on, including a free trial.
4. Check the default. "By default" means a setting exists. Find it, see who can change it, and record how it is set on your account.
5. Ask in writing. Send the vendor the questions below and keep the reply.
6. Save the evidence. Keep a dated PDF of each document and a screenshot of the relevant setting.
7. Set a review date. Terms change. Look again in six or twelve months, and whenever the vendor emails about updated terms.

Questions to send:

• Are my uploaded documents, the extracted text, or my exports ever used to train, fine-tune, evaluate or otherwise improve any model?
• Does that answer cover every company that processes my documents for you?
• Does it apply to all plans, including free and trial accounts?
• Do any staff or contractors review customer documents, and for what purpose?
• Which document makes this commitment binding, and how will I be told if it changes?

Not all evidence is equal. Rank what you collect.

The same ranking applies to every vendor, including this one. Schema Law's own commitment is taken up in the last section of this post.

Where no-training promises usually hide exceptions

Exceptions to a no-training promise usually hide in seven places: a broad "improve our services" clause, carve-outs for de-identified or aggregated data, feedback buttons, differences between plans, other companies in the chain, human review, and the vendor's right to change its terms. Read for each before you rely on the headline.

The services clause. A policy can promise no training in one paragraph and reserve the right to use content "to improve our services" in another. Ask the vendor which one governs.

De-identified and aggregated data. Some terms exclude data once it has been stripped of identifiers. A contract with the names removed can still identify a deal to anyone who knows the industry. Ask what de-identification means in practice and whether document content is ever included.

Feedback. This one is easy to trigger by accident. OpenAI's help page on how your data is used to improve model performance said, as archived in October 2026, that if you submit feedback such as a thumbs up or thumbs down, "the entire conversation associated with that feedback may be used to improve our models, even if you've opted out." Any product with a rating button, a "report a problem" form or a "share this sample with support" option deserves the same question. Policies change, so read the current page.

Plan differences. The promise may attach to business tiers only. Ask whether the free tier and the trial run under the same terms as the paid plan. The post on whether free online OCR is safe for legal documents looks at how free services pay for themselves.

Other companies in the chain. A vendor that sends pages to another provider for recognition can only promise what its own contract with that provider secures. Ask for the list and for the terms that cover it.

Human review. A policy may allow staff to read samples "for quality" or "for safety". That is access. It is a separate thing from training, but it matters for the same reason.

Future terms. "We may update this policy at any time" means the promise you verified may not be the one in force next year. This is why step 7 above sets a review date.

A common mistake is to stop at the first reassuring sentence. Search the whole document for every word in the phrase table near the top of this post, and read each hit.

Opt-out versus never-trained

Opt-out means the vendor uses your content for training unless you find a setting and switch it off. Never-trained means no such use happens on your plan and you do not have to act. With opt-out, the burden is on you, on every account, for good.

OpenAI's published pages show both models side by side, which makes them a useful illustration. As archived in October 2026:

• For individuals, the model performance page linked above says, "we may use your content to train our models", and "To opt out, turn off Improve the model for everyone under Settings > Data controls in ChatGPT."
• For business offerings, the same page says, "By default, we don't use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API to improve our models."

The opt-out has limits written into it. The page says, "After you opt out, we won't use your new conversations to improve our models." That sentence is about new conversations. It makes no promise about anything sent before you changed the setting. OpenAI's data controls page adds that for signed-out use, the choice "is saved in that browser". Check the current wording before relying on any of this, since these pages are revised often. The post on uploading legal documents to ChatGPT sets out the rest of what those pages say.

For a firm, opt-out fails in predictable ways:

• A new hire's account starts with the default.
• Someone signs in on a second device or a different browser.
• A staff member uses a personal account for one urgent job.
• Nobody recorded when the setting was changed, so earlier uploads are in doubt.

If an opt-out tool is all you have, treat the setting as a control that needs an owner. Put it on the onboarding checklist, check it on a schedule, and keep a dated screenshot per account. If you are choosing a tool, prefer one where the answer does not depend on a toggle. The post on what a closed AI system is explains a related claim vendors make.

The no-training commitment Schema Law makes

Schema Law's site states, as a commitment on every plan, "Your documents are never used to train an AI model". The homepage adds that this covers other companies' models as well as its own. Those are the company's words. Check them the way this post tells you to check anyone's.

The fuller statements, quoted from the security section of the homepage and the questions below it:

• "What you upload is never used to train or improve an AI model. Not ours, and not anyone else's."
• "Documents you upload are used only to produce your export. They are never used to train or improve any model."

Measured against the fine print list above, the wording does three things. It says "never", with no default to switch. It says "train or improve", which is wider than "train". It reaches other companies' models: "Not ours, and not anyone else's." The commitment appears in the list that applies to every plan on the pricing page, the Free plan included.

Now the limits. A website statement sits low in the evidence table earlier in this post. Nothing here independently verifies it, and this post makes no claim about how the service is built or operated. Send Schema Law the five questions from the verification section, ask which document makes the commitment binding for your plan, and keep the answer. The Enterprise plan lists "Security questionnaire and contract review" for firms that run a formal review.

If the reply you get from any vendor is slower or vaguer than its homepage, take that as information about the vendor and keep looking.