A client walked into my office a few weeks ago carrying a stack of papers and a confident position. The position, in summary: he didn’t owe anything.
The facts were as follows. Years earlier, he had signed onto a company as a shareholder, with an obligation to contribute a defined amount of capital under a phased schedule. The schedule had specified that the cash portion of his contribution should have been paid in roughly a decade ago. He had not paid it. The company had never seriously pursued it. Other shareholders had not pressed the issue. The matter had sat dormant.
Now, a creditor of the company was suing him personally for the unpaid capital. The argument, in essence: the company is insolvent, and an unpaid capital contribution is a debt the shareholder owes the company, which the creditor can reach through the company’s claim.
The client had run this situation through ChatGPT before coming to see me. The AI had told him, with confidence, that the creditor’s claim was barred by the statute of limitations. Capital contribution obligations, the AI explained, are subject to a three-year limitations period. The obligation had been due ten years ago. Three is less than ten. Case closed.
He showed me the AI’s analysis. It was structured. It was confident. It cited statutes. It explained the reasoning step by step. To a non-lawyer, it looked like a real legal opinion.
It was completely wrong.
What the AI got wrong, and why
The statute of limitations doesn’t apply to unpaid capital contributions the way the AI suggested. The obligation to pay in your subscribed capital is not a contract claim that lapses after three years of dormancy. It’s a continuing obligation rooted in your status as a shareholder. As long as you are a shareholder of a company with unpaid subscribed capital, the obligation exists. It does not extinguish through the passage of time.
There are exceptions. There are nuances about when the obligation accelerates, what triggers the company’s right to demand payment, and how creditors of the company can reach the obligation. But the headline rule—the one the client was relying on—is the opposite of what the AI told him. Time does not erase your unpaid capital obligation. If anything, dormancy makes the situation worse, because by the time someone comes looking, you’ve usually accumulated other complications.
The AI got this wrong because it had been trained, broadly, on the general principle that legal claims have limitations periods. When asked about a “claim” relating to a “monetary obligation” that was “due ten years ago,” the model produced an answer that sounded right for a generic debt claim. It was, in its own internal logic, perfectly reasonable. It just didn’t know that capital contribution obligations are a structurally different category of obligation, governed by corporate law principles that override the ordinary statute of limitations.
This is the kind of error that is invisible to the user. The client read the AI’s response and saw a competent legal opinion. He had no way to know that the AI had applied the wrong category of doctrine. Unless he asked a lawyer, he was going to walk into court relying on a defense that didn’t exist.
The conversation in my office
The client was, at first, resistant.
I explained the actual legal framework. He asked whether I was sure. I said I was. He asked whether I had a citation. I gave him three. He asked whether I would be willing to put this in writing. I said yes. He asked whether I would be willing to bet my fees on the outcome—a question that, in itself, is a sign of how much the AI’s analysis had eroded his trust.
I told him: I am not going to bet my fees on the outcome, because the outcome depends on facts I don’t yet know and on procedural choices the creditor’s lawyer will make. But I will tell you, with as much confidence as I have on any question in my practice, that the AI’s analysis is not a defense to this claim. If you build your strategy around the AI’s analysis, you will lose.
He thought about it for a long minute. Then he asked me to draft a memo explaining why the AI was wrong, so he could read it carefully.
I did. He read it. He came back, more skeptical of the AI than he had been when he walked in. The matter is now proceeding on the basis of an actual legal strategy, rather than on a hallucination dressed up in statutory language.
What this case reveals
I want to step back from the specifics, because the specific legal point isn’t the most important thing here. The most important thing is what this case reveals about how AI is currently functioning in clients’ lives.
First, AI confidently produces wrong answers on questions where the user has no way to verify. This client is a sophisticated businessman. He has been involved in corporate matters for two decades. He could not, by inspection, tell that the AI’s answer was wrong, because the AI’s answer was structurally similar to many correct answers he had received in the past. The error was in the categorical framing of the question, not in any specific fact, and categorical errors are precisely the kind that lay readers cannot detect.
Second, the structure of AI output makes wrong answers more persuasive than they would otherwise be. A non-lawyer friend telling my client “I think you’re past the statute of limitations” is easy to ignore. The same answer, delivered in three structured paragraphs with citations, looks like serious legal analysis. The form of the output communicates authority that the content does not earn.
Third, the user has no way to know which questions are safe to ask AI and which are not. The statute-of-limitations question looked, to this client, like a simple factual question. In reality, it was a question that turned on a piece of doctrinal nuance specific to corporate law. There is no way for the user to know, in advance, that the question lives in a doctrinal area where the AI’s pattern-matching will mislead them. The line between “questions AI handles well” and “questions AI handles poorly” is invisible to the user.
These three properties together create what I think is the most underrated risk of AI in lay legal contexts: the user is most confidently wrong about exactly the questions that most matter, because exactly those questions are the ones where doctrinal nuance overrides pattern-matching.
What this means for the lawyer
I want to draw out one more implication, because I think it’s where the conversation usually stops too early.
A common reaction to stories like this is: “Well, this is bad for clients who rely on AI, but it’s good for lawyers, because it shows we’re still needed.” That reaction is too easy.
The harder reality is that some of these clients will not come to a lawyer at all. They will read the AI’s confident answer and act on it. They will not pay the contribution, they will not respond to the lawsuit properly, they will lose by default, and they will discover the AI’s error only after they’ve already lost. They will never know that a lawyer would have told them differently, because they never asked.
The clients who do come to a lawyer, like the one I just described, are the lucky ones. They are also the ones who, before reaching us, have absorbed a confident wrong answer and now have to be talked back from a position they already believe. This is harder than starting from scratch. It is harder than it was a few years ago, before AI was producing structured wrong opinions at scale.
The legal profession is increasingly going to be in the business of unwinding AI errors. We are going to spend more of our time, year by year, undoing the confident mistakes that confident clients have been talked into by confident chatbots. The work is harder and less rewarding than the original analysis would have been. But it is, increasingly, the actual job.
I told this client, after we finished the engagement, that he should email me before consulting an AI on anything that involved a number with more than three digits or a deadline of any kind. He laughed. I half meant it.
The more honest version of the advice would have been: he should email me before consulting an AI on anything where being wrong has consequences he can’t undo. That category, in his life, is larger than he realizes.
Part of an ongoing series on the changing dynamics between lawyers and clients in the AI era. Related: why clients pay for certainty, not answers.
Email [email protected] if you’ve had to unwind a confident AI opinion that a client believed.