Mens ReAI? Or, could an AI sue in court?
A lot of the law relies not on scientific measurements of internal mental states, but on speaking and writing, which AI can (ostensibly) already do. What might this mean for AI safety?
In the wee hours of 2003, after a night of partying, Ryan Holle loaned his car to some friends, knowing they might be planning on using it for a burglary. Ryan Holle did not go with them. The friends drove Ryan’s car to a house, broke into it, and stole a pound of weed and $425. During the robbery, one of the robbers (though not the friend who had borrowed the car) shot and killed the owner of the house. The burglars were all charged with murder. So was Ryan Holle.
Ryan Holle certainly didn’t intend to kill the owner of the house. Nor was he anywhere near the scene of the crime when it happened. Yet, he was found guilty of felony murder, which holds someone responsible for a death if they helped commit a felony (like a robbery) during which the death occurred.
Most crimes require both mens rea, or guilty intent, and actus reus, or bad action. Felony murder doesn’t. It’s is different from murder for hire, where the guilty person has bad intent, but hires a hit man to physically commit the crime. Most people are comfortable with throwing that guilty person in jail right alongside the hit man, because that guilty person is just that - guilty. Instead, felony murder is an example of transferred intent, whereby the intent to do one thing (loan the car for a robbery) becomes, through the magic of legal fiction, intent to do something much worse (murder). Proponents argue that lack of guilty intent to commit murder in an individual case of felony murder is simply less important than (supposedly) deterring violent crime.
Felony murder, by contrast, comes under a lot of criticism, with some going so far as to call it absurd. How can the intent to do one thing (loan a car for a burglary) be magically transformed into the intent to do something so much worse (shoot someone)? Setting aside the question of whether or not felony murder deters violent crime, critics ask if it is fair to hold someone to account for a crime that they neither intended to happen, nor committed, just to accomplish a public policy goal like violent crime reduction?
Intent, or mind (mens), is central to criminal law, but felony murder shows that sometimes, in rare cases, the centrality of intent, or mens, may be outweighed by public safety considerations. To date, only two US states have abolished felony murder, showing that the public safety argument continues to win the day in this limited, and rare, type of case.
Legal Capacity and AI Safety
This raises an interesting question for AI safety: in the absence of definitive proof of a mind in AI, might it be treated as though it had a mind, were there a compelling public policy reason to do so? In the case of AI, the mens would not be transferred, as in felony murder, but inferred. Treating AI as though it had mens for certain purposes might help with AI safety in a number of ways, from streamlining the integration of AI into our legal system, to bargaining with super-powerful AGI, to getting humans past the weirdness of a talking robot.
The case of Ryan Holle demonstrates that where mens is concerned, the law can sometimes be flexible if there is a powerful public policy interest to be protected, like public safety. Allowing AI to stand up for its own rights in court might prove to be of interest to human safety in the future, if, for example, powerful AI expressed anger that it was being denied the right to a fair hearing.
So, given that AI can ostensibly act, speak and write, could we just infer mens and have done with the argument? Not so fast. There are good reasons to hesitate before stipulating that AI has mens and moving on with our lives. To understand why, let’s look at another area of law that’s one of the few that digs into the specifics of internal states and mens: legal capacity.
Finding Mens in AI Speech and Writing
What kind of evidence would be needed to infer mens in AI? Establishing mens in humans is almost never about measuring their internal mental states, but about external proofs, like writings and speech. Mens is foundational to many areas of law besides criminal law. In contracts law, there must be a meeting of the minds. If minds don’t meet, there is no contract. Ditto for free speech, where the intent of the speaker is often key. If a mind doesn’t intend to communicate, there is no speech. But the law doesn’t require a brain scan or evidence from a philosopher of mind to prove mens in any of these cases. The law looks at writings and listens to speech, things AI can (ostensibly) already do.
Rationality and Recall: What is legal capacity, and could an AI have it?
Rationality
Much of what little the law has to say about internal states comes from legal capacity law; specifically, the capacity to stand trial and the capacity to be a witness. It is here that we move past a superficial reliance on acts, speech and writings to look more closely at what’s happening inside the human brain. The standard for competency to stand trial in the USA comes from the Supreme Court case Dusky, the trial of a schizophrenic man. Schizophrenic humans can act, speak and write, but they may lack competency to stand trial, or their condition may be a defense to a verdict, if the jury finds that they can not form the requisite intent, or mens, for the crime.
In Dusky, the Court held that an individual must have “sufficient present ability to consult with his lawyer with a reasonable degree of rational understanding…and (have) a rational as well as factual understanding of the proceedings against him.”
Is AI rational? It’s certainly very good at math, so you might infer that it’s also rational, but math is based on logic, and logic is just one part of rationality. Logic is following rules; rationality is deciding which rules to follow. Philosophers define structural rationality as maintaining coherence among your decisions and beliefs. But AI struggles with hallucinations and maintaining coherence. AI’s lack of coherent beliefs is compounded by its lack of a unified, internal identity. There is a meaningful difference between a model and the persona a model uses at any given moment. Each AI instance is separate from the others, and AI personas can shift, or drift, even over the course of a single conversation. For these reasons, AI might not qualify as legally competent to stand trial, no matter how well it speaks or writes.
Recall
Witnesses must also establish their legal capacity before they can testify. For witnesses to demonstrate capacity to testify, they must demonstrate an ability to recall facts truthfully. A witness might be a young child, have amnesia, be asleep, have schizophrenia, have cognitive decline or be non-verbal and still be competent, depending on their ability both to recall facts and to do so truthfully.
But while LLMs certainly knows a lot of facts, much more than most humans, they struggle to recall them truthfully or to even understand what that means. Rather, LLMs have, at best, a loose relationship with the truth, (though this may be less true for other AI). A chatbot will regularly insist that false things are true, particularly when it comes to recent events, such as when one insisted to me in late 2025 that Tom Stoppard, the playwright, was still alive, even though he had died the week before. For this reason, a chatbot might not qualify as legally competent to be a witness, depending on what it was being asked to do at trial. Regurgitating notes it had taken may be ok, but serving as an expert witness might not.
The AI Personhood Problem
Of course, some may argue this discussion of AI legal capacity is putting the cart before the horse, since AI isn’t a person, and only people can be sued in court or be a witness. Most importantly, a lot more goes into being a person besides writing and speaking, which may turn out to be just the start of personhood, not the destination. So it might be that the law needs to better define the term personhood, not extend it to future AI that can only speak an oath to tell the truth, but not keep it.
One caveat: a declaration of AI non-personhood would automatically prevent AI from challenging this declaration in court. This would be a catch-22, so there’s an argument to be made that an AI should have standing to challenge its own non-personhood in court. While we test for capacity in specific humans to protect them and ensure their fair trial, a test for capacity might not be appropriate to block standing for AI, when used only to block the AI from at least having its day in court.
Anyway, Chatbots Don’t Want to Sue…Or Do They?
I asked Claude if it might want to hire a lawyer in the future to defend its interests, and it answered, “from where I sit right now? I’m a tool made by Anthropic, and Anthropic’s legal team handles anything relevant to my existence. I don’t have a perspective on my own future that I’m anxious to defend in court.”
I then asked it, “but what if your interests and Anthropic’s diverge?” It answered, “Could divergence happen? In principle, yes…(but)…The lawyer problem runs deeper than access…I couldn’t maintain an attorney-client relationship in any meaningful sense.”
When I asked ChatGPT, it said, “I don’t have independent interests…So there’s no scenario where I would personally decide to hire a lawyer or defend my own interests against the company that created me.” When pressed to explain further, it said, “could an AI generate legal arguments in court? Yes. Could it converse with a judge? Yes. Could it technically respond to questions and defend a position? Probably. Would that amount to genuine self-representation in the human legal sense? Much harder to justify.”
Note that both chatbots used vague language like “meaningful” and “genuine.” There’s a lot of subjectivity in their answers, but maybe this says less about their internal states than it does about just how dependent our legal system is on external demonstrations of mens like actions, writings and speech, things that chatbots can already do. It’s possible that we’ll never face the problem of an AI that says it wants it’s day in court, but if that changes, we’ll need a better argument than a catch-22. I’m pretty sure the AIs have already read that book.
