AI-Drafted Wills and the Privilege Trap: What California Families Risk When They Use ChatGPT for Estate Planning

Eighty percent of legal professionals now say AI will have a high or transformational impact on their work within five years, according to Thomson Reuters’ 2025 Future of Professionals Report, up from 77% the year before. Seventy-four percent of lawyers who use AI tools use them for legal research, and a growing share use them to draft documents outright. The same shift is happening on the client side. People who would never have attempted their own will ten years ago are now opening ChatGPT, typing in their family situation, and asking it to draft a trust.

Estate planning is a uniquely bad fit for this. A will or trust doesn’t just need to be technically correct. It needs to survive you: to be readable by a probate court, defensible against a challenge from the sibling who got less, and enforceable years or decades after you signed it, when you are no longer available to explain what you meant. Every other use of AI, drafting a letter, summarizing a contract, researching a question, lets you catch the mistake while you’re still around to fix it. An estate plan is the one document where the person who could catch the error is dead by the time the error matters.

The risk isn’t just that the AI gets something wrong, though it often does. The deeper problem is what happens to the conversation itself, whether the resulting document can survive California’s execution rules, and what a contest lawyer does with the fact that nobody licensed to practice law was ever involved. Each of those is a distinct legal problem, and each one is worse than most people realize.

The conversation you had with ChatGPT is not privileged, and it may already be evidence

When you sit down with an estate planning attorney and describe your family, your assets, and who you want to leave out of your trust, that conversation is protected. California’s attorney-client privilege, codified at Cal. Evid. Code §§ 950-962, lets you refuse to disclose a confidential communication made in the course of the lawyer-client relationship, and it lets your lawyer refuse on your behalf. Cal. Evid. Code § 917 goes further: a privileged communication doesn’t lose its protection just because it was sent electronically, or because a server somewhere along the way technically had access to it. Email to your attorney is still privileged. A shared drive with your attorney is still privileged.

None of that protection exists when you type the same information into ChatGPT, Claude, Gemini, or any other AI platform. There is no attorney-client relationship, so there is nothing for privilege to attach to in the first place. The platform is a third party, one you have no duty of loyalty from and no confidentiality agreement with beyond a consumer terms-of-service page most people never read. What you type gets stored. Depending on the platform and your settings, it may be used to train future versions of the model, reviewed by human contractors, or retained indefinitely on a server you don’t control.

This isn’t theoretical anymore. In February 2026, a federal court in United States v. Heppner ruled that documents a client created using a public AI tool were not protected by attorney-client privilege, even after the client handed those AI-generated materials to his own lawyers. The court allowed the government to obtain them. The reasoning tracks basic privilege law: sharing something with a third party before your lawyer ever sees it can destroy the confidentiality the privilege depends on, and once that confidentiality is gone, handing the material to counsel afterward doesn’t bring it back.

Apply that to estate planning. If you spend an evening prompting ChatGPT with your net worth, which child you’re cutting out and why, the affair your spouse doesn’t know about, or the reason you don’t trust your business partner with a fiduciary role, none of that is protected. If your estate plan is ever contested, that conversation can be subpoenaed. Opposing counsel doesn’t need to guess at your state of mind or your reasons for a disputed provision. They can read your own words, typed to a chatbot, laying out exactly what you were thinking and why. A skilled litigator representing the disinherited child will build cross-examination and undue influence arguments straight out of that transcript. And under Cal. Evid. Code § 912, even a communication that started out privileged can be waived if you later paste it, or describe it, to a non-privileged third party, which is exactly what pasting your lawyer’s advice into a chatbot for a “second opinion” does.

California doesn’t let a chatbot supervise a signing

Even if an AI tool produced a flawless document, and it usually doesn’t, text on a screen is not a will. California cares enormously about how a will gets executed, not just what it says. Cal. Prob. Code § 6110 requires a will to be in writing, signed by the testator (the person making the will), and witnessed by at least two people who are present at the same time, watch the testator sign or acknowledge the signature, and understand that what they’re signing is meant to be a will. Miss any piece of that and the will can fail regardless of how well-drafted the language is.

AI cannot supervise a signing. It cannot confirm that two witnesses were in the room together. It cannot assess whether the person signing had the mental capacity to understand what they were doing, an issue that comes up constantly when an aging parent is signing at a time when capacity is already in question. It cannot flag that the “witness” someone recruited off the street is actually a named beneficiary, which triggers its own problem: Cal. Prob. Code § 6112 creates a presumption that a witness who also stands to inherit procured that gift through duress, menace, fraud, or undue influence, unless there are at least two other disinterested witnesses. That presumption can cost the interested witness their inheritance beyond what they’d have received anyway under intestacy (dying without a will).

Cal. Prob. Code § 6113 adds another layer: a will has to comply with the execution law of California, the place it was actually signed, or the testator’s domicile, in order to be recognized. People who print an AI-drafted document and sign it at the kitchen table, sometimes with a neighbor as the only witness, sometimes with no witness at all because they didn’t realize one was required, are walking directly into these gaps. None of this is exotic. It’s the same failure pattern seen across every wave of do-it-yourself estate planning, from fill-in-the-blank kits to LegalZoom trusts, and it’s covered in more detail in our rundown of online estate planning pitfalls. AI just makes the document look more convincing while doing nothing to fix the execution problem underneath it.

There’s a California-specific trap layered on top of all of this: community property. California treats property acquired during marriage differently from property one spouse owned before marriage or received by gift or inheritance, and that distinction drives who can leave what to whom. An AI tool has no way to know which asset in your household falls into which category unless you explain it in exhaustive detail, and even then it routinely gets the analysis wrong, because the rule depends on facts, like how an asset was titled, whether separate funds were commingled with marital funds, and whether a spouse’s name was added to a deed, that require a professional to actually investigate rather than take your word for it.

Who drafted this, and why it matters more than you think

When a will or trust ends up in a contest, one of the first things a court examines is who actually drafted the document and what their relationship was to the person signing it. California built an entire statutory presumption around this because the pattern of a fiduciary or drafter writing themselves into a windfall is common enough to warrant one.

Cal. Prob. Code § 21380 presumes that a donative transfer to the person who drafted the instrument, or to someone who transcribed it while in a fiduciary relationship with the person making the gift, is the product of fraud or undue influence. For a transfer to the actual drafter, that presumption is conclusive; it cannot be rebutted at all except through the narrow independent-review procedure in § 21384, which requires an independent attorney to counsel the person making the transfer outside the presence of any beneficiary. The statute exists because California has seen exactly this fact pattern too many times: someone close to a vulnerable person writes, or arranges the writing of, a document that benefits them, and then produces it as though it reflects the other person’s free choice.

Now put AI into that fact pattern. If an adult child sits down with a parent’s laptop, prompts ChatGPT to draft a trust amendment that favors that child, and the parent signs whatever comes out, a court asking “who drafted this” has an uncomfortable answer. The child may not be a lawyer, may not have “transcribed” it in the traditional sense, but the substance of § 21380, that the person who authored the document and stands to benefit from it needs independent scrutiny, applies with full force. Contest litigators are already building this argument. An AI-drafted document with no independent attorney anywhere in the process removes the one safeguard, a lawyer with no stake in the outcome asking hard questions before signature, that the whole statutory scheme is built around. If you’re the family member facing that kind of accusation, or bringing one, this is exactly the terrain covered in our guide to trust contests.

There’s a second problem layered on top: construction disputes. AI-generated trust language often sounds fluent and legal without actually being precise, because the model is optimized to produce plausible text, not to resolve every contingency a drafting attorney would think to ask about. What happens if a named beneficiary dies first? What happens to a specific gift if the asset is sold before death? Is “my children” meant to include a stepchild the testator raised but never adopted? An attorney asks these questions during drafting, in the room, before anyone signs anything. An AI tool answers whatever narrow question it was asked and moves on, with no sense of what it didn’t think to ask.

When that ambiguity surfaces after death, California does allow some flexibility. The California Supreme Court held in Estate of Duke (2015) 61 Cal.4th 871 that an unambiguous will can be reformed to match the testator’s actual intent if clear and convincing evidence shows the will contains a mistake in expressing that intent, and also establishes what that actual intent was. That sounds like a safety net, and in the right case it is. But reformation litigation means a probate court hearing evidence, sometimes over multiple days, to figure out what a document actually meant, at a cost the family bears and with an outcome nobody can guarantee going in. Duke is a rescue mechanism for a genuine mistake in a document an attorney otherwise prepared carefully. It was never meant to be, and shouldn’t be relied on as, a substitute for drafting the document correctly the first time.

Once you’ve shared it, you can’t take it back

There’s a version of this risk that has nothing to do with litigation at all. When you type your financial details, family structure, and distribution wishes into a consumer AI platform, you generally have no way to verify whether that input became part of the data used to refine the model, whether it sits in a log reviewed by the company’s staff, or whether it’s retained after you delete the conversation on your end. Some platforms offer settings that limit training use; most people never touch them, and even those settings don’t guarantee the input was never logged, stored, or exposed in a breach before the setting took effect.

Attorney-client privilege has a clawback mechanism built in for exactly this kind of accidental exposure: if privileged material is inadvertently produced in litigation, there are established procedures to get it back and prevent its use. There is no equivalent mechanism for information absorbed into a commercial AI system. Once your trust terms, your reasons for disinheriting someone, or your family’s private history have been typed into that box, there’s no legal process to retrieve it, and no guarantee it stays contained to the conversation you thought you were having.

The American Bar Association addressed this directly in Formal Opinion 512, issued July 29, 2024, the current governing guidance on lawyers’ use of generative AI. The opinion requires a lawyer’s informed consent before inputting client information into a self-learning AI tool, precisely because of this uncontrolled exposure risk, and it makes clear that boilerplate consent buried in an engagement letter isn’t good enough. That’s the standard a licensed attorney is held to before using AI on your matter. Nobody is holding a consumer typing into ChatGPT at home to that standard, but the underlying risk, information that leaves your control permanently, doesn’t disappear just because there’s no ethics rule attached to your own keystrokes.

A new wrinkle: who controls your AI replica after you die

California added a genuinely new category of asset to think about in 2024, when Governor Newsom signed AB 2602 and AB 1836, both effective January 1, 2025. Together they regulate “digital replicas”: AI-generated recreations of a person’s voice or likeness realistic enough to be mistaken for the real thing.

AB 1836 amends Cal. Civ. Code § 3344.1, California’s post-mortem right of publicity statute, to specifically bar producing or distributing a digital replica of a deceased person’s voice or likeness in an expressive audiovisual work or sound recording without consent from whoever controls those rights. That right already ran for 70 years after death and was already transferable by will or trust; the new law simply confirms it covers AI-generated replicas, not just old-fashioned use of a photo or recording. Violators face damages of at least $10,000 or actual damages, whichever is greater, plus attorneys’ fees. AB 2602, meanwhile, protects living performers by voiding contract provisions that let someone create a digital replica of a person’s voice or likeness without a clearly specified scope of use, particularly where the person wasn’t represented by counsel or a union when they signed.

For most families this sounds like an entertainment-industry problem, and for now, most of the litigation under these statutes is concentrated there. But the right itself, control over a deceased person’s voice and image against unauthorized AI recreation, is a property right that passes through your estate plan like any other asset. Under § 3344.1, if you never designate who holds it, it passes to your surviving spouse, then your children, then your parents, in that statutory order, whether or not that’s who you would have chosen. If you have any public-facing presence, recorded audio or video, a business built around your name, or simply want to prevent a future where someone builds an AI chatbot trained on your voice and personality without your family’s consent, this is worth a specific instruction in your trust: who holds that right after you die, how they may license or refuse to license it, and whether you want it exercised at all. It’s a new category, and an AI drafting tool has no framework for raising the question, because it isn’t trained to know it should ask you about it. A human attorney reviewing your estate, your online presence, and your family situation is.

What AI is actually useful for, and where the line sits

None of this means AI has no place near an estate plan. Used the right way, it’s a reasonable research and organization tool. Asking a general question like “what is a revocable trust” or “what does probate mean” to get oriented before a meeting is low risk. Using it to build a checklist of documents to gather, or to draft a list of questions to bring to your attorney, is a fine use of the technology. None of that involves handing over your actual financial details or family situation, and none of it produces a document anyone is going to sign.

The line gets crossed the moment AI moves from explaining concepts to drafting or deciding. Drafting the actual will or trust language, deciding how to divide community versus separate property, interpreting how a beneficiary designation on a retirement account interacts with your trust, or making any decision about who gets what and under what conditions: these require a professional’s judgment applied to your actual facts, not a plausible-sounding answer generated from patterns in text. This is the same distinction ABA Formal Opinion 512 draws for lawyers themselves: AI can assist with a task a competent professional is actively supervising, but it cannot substitute for that professional’s independent judgment, and a lawyer who lets it do so anyway is exposed to a malpractice claim for the uncritical reliance. A consumer who lets a chatbot make those calls for their own estate takes on that same exposure, just without anyone whose job it is to catch it before it becomes permanent. For a closer look at exactly where DIY and AI-assisted planning tend to go wrong, see our piece on DIY AI estate plans.

The bottom line

An attorney-drafted estate plan in California typically costs a few thousand dollars, depending on the complexity of the estate. A contested probate or trust matter routinely runs tens of thousands of dollars in legal fees on both sides, stretches on for a year or more, and often ends with family relationships that never fully recover. AI-drafted documents don’t just carry a higher chance of failing; they fail in the specific ways that produce the most expensive kind of litigation: challenges to validity, fights over ambiguous language, and accusations of undue influence against whoever benefited from a document nobody independent ever reviewed.

The math isn’t close. A properly executed estate plan, reviewed and supervised by someone whose only job during that meeting is to protect your intent and your family, costs a fraction of what a failed AI-drafted plan costs your family to untangle after you’re gone. And the cost isn’t only financial. A privilege problem, an execution defect, or an undue influence fight doesn’t just cost money to resolve; it airs out exactly the family conflict you built the plan to avoid, in a courtroom, in front of the people you were trying to protect from it.

If you’ve already drafted something with AI, or you’re tempted to, the better move is to bring it to a review before it’s signed, not after it’s contested. A drafting attorney can usually fix a flawed provision in an afternoon. Nobody can fix a signature ceremony that already happened wrong, and nobody can un-ring the bell on a conversation that’s already sitting on someone else’s server. Working with an attorney at the drafting stage is the only version of this process where someone is actually accountable for getting it right, and where the document is built to survive the one test that matters: holding up after you’re no longer here to explain it.