AI and the Law: Can AI Write Your Contract? What Lawyers (and Everyone Else) Need to Know

There is a 73-million-person legal help gap in the United States.

According to the Legal Services Corporation, approximately 80% of low-income Americans with civil legal problems — eviction, wage theft, domestic violence, debt collection — receive no legal help at all. The average US attorney costs $200 to $400 per hour. A simple contract review runs $300 to $600. A full lease negotiation can exceed $1,500. For the majority of Americans, the legal system is functionally inaccessible.

AI has entered this gap with enormous promise and equally enormous risk. The same tools that can explain "what does 'indemnify and hold harmless' mean?" in plain English can also hallucinate six non-existent court cases and get the attorneys who submitted them sanctioned by a federal judge.

This guide is about understanding both sides of that equation — honestly.

The Access to Justice Gap AI Is Trying to Fill

Before examining what AI can do, it is worth understanding the scale of what it is trying to solve.

The Legal Services Corporation's 2022 Justice Gap Report found that for every client served by a civil legal aid organization, at least one other person with a legal problem is turned away due to limited resources. The problems left unaddressed include:

• Housing: Evictions, foreclosures, habitability issues
• Family: Domestic violence protective orders, child custody, divorce
• Income and work: Wage theft, wrongful termination, benefits denials
• Consumer: Predatory loans, debt collection harassment, fraud
• Immigration: Status issues, deportation defense, DACA renewals

These are not abstract legal questions. They are the situations that determine whether someone keeps their home, their children, or their ability to work.

The cost problem is structural. Attorneys in the US are required to pass a bar exam, maintain continuing legal education, carry malpractice insurance, and operate within a regulated professional framework. That infrastructure costs money. The $200 to $400 hourly rate is not gouging — it reflects genuine professional overhead. But it also means that legal help is, for most people, a luxury good.

AI cannot fix this entirely. But it can do specific things that reduce the gap — if people understand precisely what those things are and where they end.

What AI Can Genuinely Do for Non-Lawyers

The list of things AI does well in a legal context is real and growing. Here is an honest accounting.

Explain Legal Concepts in Plain English

This is where AI genuinely shines. Legal documents are written in a dialect of English specifically designed to be precise, not readable. AI excels at bridging this gap.

Ask an AI:

• "What does 'indemnify and hold harmless' mean in a freelance contract?"
• "What is the difference between arbitration and mediation?"
• "What does 'force majeure' mean and when does it apply?"
• "Explain what this non-compete clause actually restricts me from doing."

A good AI response to these questions is accurate, plain-language, and genuinely educational. This is legal information, not legal advice — and the distinction matters legally, as we will discuss. But information is enormously valuable. Understanding what you are reading is the prerequisite for every other legal decision you make.

Summarize Contracts and Flag Unusual Clauses

A forty-page software license agreement is impenetrable for most people. AI can read it in seconds and produce a summary of key terms, identify clauses that deviate from standard practice, and flag areas that a human reviewer should examine.

What AI can catch:

• Unusually broad intellectual property assignment clauses (e.g., a freelance contract claiming ownership of all work you create, even on your own time)
• One-sided indemnification provisions that expose you to unlimited liability
• Automatic renewal clauses buried in subscription agreements
• Mandatory arbitration clauses that waive your right to sue in court
• Non-disparagement clauses that restrict what you can say publicly after the relationship ends

What AI may miss:

• Whether a clause is enforceable in your specific state
• How a clause interacts with local law that supersedes the contract
• Whether the party you are contracting with has a track record of abusing specific provisions
• Anything requiring knowledge of facts not in the document

Generate First-Draft Templates for Common Agreements

For routine, lower-stakes documents, AI-generated templates are a practical starting point. Common use cases:

• Non-disclosure agreements (NDAs): Standard mutual or one-way NDAs for business discussions
• Simple service agreements: Freelance contracts for web design, writing, consulting
• Basic lease agreements: Month-to-month rental agreements for low-value arrangements
• Partnership MOUs: Non-binding memoranda of understanding between collaborators
• Independent contractor agreements: Basic agreements establishing contractor (not employee) status

The key word is starting point. A first draft is not a finished document. It requires review against the specific facts of your situation, applicable state law, and ideally a human attorney's eye before you sign anything.

Research Case Law and Statutes (With Verification Required)

AI can help you understand what the law says about a specific situation and point you toward relevant statutes and case law. This is useful for understanding your rights before a consultation, preparing questions for an attorney, or deciding whether a situation warrants legal action at all.

The critical caveat: AI models, including GPT-4, Claude, and Gemini, are known to hallucinate legal citations. They produce case names, docket numbers, and quotes that sound authoritative and do not exist. This is not a minor caveat — it is the single most dangerous failure mode in AI legal assistance, and it has already resulted in federal court sanctions (more on that shortly).

Rule: Every case citation from an AI must be verified on Westlaw, LexisNexis, Google Scholar, or the court's official PACER system before you rely on it for anything.

Help You Understand What Type of Lawyer You Need

One of the most practical uses of AI in legal contexts is navigation. The US legal profession is deeply specialized. Someone with an employment dispute may need an employment attorney rather than a general practice lawyer. Someone facing a landlord who won't return a security deposit may need a tenant's rights specialist, a small claims court guide, or simply the local housing authority complaint process.

AI can help map your problem to the right kind of professional — and, importantly, help you identify when you do not need an attorney at all (many small claims matters, for instance, are designed for self-representation).

Prepare for a Consultation with an Actual Attorney

A $300 attorney consultation is far more valuable when you arrive having already organized your facts, identified your key questions, understood the basic legal framework, and assembled your documents. AI can help with all of this preparation — making the expensive human hour go further.

What AI Absolutely Cannot Do

This section is not a disclaimer. These are real limitations with real consequences.

Give Jurisdiction-Specific Legal Advice

US law is a patchwork. Employment law, landlord-tenant law, family law, and contract law vary significantly by state. What is enforceable in Texas may not be in California. What requires written notice in New York may be implied in Georgia. Non-compete agreements are broadly unenforceable in California but routinely enforced in Florida.

AI training data is not jurisdiction-specific by default. A model trained on the general corpus of legal text may produce advice that is accurate for some states and wrong for others — and it often will not reliably distinguish between the two. If you ask "is my non-compete enforceable?" and do not specify your state, jurisdiction, and specific circumstances, any answer you receive is incomplete at best and dangerously wrong at worst.

Even when you specify jurisdiction, AI may be outdated. Laws change. Courts issue new rulings. Statutes are amended. A model with a training cutoff of late 2024 may not know about a 2025 ruling that changes everything about the question you are asking.

Represent You in Court

This one requires no explanation, but it is worth stating: AI cannot appear in court, file motions on your behalf, negotiate with opposing counsel, or exercise any of the professional judgment functions of legal representation. Tools that help you prepare your own court filings are useful; they are not representation.

Replace Licensed Legal Advice for High-Stakes Situations

There are categories of legal matter where the stakes are too high and the fact-specificity too great for AI to meaningfully substitute for a licensed attorney:

• Criminal defense: Jail time is irreversible. A missed suppression motion, a bad plea deal, or an overlooked sentencing guideline can define the rest of someone's life.
• Divorce and child custody: These proceedings determine where children live and with whom. They require local court expertise, knowledge of the specific judge, and strategic judgment that no AI possesses.
• Immigration: Errors in immigration filings can result in deportation, multi-year bars from re-entry, and permanent status consequences for family members.
• Business disputes over significant sums: A contract dispute involving $100,000 requires an attorney who understands the specific jurisdiction, the applicable statute of limitations, and the strategic calculus of litigation versus settlement.

Guarantee Accuracy

AI models hallucinate. This is a documented, persistent, and not-yet-solved characteristic of large language models. They produce confident-sounding text that is factually wrong. In most contexts, this is an inconvenience. In legal contexts, it can result in sanctions, lost cases, or agreements that expose you to liability you did not intend to accept.

Sign Documents or Serve as a Witness

AI cannot execute a contract, serve as a notary, or act as a witness to a signature. These are roles that require legal personhood, which AI does not have.

The Case That Changed Everything: Mata v. Avianca (2023)

If you want to understand why the legal profession's relationship with AI changed permanently in 2023, you need to understand this case.

The background: Roberto Mata was suing Avianca Airlines for injuries sustained on a flight. His attorney, Steven Schwartz of Levidow, Levidow & Oberman, used ChatGPT to research case law supporting the claim.

What happened: ChatGPT produced a list of cases, including Varghese v. China Southern Airlines, Shaboon v. Egyptair, Zicherman v. Korean Air Lines, and three others. They sounded real. They had plausible docket numbers. They cited appropriate courts and years. Schwartz filed a brief relying on these cases.

The problem: None of these cases existed. When Avianca's attorneys tried to locate them, they could not. When the federal judge, P. Kevin Castel of the Southern District of New York, asked Schwartz to produce the cases, Schwartz submitted fabricated case summaries — printouts of ChatGPT responses presented as if they were real legal documents.

The consequences: Judge Castel sanctioned Schwartz, his partner Peter LoDuca, and the firm for filing a brief "replete with citations to non-existent cases." The attorneys were required to pay opposing counsel's legal fees and notify each judge named in the fake opinions. The case received international coverage and became the defining example of AI hallucination in a legal context.

The lesson is not that attorneys should avoid AI. The lesson is that AI-generated legal citations must always, without exception, be verified through authoritative legal databases before they appear in any document that matters.

Judge Castel's opinion was direct: the attorneys' error was not using AI. Their error was trusting AI output without verification.

AI Legal Tools Built for Professionals

The legal technology industry has developed purpose-built AI tools with guardrails that general chatbots lack. These are what law firms and corporate legal departments are actually deploying.

Harvey AI is the headline name. Backed by OpenAI and used by elite law firms, Harvey is trained specifically on legal materials and designed to minimize the hallucination risk that plagued early general-purpose AI deployments. It is not a consumer product — it is a professional tool sold to law firms that have the expertise to review its outputs critically.

Casetext's acquisition by Thomson Reuters for $650 million in 2023 was the clearest signal that established legal data companies understood what was happening. By integrating Co-Counsel with Westlaw's verified database, Thomson Reuters created a research tool where AI output is grounded in confirmed, verifiable sources — a direct response to the Mata problem.

E-discovery is where AI has arguably made the deepest inroads in the legal profession. Relativity AI and similar tools review millions of documents for litigation, identifying privilege, relevance, and key evidence at a speed and scale no human team could match. E-discovery AI is not making legal judgments — it is doing pattern recognition on document sets — which means the hallucination risk is lower and the productivity gains are real.

AI Tools for Consumers and Small Businesses

The consumer legal AI space is more uneven in quality and more contested in terms of what it can legally claim to do.

DoNotPay: Promise, Reality, and Legal Trouble

DoNotPay launched in 2015 as an automated tool for contesting parking tickets. By 2023, founder Joshua Browder was calling it "the world's first robot lawyer" and claiming it could handle everything from canceling subscriptions to fighting bank fees to negotiating medical bills.

The reality was more complicated.

In early 2023, DoNotPay announced it would pay $1 million to anyone who would use its AI to argue before the US Supreme Court, with an AI voice in their ear during oral argument. The stunt was called off after state bar associations threatened sanctions against anyone who participated.

Later in 2023, DoNotPay settled a class action lawsuit brought by the National Association of Consumer Advocates, which alleged the company had overstated what its tools could do. Browder's representations that the AI could perform at "the level of a top human lawyer" were not supported by the product's actual capabilities.

What DoNotPay actually does reasonably well:

• Automated generation of dispute letters (billing errors, subscription cancellations)
• Small claims court filing guides
• Basic FOIA request templates
• Generating initial demand letters for common consumer disputes

What it cannot do: Provide the genuine legal analysis, jurisdiction-specific advice, and strategic judgment that its marketing sometimes implied.

The DoNotPay story illustrates a broader pattern in consumer legal AI: the gap between marketing claims and product reality is often large, and the people most likely to be harmed by that gap are the people who most need legal help and have the fewest resources to absorb a bad outcome.

Other Consumer-Facing Tools

Spellbook integrates directly into Google Docs and Microsoft Word to provide AI-powered contract redlining. It suggests changes, flags problematic clauses, and proposes alternative language. For small businesses negotiating vendor agreements, it is a genuinely useful tool — with the standard caveat that its suggestions require human review.

LegalZoom has long provided document generation for common legal needs (LLCs, wills, simple contracts). Its recent AI integrations speed up the drafting process and add some automated review capabilities. Critically, LegalZoom offers an attorney review option for an additional fee — this hybrid model (AI draft, human review) is the most defensible approach for documents you will actually rely on.

Lawdingo and Avvo focus on lawyer-matching with AI-assisted preliminary assessment to help users identify the right type of attorney for their situation.

General-purpose AI (Claude, ChatGPT, Gemini) can draft surprisingly competent first-pass legal documents for low-stakes situations. Ask Claude to draft an NDA for a two-person business meeting and you will get a workable template. But these tools have no legal-specific guardrails, no jurisdiction training, and no verification layer for citations — which means their outputs require more careful review than purpose-built legal AI.

The Unauthorized Practice of Law Problem

Here is the legal question that hangs over the entire consumer AI legal market and has not been resolved: does AI providing legal assistance constitute the unauthorized practice of law?

In the United States, giving legal advice is restricted to licensed attorneys. "Legal advice" means applying the law to specific facts and recommending a course of action. "Legal information" — explaining what the law says in general — is permitted for anyone. The line between the two is often not clear in practice, and the distinction is enforced by state bar associations with varying levels of rigor.

When DoNotPay's AI tells you "based on your situation, you should send a dispute letter citing the Fair Credit Billing Act and request a chargeback within 60 days," is that legal information or legal advice? The answer matters legally.

DoNotPay's position was, essentially, that it was a robot and therefore UPL rules designed for humans did not apply. This argument did not succeed in court. Courts have generally held that the practice of law is defined by what activity is occurring, not by whether the entity performing it holds a law license or has a human body.

The FTC, state bar associations, and consumer protection regulators are all watching this space. The regulatory question is genuinely unresolved, and the answer will shape how the industry evolves over the next five to ten years. For now, the practical implication is this: be skeptical of any AI tool that claims to provide legal advice rather than legal information, especially for high-stakes situations.

How to Use AI for Legal Documents Without Getting Burned

With that full picture in mind, here is a practical framework for using AI in legal contexts responsibly.

The Tiered Approach by Stakes

Specific Practices That Reduce Risk

1. Verify every case citation independently. Never include a case citation from an AI in any document without checking it on Westlaw, LexisNexis, Google Scholar, or PACER. This takes thirty seconds. The alternative is what happened in Mata v. Avianca.

2. Be explicit about jurisdiction in your prompts. Do not ask "is this non-compete enforceable?" Ask "is this non-compete clause enforceable under California law for a software engineer earning $150,000 per year?" The more specific you are, the more useful and accurate the response.

3. Compare AI output to official government forms. Courts provide free templates for many common filings — small claims, eviction responses, name changes, simple wills. These official forms are written by lawyers, approved by the court system, and updated when the law changes. They are often better starting points than AI-generated documents, and comparing AI output to official forms is a useful quality check.

4. Use AI for preparation, not replacement. The highest-value use of AI in legal contexts is making your time with a human attorney more productive. Use AI to understand the issues, organize your facts, prepare your questions, and draft preliminary documents — then bring all of that to a consultation where a licensed attorney can apply professional judgment.

5. For anything over $5,000 or with criminal implications, use a human lawyer. This threshold is a rough guide, not a law. But it reflects the point at which the cost of an attorney review is small relative to the stakes of getting it wrong.

AI, IP Law, and Who Owns What AI Generates

One area where AI and the law intersect in a legally unsettled way is intellectual property. If you use AI to generate a contract, a marketing document, or any other written work product, questions of ownership arise.

The US Copyright Office position (as of 2025-2026): AI-generated content without meaningful human creative input is not eligible for copyright protection. Works with "sufficient human authorship" — where a human makes creative choices about what AI generates, edits, and selects — may be protectable. The exact line is still being defined in litigation.

The major pending cases:

• Getty Images v. Stability AI: Getty alleges Stability AI used millions of its copyrighted images to train Stable Diffusion without license or compensation. The case is proceeding through UK and US courts and will set precedent for training data liability.
• The New York Times v. OpenAI and Microsoft: The Times alleges ChatGPT was trained on and reproduces its journalism without license. The case raises questions about fair use, training data, and output reproduction that will define the legal framework for foundation model companies.

What this means for businesses using AI-generated content:

For AI-generated text (contracts, marketing copy, blog posts), the copyright situation is murky but the practical risk is manageable: if you substantially edit and customize AI output, you likely have some claim to the result. For AI-generated images used commercially, the Getty v. Stability AI case creates real uncertainty about whether the training data underlying those images was properly licensed.

The safest approach for businesses: use AI as a drafting and ideation tool, make meaningful editorial choices, and document your human involvement in the process. For high-stakes content (marketing campaigns, product materials, anything you might need to defend in court), consult an IP attorney about your specific situation.

The Broader Pattern: AI Companies Facing Legal Accountability

It would be incomplete to discuss AI and law without noting that AI companies are themselves increasingly the subject of legal action.

The class action against Anthropic over Claude Max's advertised usage limits — filed in California federal court in June 2026 — alleges that the Claude Max 5x and 20x subscription plans did not deliver the usage capacity implied by their pricing and marketing. The case is a consumer protection action, not an AI-specific legal theory, but it illustrates that AI product companies are subject to the same false advertising and consumer protection laws as any other company.

This matters for the legal AI space specifically. When DoNotPay overstated what its AI could do and was sued for it, the legal theory was straightforward consumer protection. When AI legal tools make claims about their capabilities — accuracy rates, citation reliability, compliance with jurisdiction-specific law — those claims are potentially subject to the same scrutiny.

The legal system is catching up with AI. The Mata v. Avianca sanctions, the DoNotPay settlement, the FTC's increasing attention to AI product claims, and cases like the Anthropic Max lawsuit are all data points in the same direction: AI companies are not exempt from legal accountability for what they claim their products do.

Understanding how AI companies approach security and manipulation of their own systems is part of understanding why AI legal tools built for professional use have different safeguards than general-purpose tools — the professional legal market demands accountability that the consumer market is still working out how to require.

A Realistic Picture of Where This Goes

The access to justice gap is real and AI has genuine potential to help close it. The technology is already doing useful work: explaining legal concepts, drafting templates, summarizing documents, and helping people understand their situations well enough to make better decisions about when they need a lawyer and what kind.

The risks are also real. Hallucinated citations are not a theoretical concern — they resulted in federal court sanctions that will be cited in legal education for decades. Jurisdiction-specific accuracy is a genuine limitation of current models. The line between legal information and legal advice is contested territory.

The honest picture is this: AI is a useful tool for legal self-education and preliminary document drafting, with real limitations that matter most for the people who most need legal help. The person facing an eviction with no money for an attorney benefits enormously from AI that can explain the process, identify the relevant statutes, and help them draft a response. That same person is also the one most exposed if the AI is wrong about their jurisdiction's notice requirements.

Use the tools. Use them carefully. Know where they end. And for the situations where the stakes are highest — the criminal charge, the custody dispute, the immigration filing — find a human lawyer, even if that means legal aid, a law school clinic, or a pro bono program.

The access to justice gap is not a problem AI will solve alone. But used with clear eyes about what it can and cannot do, AI can make legal help meaningfully more accessible for people who have had effectively none.