What does Generative AI have to do with the truth?
I had a strong reaction when I read a recent article by Tom Barnett in Fast Company. It was called: “Artificial stupidity: Can GPT handle the truth.” My questions were immediate, and raised my blood pressure: can lawyers handle the truth? can humans handle the truth? what is “the truth”? Help me Tom Cruise! Seriously though, if I am using a natural language model based on machine learning from many examples of humans trying to communicate, won’t that model struggle with the same things we struggle with daily? Do we humans have the answer to the question: What is “the truth”? Can AI handle the truth? Does it need to?
Clearly, it’s the idea of truth seeking that bothers me. Focusing on whether a GPT AI can handle the truth just blithely skips over a major philosophical debate in the profession. And does so seemingly just for the opportunity to bash a trending buzzword. This really is not a new debate. In a famous 1967 opinion, Justice Byron R. White said in his dissent in United States v. Wade (388 U.S. 218): “as part of our modified adversary system . . . we countenance or require conduct which, in many instances, has little, if any, relation to the search for truth.” Let’s apply that lesson to our modern technology. Is a lawyer who is using generative AI looking for the truth, or the best outcome for the client?
The justice system is an adversarial system based on client interests. Believing such a system can routinely produce ‘truth’ is hopelessly naïve. It is a system with bowling lane bumpers for professional conduct, and a wide lane for creative problem-solving. GPT can handle the truth, it’s just that lawyers aren’t asking for it.
The focus on whether AI can handle the truth overlooks a deeper philosophical debate about truth in the legal profession, hinting at a mismatch between technology’s capabilities and the complexities of human truth-seeking.
Truth v. Facts
There’s a great article in Psychology Today that leads with this amazing quote from Mark Twain: “It ain’t what you don’t know that gets you into trouble.” The article continues with, “It’s what you know for sure that just ain’t so.” The article goes on to discuss how we think of truth. We mix the idea of truth with other ideas like reality, knowledge, accurate information, certainty, and facts. In the realm of legal opinions and legal reasoning, what passes for truth can be truly incredible. Legal precedents on slavery, segregation, forced sterilization, and Japanese internment are prominent examples of what we once accepted as true. We are ‘sure we know for sure’ what a legal answer is, and we are often wrong. ‘Legal truth’ is not ‘truth’.
I do not want to drift into an analysis of psychology and philosophy. That isn’t the point of this piece. The point is grounded in the tech. When a language model is trained on the probability of word occurrence in a given sentence, it does not seek the truth. No one should believe that it does. It can produce a fact. However, even a task as simple as generating a fact is fraught with peril for the untrained user.
If I ask a generative AI what color a banana is, I’ve asked a terrible question. The “truth” can be the obvious answer of yellow, but it could also be green, green-yellow, yellow-brown, or brown (among plausible others!). AI can handle the truth of determining a banana’s color. I just have to ask the question in the right way. I also need to provide the right data. My use of the tool may lead to an untruth. That is not due to the inherent ability of the tool.
Fundamentally, there is no objective truth for comparing the output of a generative AI. There is also no objective truth for comparing the output of a lawyer. Attorneys produce output that requires legal minds to wrestle over creative interpretations of existing precedent. This process extends our understanding. Legal reasoning is much more about creative problem solving than the regurgitation of legal facts. Lawyers have been creative problem-solving with half-truths for as long as there have been legal problems to solve. Adding a new tool isn’t going to hurt, and it might even help.
Legal truths often diverge from objective facts, a distinction that highlights the subjective nature of legal reasoning and the limitations of AI in providing the fact-rich information some people desire.
Developing Creative Solutions to Legal Problems
In the Fast Company article, the author compares ChatGPT to a hypothetical law associate who is always making stuff up. What the article misses is describing what the hypothetical law associate is actually working on. The author discusses an out-of-control law associate. This associate prepares convincing material. However, this material is “totally false” for the article’s hypothetical law firm. But what if that’s what the client needs, and the law firms wants? Let’s not take it a step too far. The model rules of professional responsibility do set good limits. “Totally false” is too far. However, “random hallucinations” and “making things up out of whole cloth” could be within the bounds of professional practice. Sometimes we need creative solutions to novel legal problems.
Our laws are old, and new laws are not always written well. There is a constant need for new and creative interpretations of existing laws. The pace of AI development aptly demonstrates a critical point. Law and policy changes often lag behind the pace of technological growth. This aggressive pace means even more creative interpretation must be done with the laws that currently exist. All together this means there are many attorneys all over the world trying to develop creative solutions to new problems. As a result, there are going to be some random hallucinations and made-up solutions—and that’s just fine.
The question then is not whether AI can handle the truth. The real query is whether its creative reasoning based on user input can be useful. Can this benefit the user and everyone else? Sometimes there is no truth, sometimes what is needed is creativity and an extension of existing (and even attenuated) logic. Guess what? AI is good at that. When Wharton MBA students were pitted against AI to generate innovative ideas, AI won. One person’s hallucination is another person’s innovative new idea.
Willful Blindness and Naïveté vs. Creative Problem Solving
Do not assume the opposing view is correct. It would be a mistake to look at the question of whether AI can handle the truth from the completely opposite perspective. Generative AI can indeed be foolish. It can be used poorly by users who are poorly-educated on generative AI or inattentive. This misuse can cause real damage. But these users share a common issue with some commentators. They rely on generative AI to create reality, certainty, and facts. However, their inputs are not grounded in reality, certainty, and facts. Ask AI to dream of a new legal future, and it will dream with you.
When the law firm Levidow and Oberman made headlines in June 2023, they submitted a brief with made up case cites. Many people looked at their actions. They did not consider their intentions. They wanted to find a creative solution to a highly specific legal problem. Their client was making a rather difficult personal injury claim. This claim came after the statute had tolled. The case law wasn’t there. They needed a new interpretation of an international convention with bankruptcy law. ChatGPT showed what the perfect case would look like.
Some commentators rightly observed that this was a case of willful naïveté regarding the technology. This occurred in the perfectly reasonable pursuit of a creative legal solution. An article in Above The Law addresses the brief. It states: “Perhaps finding a whole page of directly quotable support for a hyperspecific legal question should’ve tipped someone off.” In this case, it did not tip the lawyers off, but we can identify with the lawyers’ likely elation. Many lawyers can picture the excitement the brief’s writers felt. They experienced joy in finding a case perfectly on-point for their legal issue. When the cases are not fake, it is thrilling to find other lawyers who share your creative reasoning. It is also exciting to see them make similar connections.
Lawyers frequently use outside interpretations and logic. They often rely on these to make creative new arguments and inferences. Much like arguing from law review articles, unpublished opinions, or international law, these new arguments are not based on “truth.” They do not live in reality and are not based on certainty. Instead, they extend from a lawyer’s fact-based logic and reasoning. A natural language model can construct these extensions probabilistically. I can ask a generative AI model to dream of a new legal future. It has the capability to take me there. Before AI we called this: creative problem-solving and legal reasoning.
AI’s real value in the legal field may lie in its ability to aid in creative problem-solving rather than in uncovering objective truths, emphasizing the tool’s role in generating novel legal interpretations.
Seeking Truth: Lawyers as Finders of Fact
Lawyers are not mathematicians, and legal problems are not mathematical proofs. Professor David Luban wrote in a 1983 article: “A trial is not a quiz show with the right answer waiting in a sealed envelope. We can’t learn directly whether the facts are really as the trier determined them because we don’t ever find out the facts.” Our system rests on a foundation of truth. This can only be accurate if we know what the truth is. We do not.
There is no such thing as objective truth in the law. We do not need to cite examples from the greatest and most ill-conceived cases of our nation’s long history. These examples would just show how accurate this statement is. Laws age, and science advances. Our understanding of the limitations of human beings improves. How we view the world and our place in it changes over time. So, segregation ends. Fingerprinting and bullet identification become open to question. Eyewitness testimony comes down from its pedestal. International humanitarian law becomes a focal point of modern war. Generative AI will just be another step in the process for seeking understanding from the data available.
Perhaps, what lawyers and commentators need is better education on AI tools. They should understand what these tools can do before they ask if AI can handle the truth. However, this process of educating users to be most effective faces challenges. Regressive articles attempt to shoehorn moralistic principles and ill-defined terms into a topic where they don’t belong. It is not the job of AI to be more “truthful” when it comes to legal problems. AI is not meant to be a better finder of fact. That responsibility has, and always will, rest with the lawyer/user.
My point is not to question reality, or denigrate the importance of seeking truth and facts. Sometimes the traffic light was either red or green. Instead, my purpose is to make it clear that you cannot blame a system for tasks outside its design. When I ask a generative AI for a fact, my question deserves scrutiny first. After that, the answer from the model should be examined.
The justice system, designed around adversarial principles and client interests, inherently struggles with the notion of producing ‘truth,’ a challenge that extends to the use of AI in legal contexts.
Conclusion: Lawyers are Not Prompt Engineers
Our expectations and understanding need to change before we ask if AI can handle the truth. A prompt engineer should be smart enough to ask the earlier banana question differently. They need to craft a question that will lead to a consistent answer. But what about creative reasoning questions? When Pablo Picasso was asked to draw a face in a famous 1956 video, the result surprised everyone. I will not spoil the surprise. There, the artistic user took the question and developed a creative solution independent of our interpretation. What should we expect, he is an artist. I would similarly argue there is an art to the law. Our understanding of what a lawyer does needs to change first. Only then can we discuss the efficacy of the lawyer’s tools.
A lawyer is after the truth needed for their client, constrained within the rules of professional responsibility for their given jurisdiction. Existing law is not a barrier when justice for a client is in the balance. A good lawyer will work tirelessly. They explore how the law should change. They adjust the law for the facts of the client. The result is a problem in need of a creative solution, and a lot of questions from a lawyer. The questions these lawyers are asking would make an AI prompt engineer cringe. Lawyers are not trained in prompt engineering in law school. For their clients, they exercise the creative parts of an AI model. This approach is instead of using the fact-based and truth-seeking parts.
Lawyers engaged in creative problem-solving are going to get creative answers. However, AI can still be relied upon to generate many facts. When asking AI to strive for certainty, the questions need to be well-formulated. They must be free from room to roam. Asking a generative AI model if it understands the truth is a poorly formulated question. It provides too much room to roam. This is a mistake the Fast Company article makes and relies upon. Try thinking like a prompt engineer and ask instead: “what is the consensus definition of the word truth?” and see just how far some of us need to travel on our journey with artificial intelligence.
Leave a Reply