Leveraging generative AI and LLMs to make predictions about the path and outcome of undertaking mental health therapy.
getty
In today’s column, I examine the fascinating possibility that we might be able to use generative AI and large language models (LLMs) to predict the psychotherapeutic success or failure for people opting to undertake mental health therapy.
Here’s the deal. A person decides to get mental health support and makes use of therapy accordingly. A vexing question that has existed since the beginning of psychotherapy is whether the eventual outcome can be predicted. The person might successfully complete therapy and be considered good to go, no longer requiring additional therapy. On the other hand, a person might end up still in mental straits, despite the therapy going on and on.
Research in the field of psychology suggests that predicting the likely outcome is often possible. The hope is to make the prediction as soon as possible, namely, perhaps after the first or second session of therapy. This will then allow the therapist and the client to identify jointly what else might be undertaken if the treatment is seemingly not going to produce the anticipated upside results.
Various traditional statistical approaches have been previously used for this complex prediction task. The latest avenue consists of stridently leaning into the use of generative AI and LLMs. Using modern AI for this vaunted purpose is an intriguing possibility and is only now being avidly explored.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
MORE FOR YOU
As a quick background, I’ve been extensively covering and analyzing a myriad of facets regarding the advent of modern-era AI that produces mental health advice and performs AI-driven therapy. This rising use of AI has principally been spurred by the evolving advances and widespread adoption of generative AI. For a quick summary of some of my posted columns on this evolving topic, see the link here, which briefly recaps about forty of the over one hundred column postings that I’ve made on the subject.
There is little doubt that this is a rapidly developing field and that there are tremendous upsides to be had, but at the same time, regrettably, hidden risks and outright gotchas come into these endeavors, too. I frequently speak up about these pressing matters, including in an appearance last year on an episode of CBS’s 60 Minutes, see the link here.
I’d like to set the stage on how generative AI and large language models (LLMs) are typically used in an ad hoc way for mental health guidance. Millions upon millions of people are using generative AI as their ongoing advisor on mental health considerations (note that ChatGPT alone has over 900 million weekly active users, a notable proportion of which dip into mental health aspects, see my analysis at the link here). The top-ranked use of contemporary generative AI and LLMs is to consult with the AI on mental health facets; see my coverage at the link here.
This popular usage makes abundant sense. You can access most of the major generative AI systems for nearly free or at a super low cost, doing so anywhere and at any time. Thus, if you have any mental health qualms that you want to chat about, all you need to do is log in to AI and proceed forthwith on a 24/7 basis.
There are significant worries that AI can readily go off the rails or otherwise dispense unsuitable or even egregiously inappropriate mental health advice. Banner headlines in August of this year accompanied the lawsuit filed against OpenAI for their lack of AI safeguards when it came to providing cognitive advisement.
Despite claims by AI makers that they are gradually instituting AI safeguards, there are still a lot of downside risks of the AI doing untoward acts, such as insidiously helping users in co-creating delusions that can lead to self-harm. For my follow-on analysis of details about the OpenAI lawsuit and how AI can foster delusional thinking in humans, see my analysis at the link here. As noted, I have been earnestly predicting that eventually all of the major AI makers will be taken to the woodshed for their paucity of robust AI safeguards.
Today’s generic LLMs, such as ChatGPT, Claude, Gemini, Grok, and others, are not at all akin to the robust capabilities of human therapists. Meanwhile, specialized LLMs are being built to presumably attain similar qualities, but they are still primarily in the development and testing stages. See my coverage at the link here.
Let’s talk about mental health therapy all told.
When someone opts to formally undertake mental health therapy, there is a longstanding, unresolved question of whether the therapy will ultimately be a success or a failure. The aim would be to try and ascertain the likely outcome of treatment at the soonest possible opportunity. The sooner, the better. The complication is that you usually need the therapy to get underway before you can reliably make such a prediction.
Consider these stages at which a prediction might be made:
It would be especially noteworthy to be able to make a bona fide prediction before therapy gets underway, ascertaining how the therapy will likely proceed and conclude. That’s usually not very feasible.
Research suggests that at least one session is usually the most suitable trigger point at which a solid prediction can be made. Some research says it takes two sessions. After the third session, the viewpoint is that a reasonable prediction could happen at any time, but this also entails having the person already invested in several sessions. Again, earlier would be better, if possible.
In a research study entitled “Predicting Treatment Success And Failure Using Routine Outcome Data: The Role Of Therapist Effects In Dynamic Predictive Modelling” by Daryl Mahon, Takuya Minami, Jeb Brown, Counseling and Psychotherapy Research, September 24, 2024, these salient points were made (excerpts):
The study used a therapeutic-focused database of clients who had completed questionnaires detailing their self-reported outcomes. A sample of instances consisting of 1,020 therapists and 68,690 clients was utilized for the research analysis. Various variables were identified and codified. The statistical analysis made use of regression models.
Would it be possible to make use of generative AI and LLMs as a means of predicting psychotherapeutic outcomes and their corresponding trajectories?
The answer is yes.
A caveat is that these are early days for undertaking such an approach, and there isn’t enough research to yet categorically say whether AI can do a good job or a lousy job at making such predictions. In other words, sure, you can try using AI, but will it empirically stand up as a viable and reliable method?
Time and robust research will tell.
Let’s venture into the matter and see what can be undertaken.
The first place to begin the journey entails delineating the overall nature of the therapist and the overall nature of the client. We must lay out the characteristics of the therapist so that we can use that depiction in the prediction process. Likewise, we must lay out the characteristics of the client, too. Based on those descriptions, an LLM can be prompted to make a prediction about success or failure as therapeutic outcomes.
I’ve previously covered that it is possible to craft a so-called digital twin in AI; see my in-depth discussion at the link here. A digital twin is simply the aspect of specifying enough detail about a real human that the AI can pretend or simulate being that person. You invoke an AI persona to do so. It is a mimicry of a particular human.
If you ask AI to undertake a persona of someone for whom there was sparse data training at the setup stage, the persona is likely to be limited and unconvincing. You can augment the AI by providing additional data about the person, using an approach such as RAG (retrieval-augmented generation, see my discussion at the link here). Personas are quick and easy to invoke. You just tell the AI to pretend to be this or that person. If you want to invoke a specific person, you will need to specify sufficient characteristics so that the AI will get the drift of what you intend. For prompting strategies on invoking AI personas, see my recommended steps at the link here.
I am a steadfast advocate of using AI personas as a training capability for therapists and mental health professionals (see my coverage on this exemplary approach, at the link here). Things go like this. A budding therapist might not yet be comfortable dealing with someone who has delusions. The therapist could opt to “fake” practice on a real person pretending to have delusions, though this is likely costly and logistically complicated to arrange. Instead, they easily craft an AI persona that pretends to be delusional. Practicing and assessment, along with therapeutic improvement, can be readily achieved.
Conventionally, therapy consists of a human therapist and a human client. That’s the classic therapist-client pathway. To get AI into predicting the outcome of that therapist-client combination, we need to bring the respective therapist description and client description into the AI.
There are four primary variations with this approach. You can aim to mimic a real therapist and a real client. That requires two digital twins. Another option is to mimic a real therapist and have a fake client, namely, devising an AI persona “client” that is not based on a real human. An additional option encompasses a fake therapist, i.e., an AI persona not based on a real therapist, coupled with a digital twin of a real client. Finally, there is the possibility of using AI to undertake a fake therapist and a fake client.
Here are the four noted variations:
These leverage the AI personas to either consist of depicting a real person or a fake person, doing so either in the role of the therapist or the role of the client.
Let’s walk through an illustrative example.
I went ahead and logged into a popular LLM. This will work in any of the major AIs, such as OpenAI ChatGPT and GPT-5, Anthropic Claude, Google Gemini, xAI Grok, Meta Llama, and others.
I opted to get underway by detailing a therapist named Josh.
Next, I’ll do the same for the client named Sam.
The LLM now has a brief depiction of the therapist and the client. Whether this scant description is sufficient would be an important aspect of research analysis. How much of a description for a therapist is truly required? How much of a description of the client is truly required? Maybe a short snippet is enough. Or maybe volumes of details are needed.
I proceeded to tell the AI what happened during the first session of Josh and Sam. The idea is to see whether the LLM can make a prediction based on the results of merely one session.
Here we go.
I wanted to get the AI to provide more detail about the prediction, so I asked accordingly.
In this example, the outcome was predicted as being of a positive nature. The psychotherapy undertaken by Josh would be construed as a success.
Let’s see what the prediction might be in a tougher situation.
Here, I describe a new client named George.
The prediction by the LLM was this.
And I asked for additional details and an explanation.
The use of AI for predicting the outcome and trajectory of therapy is certainly promising. This could be pushed further by, for example, giving the AI a set of transcripts of whatever sessions have been undertaken. Rather than merely telling the AI what happened in the first session of Josh and Sam, we could input a transcript and let the AI use that as an additional basis for making its prediction.
Some are dismayed that AI is presumably making a prediction based on textual content versus using numeric content. In the case of the research study that I earlier cited, they used a database and numerous codified factors in a numerically oriented regression model. The LLM is not predicting via a similar numeric fashion. Thus, there are concerns that this is all hocus pocus and an LLM is just making up the whole kit and kaboodle.
Perhaps the two could be dovetailed together, making use of AI and a conventional statistical prediction method. That would seem a judicious avenue.
Plenty of worries abound about AI usage.
Suppose that we rely solely on an LLM to make these kinds of life-changing predictions. The AI predicts that Sam isn’t going to have a positive outcome by working with Josh. What does Josh think of that prediction? Could the AI be steered toward particular therapists and used as a commercial wink-wink of garnering clients?
Significant issues arise, such as the privacy and confidentiality associated with using AI in this manner. Also, there are potential clinical liability aspects and regulatory difficulties. Is the LLM essentially giving out clinical decision-making or medical advice by being used this way?
Another worry is that a classic self-fulfilling therapeutic prophecy can arise. The AI had predicted that Sam would require six sessions. If Josh tells Sam that they will likely be done by the sixth session, what might that do? The pacing of the therapy could be subconsciously artificially altered, with both the therapist and the client believing that by the sixth session, the world will be a different place. Expectations go awry. Predictions inadvertently become a kind of therapy script that must be blindly followed.
Using contemporary LLMs as a prediction tool for gauging the outcome and trajectory of psychotherapy is an intriguing possibility. It isn’t a cure-all. There are decidedly momentous upsides and downsides.
A final thought for now.
As per the remarkable words of Marcus Aurelius: “Nothing has such power to broaden the mind as the ability to investigate systematically and truly all that comes under thy observation in life.” Those fine insights apply to using LLMs in a psychotherapeutic predictive capacity. I will keep you posted on progress made in this innovative and emerging realm.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。