Listening to the words

It’s hard to tell who is at real risk for suicide.

“The problem is that if you ask people if they plan on killing themselves, they probably will say no, even if they do want to kill themselves,” Yosef Sokol said.

“People who really want to die don’t want you to stop them. They also don’t want to be hospitalized. They don’t want the stigma that comes with people thinking that there’s something wrong with them because they want to die.

“On the other hand, people who do talk about suicide often do not plan to kill themselves. Being hospitalized involuntarily for suicidal ideation is remarkably unpleasant, and therapists would like to spare patients who do not need it from that experience.”

But how can you tell who’s who?

Rabbi Dr. Sokol, who lives in Passaic, is a clinical psychologist with formidable credentials. He’s an assistant professor at the school of health sciences at Touro University and a researcher at the Veterans Administration Medical Center in the Bronx, where he not only works on his own projects but oversees others. And he’s a rabbi — his smicha is from Yeshiva Bais Yosef — who is deeply informed by Jewish wisdom. He does not talk about that part of his background in his clinical work, which is based on science and clinical research, but it infuses his worldview, as his book, “Becoming Godly: Integrating Torah and Psychology to Guide Us in Emulating HaShem,” makes clear.

So. To go back to the problem, how do you differentiate people whose suicidal ideation may well lead to suicide from those who are drawn to the idea but would stop short of attempting it, and therefore would benefit from a different form of treatment?

To be concrete, people call suicide helplines for a variety of reasons. “Sometimes you’re looking for support, you’re looking to vent, you’re looking to tell somebody you’re not doing well, and you don’t always want to be hospitalized,” he said. “But you tell the person who answers the hotline” — who “are there because they want to help people, and I respect them greatly for doing that, but they often are not very well trained”  — that “‘I have a plan, and I plan on carrying it out within the next week or the next day,’ then they may have to elevate it, escalate it, call an ambulance to take them to the hospital.”

People often call helplines “when they’re deeply, deeply ambivalent. They know they want some kind of help, but they don’t know what kind. They know they don’t want the kind that’s dramatic and stigmatizing.

“We need better methods of assessing risk.”

The generally used standard questionnaire, called Suicide Behavior Questionnaire Revised, developed in 2001, was broadly useful but easy to game, Dr. Sokol said. So he developed a set of more probing questions, another scale called Future Self-Continuity in 2020; those questions have become the standard in the field.

But “I made the standard, so I know that it is flawed,” he said. “It asks you to rate yourself from one to six about how vividly you can see yourself in the future.” The questions are good, but “still you are asking people to rate themselves from one to six. How deep can that get?”

But now he’s going beyond that by developing a list of 15 questions (and subquestions) using a large language model to analyze the language that people use as they answer them. The LLM can analyze the answers more thoroughly in far less time than a therapist can.

The questions that Dr. Sokol uses to look for suicidal ideation all have to do with how people can imagine themselves in the future — in five years, in 10 years. “If people feel that they are not connected to their future self, that’s a strong risk factor,” Dr. Sokol said. “People will want to end their lives because they have no connection to who they’re going to be. So they feel that they might as well end it.”

“Tell us how you are similar and how you differ from your future self,” the first question asks. “Describe your future self as vividly as you can.”

Subsequent questions ask the respondent to imagine a typical day in the future, and to describe it in detail. Questions ask about friendships, familial relationships, and communities. They ask about respondents’ thoughts on how their personality might grow or change, about goals and accomplishments, about hobbies, beliefs, and values.

These questions are not generic. They are highly specific. They are about “future self-continuity,” Dr. Sokol said. You — the person who answers these questions — will continue to be you, but “your life will be very different. You’re going to be very different. So are you still you? Do you still feel that is you?”

These questions are fascinating for anyone to contemplate, but they have particular relevance as a diagnostic for people who might be considering suicide.

Next, he uses AI — Claude 3.5 Sonnet. That way, “I can have you talk about your future and then examine the language you use. Is it specific? Is it concrete? Or is it vague, nebulous, generic terms? “So that’s what I had the LLM do. I created a rubric that’s very, very specific and detailed.

“A human could do this. I’m a therapist, and I’m pretty well trained on this. I could sit down with a transcript, and it would probably take me at least 20 minutes per transcript. Claude could do it in a second or less, and it’s much more reliable.” That’s at least in part because it’s entirely objective, something that a human being by definition cannot be.

Dr. Sokol’s paper about the LLM has just been published in the Journal of Personality Assessment.

He’s now working on “my most novel paper that I’ve ever written,” he said. It’s about “a whole new way of looking at language, in terms of suicide or just anything.”

To oversimplify greatly, this paper examines how one word is likely to follow another word, and how that combination of words is likely, according to the LLM, given the enormous, beyond-human-capacity amount of data the LLM has, to result in a predictable outcome, including suicide. It’s about how stories — which are lives — are likely to unfold.

“We’re never going to get that good at it, because people are way too complicated,” Dr. Sokol said. “We’re going to get just good enough at it so that we can use it well.”

That idea can have scary theoretical, almost science-fiction-like implications — do you lock up someone who is likely to commit a crime preemptively, before the crime is committed?

But that’s getting far ahead of where we are now. For now, the LLM is likely to be able to help predict suicidal ideation, or its probable absence, and that will help therapists treat patients better.

Dr. Sokol will continue to experiment with the LLM, to treat patients and train other therapists, and he will continue to think and write because it is wildly exciting, he said.

And he will continue to draw on his life as an observant Jew in his work on “how to use the idea of psychology and Torah to understand what it means to be a better person,” because everyone’s life is a combination of everything they know, have lived through, and imagine for themselves and the world around them.