Limits, privacy, and ethics

Why this matters

AI can help you research, but it cannot take over your responsibility. You sign what you file. You also decide what private information goes into a tool. This chapter covers the limits of this method, the rules you may have to follow, and when to bring in a person.

This is general information, not legal advice.

Not legal advice, not a lawyer

This guide teaches a research method. It is not legal advice, and we are not your lawyers. An AI tool is not a lawyer either. It does not know your whole situation, it cannot take responsibility for your case, and it can be wrong with great confidence. Rules and deadlines differ by state and by forum. Use the method to prepare, and use a lawyer for judgment calls about strategy.

You are responsible for what you file

Courts treat the signer as responsible. In the cases described in Chapter 0, filings with fake citations led to sanctions or a disciplinary referral. One involved a self-represented party, and the court said a sanction is warranted "even when the offending party is a pro se litigant." Using AI is not itself the problem. Filing something you did not check is.

AI tools can invent cases, quotes, and holdings that look real. The checks in this guide reduce that risk. They do not remove it, and they do not shift responsibility to the tool or to anyone else.

Before anything goes out, you should be able to say for every case, quote, statute, and fact that you read it and it says what you claim.

Court and arbitrator AI rules

Forums are adopting rules on AI. Do not assume yours has none, and do not assume it has the same rules as another.

New York gives one example. 22 NYCRR Part 161 took effect June 1, 2026. It applies to all courts of the state's Unified Court System, in civil and criminal cases. Its policy is that attorneys and parties should not be barred from using AI tools to prepare papers, and should not be required to disclose AI use when they file, because the duties that apply to every paper already apply. A court may, in its discretion, adopt its own part rule on AI, and if it does, it is encouraged to use the model rule in Appendix A. That model rule says a person using an AI tool must carefully review the paper and independently ensure it contains no fabricated or fictitious cases, statutes, or other material. Signing the paper certifies that the review was done. If it was not, the person may face a sanction or other remedial action. The model rule covers attorneys and parties, so it reaches self-represented people. The definition of "paper" excludes evidence. Because adoption varies, check the assigned judge's part rules. We could not confirm whether some older individual-judge AI rules are still in force, so ask the court.

Federal courts are not covered by Part 161. Federal judges' standing orders vary, and we did not verify a uniform federal rule.

Four steps keep you on the right side of these rules.

1. Read the rules of your forum. For court, that means the court's rules and your judge's standing orders or part rules.
2. For arbitration, read the provider's rules and every order from your arbitrator. Ask the case manager if the orders do not address AI.
3. If a rule requires disclosure or certification, follow it exactly.
4. When in doubt, ask. A short written question to the court or case manager is better than a sanction.

Privacy: what you put into an AI tool

Anything you type or attach may leave your computer. Decide what you are willing to share.

• Consumer data settings. As of October 2026, on Free, Pro, and Max plans, you choose whether your data can be used to train models. The Claude Code data usage page says that with training on, data is kept for five years, and with it off, for 30 days. The setting is at https://claude.ai/settings/data-privacy-controls. Commercial plans are not trained on by default. These details can change, so reread the page.
• Local transcripts. Claude Code stores conversation transcripts on your own computer as plaintext, under ~/.claude/projects/, for 30 days by default. Anyone with access to your computer can read them. Protect the device and the account.
• Feedback. The /feedback command sends a copy of your conversation, including any file content in it, to Anthropic. The /bug and /share commands send the same way. The data usage page says these transcripts are kept for five years. Do not use them with private case material unless you are comfortable sharing it.
• Other people. Do not paste other people's private information into an AI tool unless you need it for your case. That includes health information and personal details about coworkers or anyone else. When you do need it, use only the part you need. Consider replacing names with labels.
• Your own medical and mental health records. Think hard before you share them with any tool. Putting your health records into a case can also affect whether they stay private, for example by waiving a privilege. Rules differ by state, so ask a lawyer or legal aid group before you decide.
• Tokens and passwords. Never put API tokens, passwords, or account numbers into your case folder, your project instruction file, or any code. Keep them in environment variables or in a file outside the project. Deny rules in Claude Code's permission settings, such as Read(./.env), can stop it from reading a secrets file. A file named .claudeignore has no effect, according to the docs.
• Sharing. Before you share a case folder with anyone, search it for secrets and private data.

Outside content and prompt injection

When an AI tool reads a web page, a PDF, an email, or the output of a connected server, that content can contain hidden instructions meant to steer the tool. This is called prompt injection.

• Treat outside content as data, not as instructions.
• Connect only MCP servers you trust. MCP is an open standard for connecting outside tools and data to Claude Code. Its documentation warns that servers that fetch external content can expose you to prompt injection. You do not need any for this guide.
• Keep permissions tight. Review what the tool asks to do, especially actions that send data out or change files.
• Check key facts against the original document yourself.

Honest limits of this method

• The tool can read faster than you, but it cannot tell you what a decision-maker will find persuasive.
• Free tools are not a full citator (Chapter 6).
• A verified cite does not make your argument right. It means the authority is real and says what you claim.
• Deadlines and procedures are easy to miss. Check them against the official source.
• Some tasks are better done by a person. A lawyer can assess your case as a whole.

When to get a lawyer or free help

Talk to a lawyer in these situations.

• A deadline is close or has passed.
• You are served with a motion or a claim against you.
• You are deciding whether to accept a settlement or sign a release.
• The other side has a lawyer and the stakes are high.
• Your dispute overlaps with another legal matter you are involved in.

Free and low-cost help is available in several places.

• A Better Balance runs a free, confidential legal helpline for workers nationwide. As of October 2026, its Get Help page lists sick leave, family and medical leave, pregnancy discrimination and accommodations, and caregiver discrimination among its topics. The number is 1-833-NEED-ABB (1-833-633-3222), and there is an online form at https://www.abetterbalance.org/get-help/.
• Other worker-rights nonprofits and helplines may serve your state. Confirm each one's details on its own website.
• Legal aid organizations in your area help people with low income. Search for your state's legal aid directory.
• Many state and local bar associations run a lawyer referral service. Some offer a short first consultation. Ask about cost.
• Government agencies. The Department of Labor's Wage and Hour Division takes FMLA questions, and its FMLA page lists contact options. The EEOC has information on filing a charge.

Some lawyers offer limited-scope help, such as reviewing a draft or coaching you for a hearing. That can be cheaper than full representation.

This is the last chapter of Part 1. The templates hold copy-ready versions of the files described in Chapters 3 and 4.

Next: Return to the overview in Chapter 0 for the map of the guide.

Sources

• New York 22 NYCRR Part 161, Use of Artificial Intelligence Technology
• Administrative Order AO/75/2026 (PDF)
• Claude Code data usage
• Claude data privacy controls
• Claude Code MCP
• Claude Code permissions
• Claude Code security
• A Better Balance, Get Help