# OCR Without Sending Documents to OpenAI: How to Check

- Author: Schema Law (https://www.schemalaw.com/)
- Published: 2026-09-21T15:30:02Z
- Updated: 2026-10-11T19:22:20Z
- Category: Confidentiality
- Canonical: https://www.schemalaw.com/blog/ocr-without-sending-documents-to-openai

> Want OCR without sending documents to OpenAI or another model provider? Where to look, how to read a subprocessor list, and what to do if none is acceptable.

To get OCR without sending documents to OpenAI or any other outside model provider, first find out who actually reads the pages. Check the tool's subprocessor list, privacy policy and data processing terms for named AI companies, then ask the vendor in writing. If no third party is acceptable, use recognition that runs on your own computer.

OpenAI stands in here for a whole class of company. The same check applies to any business that sells document-reading models to other software makers. This post explains how to trace where a page goes. It does not say that using such a provider is wrong, and it is not legal advice. Whether a given chain of vendors suits a given client's documents is a judgment for the responsible lawyer under the rules of your jurisdiction.

## Why would an OCR tool send documents to OpenAI at all?

An OCR tool sends documents to a company like OpenAI because building a recognition model is expensive and renting one is quick. The tool's maker writes the upload screen and the export formats, then passes each page to a model provider's programming interface and receives the text back.

Older OCR engines matched character shapes. Some newer tools use general-purpose models that take a picture of a page and describe what is on it, tables and all. The providers of those models sell access by the request. OpenAI's own developer documentation on [file inputs for PDF documents](https://developers.openai.com/api/docs/guides/pdf-files) describes the mechanism: "On models with vision capabilities, such as `gpt-4o` and later models, the API extracts both text and page images and sends both to the model." A converter built on such a service is, in part, a front end for someone else's model.

Three consequences follow for a law office.

- **The brand on the website may not be the company that reads the page.** A small company can offer recognition it did not build, because it rents it.
- **Your document then exists in at least two places.** The converter holds a copy, and the model provider receives one, each under its own terms.
- **The converter's promises bind the converter.** What the provider does is governed by the contract between those two companies, which you have not seen.

None of this makes a tool unsafe. It makes the chain longer, and your diligence has to follow the whole chain.

Map that chain before personnel files or deal drafts leave the firm. Third-party risk includes the model company and every host behind it. Ask for the data processing agreement that covers each hop, and file the company names with the matter. Client confidentiality is simpler to describe when you can say who received the page image.

## How do you find out which model provider an OCR tool uses?

You find the model provider behind an OCR tool by reading four documents and then asking. The subprocessor list, the privacy policy, the data processing agreement and the technical documentation each tend to name outside companies. A written question to the vendor settles whatever those leave open.

| Where to look | What it may tell you | Usual location |
| --- | --- | --- |
| Subprocessor list | Each outside company that handles customer content, with a purpose | Footer links titled Legal, Trust, Privacy or Security |
| Privacy policy | Categories of recipients, such as "AI service providers" | Footer |
| Data processing agreement | The rules for adding and changing subprocessors | Legal page, or on request |
| Product or developer documentation | Model names, or phrases such as "powered by" | Help center, changelog, release notes |
| Security questionnaire answers | Direct replies about third parties | From the sales or security contact |

### Search the documents for the right words

Open each page and search within it for: *subprocessor*, *sub-processor*, *third party*, *service provider*, *AI*, *model*, *LLM*, and the names of the large providers. Vague phrases deserve a follow-up. "We work with trusted partners to deliver our services" tells you there is a chain and nothing about who is in it.

### The question to send

If the documents do not name anyone, send this, and ask for the reply by email so you can file it:

> When we upload a document, which companies other than yours receive the page images or the extracted text? For each one, please give the company name, what it does with the content, how long it keeps it, and whether it may use it to train or improve any model.

A refusal to name any outside company is an answer too, and the firm can weigh it.

### What silence means

A missing subprocessor list does not prove that pages stay in-house. A vendor may simply not have published one. Equally, a list that names no AI company does not prove the absence of one if the list is out of date. Look for a "last updated" line and compare it with the product's recent feature announcements.

## Reading a subprocessor list for a document converter

Read a subprocessor list by sorting its entries into those that can see document content and those that cannot. For each entry that can, note what it does, where it operates and which product features it applies to. The model provider, if there is one, will be in the first group.

A subprocessor is a company your vendor engages to process customer data on the vendor's behalf. Lists follow a common layout. OpenAI publishes one for its own business services, and it shows the form well. The [OpenAI sub-processor list](https://openai.com/policies/sub-processor-list/), as archived in July 2026 and marked "Last updated: June 2, 2026", opens with the sentence "OpenAI engages the following entities to provide processing activities for Customer Data" and sets out four columns: entity name, OpenAI product or service, location of processing, and purpose of processing. The purposes shown include cloud infrastructure, customer support and content moderation. That page is a reminder that the chain does not stop at the model provider. The provider has subprocessors of its own.

### A method for any converter's list

1. **Mark the content handlers.** Hosting, storage, recognition or "AI processing", and customer support tools that receive attachments can all hold your pages. Billing, analytics and email delivery usually cannot.
2. **Find the recognition step.** Look for purposes worded as OCR, document parsing, machine learning, inference or AI features.
3. **Note the location column.** If a client's guidelines restrict where data may go, this is where you find out.
4. **Check the product column.** An entry may apply only to an optional feature, such as a summary button, that you can leave switched off.
5. **Record the date.** Save a copy of the list as it stood when you approved the tool.
6. **Look for a way to hear about changes.** Some vendors offer an email subscription for list updates.

### A worked example

Mara Ellison, a paralegal at a fictional employment firm, is asked to clear a converter called PageLift for a set of personnel files. Its list has six entries: a cloud host, a payment processor, an email delivery service, a help desk tool, an error-tracking service and "an AI model provider, for document text extraction, United States". She marks the host, the help desk and the model provider as content handlers. She notes that the error tracker might capture snippets and adds a question about that. She then opens the model provider's own published terms for business customers, saves all three documents to the vendor file, and sends the supervising lawyer a half-page summary. The lawyer decides. Mara's job was to make sure the decision was about the real chain.

What the contract should say about adding new subprocessors after you sign is a separate subject, covered in the post on [clauses to read in an OCR vendor data processing agreement](https://www.schemalaw.com/blog/ocr-vendor-data-processing-agreement).

## API terms versus chatbot terms at a model provider

A model provider's terms for its programming interface can differ from its terms for its consumer chatbot. A converter that sends pages to a provider does so through the interface, so the chatbot's privacy settings and headlines about them do not describe what happens to your document.

OpenAI's published material shows the split clearly. Its help article on [how your data is used to improve model performance](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance), as archived in October 2026, says of services for individuals: "we may use your content to train our models", with a setting to opt out. Of business services, the article says: "By default, we don't use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API to improve our models."

Its developer guide to [data controls in the OpenAI platform](https://developers.openai.com/api/docs/guides/your-data) adds detail on retention. It says that by default "abuse monitoring logs are generated for all API feature usage and retained for up to 30 days", with stated exceptions that include longer retention required by law, and it describes a Zero Data Retention arrangement that excludes customer content from those logs. It also warns that some endpoints "may retain application state when used, even if you have Zero Data Retention enabled."

These pages change, so read the current versions. The points to carry away are general.

| Question | Consumer chatbot | Programming interface used by a converter |
| --- | --- | --- |
| Who is the provider's customer? | You | The converter's maker |
| Who chooses the settings? | You, in your account | The converter's maker |
| Is training on content the default? | Depends on the provider and your settings | Depends on the provider's business terms and any opt-in by the converter's maker |
| Is content retained after the request? | Per the consumer policy | Per the business terms, and per the features the maker uses |

The right-hand column is the one that matters, and you control none of it. The useful questions for your vendor are therefore: which terms govern your account with the provider, have you opted in to any data sharing, and do you have a reduced-retention or zero-retention arrangement? For how to test the answers on training, see the section on [how to verify a no-training claim](https://www.schemalaw.com/blog/ocr-that-does-not-train-on-your-data#how-can-you-verify-a-no-training-claim). For the risks of pasting a document into the chatbot itself, see the post on [uploading legal documents to ChatGPT](https://www.schemalaw.com/blog/is-it-safe-to-upload-legal-documents-to-chatgpt).

## Does a PDF converter that runs in the browser still upload the file?

Usually, yes. A converter you open in a browser is a web page that sends your file to the company's servers, and from there possibly onward. "Runs in the browser" describes where you click. Only a tool that does all its processing on your device, inside the page, avoids an upload.

If you are looking for a PDF converter that doesn't upload to third parties, there are two separate uploads to rule out: the one from your computer to the converter, and the one from the converter to anyone else. The second is invisible from your desk. Only the documents and the vendor's answers reveal it.

The first you can observe.

### Three ways to see whether the file leaves your machine

1. **The offline test.** Load the converter's page, then disconnect from the network. Convert a blank test page. If it works, processing is local. If it fails or waits, the page needs a server.
2. **The network panel.** Open your browser's developer tools, choose the Network tab, and convert a test file. A request whose size is close to the file's size, sent when you click convert, is the upload.
3. **The site's own words.** Look for a plain statement such as "files never leave your device". Phrases like "secure upload" or "files deleted after one hour" confirm that an upload happens.

Use a blank or public page for all three. A test is not a reason to send a client's file anywhere.

### Limits of the tests

A page that passes the offline test today can be changed tomorrow, because a website is delivered fresh on every visit. A tool may also work locally for simple conversions while sending harder pages to a server. Read what the site says about each feature.

Whether a browser-based service or installed software suits your office overall is a broader comparison, set out in the section on [whether you need to install OCR software at all](https://www.schemalaw.com/blog/online-ocr-vs-desktop-ocr-for-law-firms#do-you-need-to-install-ocr-software-at-all).

## Options when no third-party model provider is acceptable

When no third-party model provider is acceptable, you have four routes: recognition that runs on your own machines, a vendor that operates its own recognition without outside processors, a vendor whose chain the client has approved in writing, or leaving the document as a scan. Which fits depends on why the restriction exists.

Start by pinning down the restriction. "No OpenAI" in a client's guidelines may mean no consumer chatbots, no named company, no generative models at all, or no processing outside a country. Ask the client which, because the options differ.

| Route | What it involves | Trade-off |
| --- | --- | --- |
| Local OCR | Software installed on a firm computer does the recognition | You maintain it, and results on difficult scans vary by engine |
| Vendor with no outside processors | A hosted tool whose maker states that it runs recognition itself | You still rely on one vendor, and must verify the statement |
| Approved chain | A hosted tool with named subprocessors the client has accepted | Needs the client's written approval and a watch on changes |
| No conversion | Work from the page images | No search, no copy and paste |

### Local OCR

Installed software keeps the page on your computer. One freely available engine is Tesseract, which its own [Tesseract user documentation](https://tesseract-ocr.github.io/tessdoc/) describes as "an open source text recognition (OCR) Engine, available under the Apache 2.0 license" that "can be used directly via command line". Before you rely on any installed product, check its settings and documentation for any feature that sends pages to an online service. Test it on your own difficult pages and proofread the output.

### A vendor that makes the statement in writing

If a hosted tool says no outside company receives your pages, get that sentence in the contract, with a duty to tell you before it changes. A statement on a web page can be edited without notice.

### Where Schema Law fits

Schema Law is a hosted service: you upload in the browser and export the result. On use of documents, its homepage says: "What you upload is never used to train or improve an AI model. Not ours, and not anyone else's." The commitments listed for every plan on the [pricing page](https://www.schemalaw.com/pricing) add "Documents stay inside your own account". Those sentences, with the homepage's [security commitments](https://www.schemalaw.com/#security), are the published position. The site does not name the companies or models involved in reading a page. They concern what may be done with a document. If a client's restriction turns on which companies are in the chain, put the question from earlier in this post to Schema Law in writing, as you would to any vendor. The Enterprise plan lists "Security questionnaire and contract review".

### Write the decision down

Whatever route you choose, note in the matter file what the restriction was, what you checked, on what date, and who approved the tool. If the chain changes later, that note tells you which matters to revisit.

## Try it on a real document

Upload a scanned PDF and export it as Word, a searchable PDF or plain text on the free plan. Your documents are never used to train an AI model. [Start free](https://www.schemalaw.com/api/auth/google)