What to Check Before You Upload Court Papers to an AI Chatbot

It's late. You've read a letter from the court three times, and one paragraph still makes no sense.

You open a chatbot. You could ask what a legal term means. You could upload the letter. You could ask it to write your response.

Those are different requests. Each gives the tool a different amount of information and a different role in your case.

Before you press send, there are three questions worth separating. Who receives your information? Is the answer accurate? Does it actually fit your situation?

I'm attorney David Martino, and welcome to The Steadycase Podcast. This show offers general legal information for everyone listening and practical ways to keep your records organized. An attorney who represents you can advise you about your own case.

I use AI in producing this podcast. I understand why someone would turn to a tool that offers an explanation immediately, especially when the alternative is another evening of confusion.

The useful question is how much confidence a particular answer deserves. Producing a clear sentence and establishing that the sentence is correct are separate steps.

In the previous episode, we talked about getting text messages off the phone and into a dated file. Today, we'll follow what happens when a copy goes into an online AI service, and what deserves attention when an answer comes back.

There is research on this. Clio's 2025 survey of a thousand U.S. adults found that fourteen percent had used AI to answer legal questions. That figure covers legal questions generally. It doesn't tell us how many parents are using chatbots during custody cases.

A newer UK study gives us a closer look at the conversations. Among participants who reported a recent legal dispute, about one in six had used a chatbot for help. Some of the questions concerned family law. Some people already had lawyers and asked a chatbot to explain or check what their lawyer had said.

The two studies cover different populations. Neither gives us a national count of American custody litigants using AI. They do document something recognizable: people are seeking explanations, drafting help, and reassurance during legal problems.

Start with the information you provide.

When you send a file to an online service, treat it as a copy handed to another organization. Its handling depends on the service, the feature, and the kind of account.

For example, one large AI company's consumer privacy page says human reviewers examine some chats. The company warns against entering confidential information that someone wouldn't want a reviewer to see. Reviewed chats can be retained for up to three years, even after the user deletes their activity.

Other services and account types have different arrangements. One provider says it doesn't use business-service or API inputs and outputs for model improvement by default. Its consumer accounts have different settings.

That distinction is important when AI is built into another app. The consumer chatbot's policy may not describe the arrangement that app has with the provider.

Connecting a chatbot to an outside service can also mean sharing information with another organization. Check both the chatbot's privacy policy and the connected service's policy, including what each receives and retains.

These are examples from policies checked in October 2026. Check the current terms for the particular service and feature.

Training is only one question. There are also questions about storage, access, and deletion. A promise that your information won't train a model doesn't tell you whether the service keeps a copy. It doesn't tell you whether the answer is correct, either.

The American Bar Association asks lawyers using these tools to examine their terms and confidentiality risks. That guidance is written for lawyers. The underlying questions are useful for anyone considering an upload. Who receives the information? Who can access it? What happens to stored copies after deletion or account closure?

A court order can impose restrictions beyond a company's privacy policy. A document's availability in your files doesn't settle whether you may send it to an outside service. Court rules, orders, and any restrictions on the records need separate attention. A self-help center can help locate court resources. A lawyer can explain how restrictions apply to particular documents.

There is a technical problem, too. Covering words on a screen may leave the underlying text in the file.

Two federal courts warn that drawing a box over text, changing its color, or using certain PDF commenting tools can leave information recoverable. Proper redaction removes the information. The appearance of a black rectangle alone doesn't establish that removal.

Keep the complete source separate from the copy prepared for sharing. A redacted document should have a distinct file name. The document that looks clean on your screen still needs a check for information remaining in the file.

Here's a hypothetical parent deciding what to do with that court letter.

The parent saves a complete scan in the case archive. The question concerns one unfamiliar term, so the parent first considers whether a general explanation requires any upload at all.

That decision comes before preparing a copy. Removing a name doesn't automatically make an entire case document appropriate for an outside service.

If the parent later shares a permitted copy, the archive should identify that exact version. A log entry records the date, the service, and the account type. It points to the saved copy that was actually sent.

Suppose the upload contained only one page, with some information removed. A log pointing only to the complete five-page source would leave the parent guessing about the upload.

Keeping both versions answers two different questions. What did the parent receive? What did the parent share?

This is an organization habit. A detailed log doesn't provide permission to disclose a restricted document.

Now suppose the service has strong privacy protections. We still have to examine its answer.

An AI summary can omit information or produce an inaccurate statement. A generated citation can look convincing without identifying a real case. Those risks exist separately from the provider's training policy.

Consider another hypothetical example. A text message says, “I'll be there at six thirty.” The generated summary says, “The other parent refused to return the child.”

The second sentence adds a conclusion. The quoted message alone doesn't establish a refusal. Other records might change our understanding, but those records would need their own review.

That's why a summary needs comparison with its source. Did it preserve the date? Did it identify the right speaker? Did it turn an allegation into an established fact?

This example describes a possible error. It isn't a finding about any particular product.

There is another step beyond checking whether a sentence is true.

A real court rule may govern a different state. A real form may address a different kind of proceeding. A quotation may be accurate while the conclusion drawn from it goes beyond the decision.

A tool that puts documents in date order performs a different task from a tool that writes a declaration or selects grounds for a motion. The latter may introduce factual assertions and legal judgments that require much closer review.

A completed form can look finished while leaving those questions unanswered.

The distinction is useful even when someone has an attorney. A generated draft can help identify questions for a conversation. It doesn't establish that the proposed filing fits the case.

The consequences are documented.

In May 2026, a New York appeals court considered a custody appeal by a self-represented father. His brief relied on an AI-generated citation to a nonexistent decision. The court imposed a two-hundred-fifty-dollar sanction for his failure to verify the authority.

There is an important qualification. When deciding the custody issues, the court expressly said it drew no unfavorable inference from his use of AI. The sanction and the custody decision were separate parts of its analysis.

Self-represented people appear frequently in the reported decisions. Researcher Damien Charlotin maintains an international database of decisions involving hallucinated material. In the October snapshot we checked, it recorded more decisions involving self-represented litigants than decisions involving lawyers.

That's a comparison within a collection of reported decisions. It doesn't establish which group is more likely to make an error or get caught. The database doesn't count every AI-assisted filing, and the entries don't all involve a monetary sanction.

In Florida, the state's Supreme Court amended its rules in 2026 so that people signing court filings represent that the cited legal authorities exist and are accurately cited. The requirement applies to attorneys and unrepresented parties.

That is an accuracy requirement. It isn't a ban on AI. Other courts have their own rules, so a national podcast cannot supply one filing checklist for everyone.

First: If the service promises not to train on my records, can I trust its draft?

That promise addresses one use of your information. It doesn't establish that the draft preserves your facts, selects the appropriate form, or draws a sound legal conclusion. Privacy protection and output review are separate checks.

Picture a draft that quotes a text message correctly, then treats that message as proof of something it doesn't say. The service could keep its privacy promise and still give you a sentence that needs correction. Put the draft beside the records it describes. Check the dates, the quotations, and which statements are facts versus conclusions. Then ask whether the form and rule fit the court and the kind of proceeding.

Second: If I delete the chat, have I deleted the uploaded file?

It depends on how the service stores files. One provider says that deleting a chat doesn't delete a file saved separately in its library. Another retains some human-reviewed chats after activity deletion. Check the service's current file and chat policies separately.

Think of the conversation and a file saved in a library as two things you may need to manage separately. If a connected service also received information, its copy may have its own rules. Look for the controls and retention terms that apply to the particular account, file, and connection. Your own record of the exact copy shared helps you identify what you're trying to find or delete.

Third: Can I ask the chatbot to check its own answer?

You can ask it for a source or a different explanation, but a second answer is still generated output. If it repeats a citation, repetition doesn't verify that the case exists or says what the chatbot claims.

For a summary, set it beside the underlying record. Imagine a court letter says one party requested a change, while the summary says the court ordered it. Those are different statements. For a legal citation, open the actual decision or rule and check the passage. Whether that authority applies to a particular case is another question; a licensed attorney can help with that assessment.

Fourth: Can I use ChatGPT, Claude, or Gemini to help with my custody case?

These services can explain a general term or suggest questions for a lawyer. A person can ask what a term means without uploading the court document. When someone does send a prompt or file to an online service, the provider receives it for processing.

Here is a snapshot of the companies' published consumer policies as of October 2026. The dated links are in the show notes.

On a personal ChatGPT plan, model improvement is on by default, but the user can disable that setting. Regular chats stay in the account until the user deletes them. OpenAI generally schedules deleted chats for removal within thirty days, with exceptions.

For Claude Free, Pro, and Max, Anthropic says using chats to improve Claude generally depends on the user's choice, with stated safety and feedback exceptions. A deleted chat is generally removed from its back-end systems within thirty days. If the user allows model improvement, de-identified data may remain in its training pipeline for up to five years.

For Gemini Apps, Google's Keep Activity setting saves chats and allows activity to improve its services. The default auto-delete period is eighteen months. Chats reviewed by humans can remain for up to three years even after activity deletion. Turning Keep Activity off changes the handling of future chats, but Google says it retains them for seventy-two hours to operate and protect the service.

Business accounts and connected services can have different terms. Check the current terms for the account and feature. A provider's privacy settings do not establish that a generated answer is correct.

Fifth: What are the risks of putting personal information in a chatbot?

Imagine a parent pasting a court letter that names a child, gives a home address, and describes a medical appointment. The parent might need help with one sentence, but the service receives the entire letter. The document may also contain information about people who never chose to use the chatbot.

One risk is retention. A conversation or uploaded file may stay in an account after the person closes the app. Deleting a chat may leave a separately saved file, a connected service's copy, or material retained for review under the provider's policy.

Another risk is use and access. Could the information help train a model? Depending on the service and settings, yes. Authorized people or service providers may also review some conversations for safety, support, or quality. Both OpenAI and Google caution users about sensitive information in consumer chats.

Using content to improve a model is different from posting the original letter on a public website. A provider that uses content to improve a model does so under its own policies. Turning model improvement off does not automatically delete a saved chat or govern a connected service's retention. Anthropic also says changing that setting cannot remove data from a training run already in progress.

A useful first step is to ask the general question without case facts or the file. If the exact wording matters, consider a short excerpt without including personal details. Keep a copy of the exact material sent to the service. Drawing a black rectangle over text in a PDF editor is different from using its Redact feature. The rectangle may hide words on the screen while leaving them in the file. A proper Redact feature removes the selected content when you apply and save it. The final file may still contain other sensitive details or hidden information, so check the copy before sharing it.

AI can make unfamiliar language easier to approach and help someone prepare a clearer question. The research shows why people seek that assistance.

Useful assistance still needs careful review. Before sharing a record, understand the service's handling of the information and any restrictions on disclosure. After receiving an answer, check both its accuracy and its relevance to the situation.

Keep the complete source in your archive. Keep a separate record of the exact copy you shared. Treat a generated summary or draft as another document requiring review.

The free Document Everything Method is at getsteadycase.com/document-everything. The research and court decisions discussed today are linked in the show notes.

Have a topic you'd like us to cover? Email [email protected]. Please leave out the details of your own case.

Thanks for listening.