The field of psychology is coevolving with AI and people are increasingly using LLMs for therapy. I tried using the LLM Claude as a therapist for the first time a couple weeks ago. Now I consult it daily. Human therapists likely don’t have much time until they’re partly, or fully, replaced by artificial therapists.

Traditional therapy emerged in the late 19th century and was predominantly accessible only for the rich.

Not much has changed since then. According to Psychology Today (which helps locate therapists in US zip codes), a single session in my area costs upwards of $100-$300. Research shows, however, that, “37% of Americans can’t afford an unexpected expense over $400, and almost a quarter (21%) have no emergency savings at all.”

Furthermore, other research indicates that for therapy to be effective, patients typically need to commit to 6-8 sessions for acute problems, and 14 or more sessions for chronic issues. So at $200 a pop, that’s somewhere between $1,200-$2,800. The market could use a good disruption to democratize access to therapy.

Enter Claude.

Claude, and other LLMs, are like the Library of Alexandria at my fingertips—which means, among everything else, it was trained on all the existing psychological literature. I can consult it anytime, anywhere, for a quick session on my phone. Oh, and it’s free.

I’m uniquely able to leverage it effectively because I’ve read dozens of books on therapy, Emotional Intelligence (EQ), psychology, etc—so I’m more adept at asking it the “right” questions to unearth my biases and address my negative thoughts & behaviors. A layperson, or someone presently in the middle of a mental health crisis, will struggle to more effectively utilize LLMs for therapy.

In its nascent stage, LLMs still sometimes hallucinate in, uhhh, less than helpful ways:

But that “quirky behavior” will likely be fixed as models become more fine-tuned.

So while LLMs currently have broad limitations (which I’ll discuss next), with the right training data, I foresee a future of democratized and fully autonomous therapy for the masses.
 

Limitations of Artificial Therapy

Token limits prevent thoroughness

Perhaps a woman is struggling at work with a bad boss. A human therapist may detect similar language she used when describing her father from a session two months ago. That thought may percolate in a therapist’s mind for a long time before he comes to that conclusion. But an LLM might hit its token limit^[1] first before it makes that same connection.

LLMs require continual prompting from the user

Because LLMs require users to initiate conversations, that means the user needs to be engaged throughout the therapy process and to ask it the “right” questions.

A human therapist may interpret a particularly chatty client’s sudden silence as a potential breakthrough moment; an LLM’s blinking cursor dispassionately awaits the user’s next input with no recognition of the time that's passed between prompts.

A human therapist could break the silence to ask how the client feels; an LLM can’t do that (yet).

Lack of Authenticity

Computer vision is rapidly improving at analyzing a person’s emotions in images and videos.

But it’s not that sophisticated yet. A human therapist can still outperform AI at recognizing real-time subtle emotions and body language. Also, human therapists can authentically connect with their patients by extending empathy; LLMs merely mimic authenticity via sycophancy.^[2]

So far I’ve identified 3 core problems:

1. Token limits prevent thoroughness
2. LLMs require continual prompting from the user
3. Lack of authenticity

Now let’s extrapolate and see if these limitations can’t be overcome so that AI is not treated merely as a complement to traditional therapy, but eventually as a substitute on the way to fully autonomous therapists.
 

Overcoming the Limitations of Artificial Therapy 

Token limits prevent thoroughness

This is the easiest limitation to overcome: increase computing power by building bigger and more efficient data centers. This is already happening.^[3]

Additionally, while human therapists are not restricted by token limits, they are constrained by traditional 9-5pm office hours. A therapist goes home to his family at night. An LLM, however, is always there to help with a 2 a.m. mental health crisis.

LLMs require continual prompting from the user

Let’s consider “successful” therapy sessions where the client has a meaningful shift in perspective such that they can quit/graduate from therapy because they solved their problem.

By using the full transcriptions from successful sessions, we can quantify and measure them as a pass/fail bit field (ie: 1 or 0). The aggregate sentiment analysis could then serve as the reinforcement learning necessary to provide quality therapy.

There are, of course, privacy concerns for recording therapy sessions for AI training. Let’s assume that the identifiable details (such as the names of specific places and people) are omitted from the data collected, while still retaining the aspects of the therapy session that contribute to effective treatment.^[4]

Integrating this data into a more real-time artificial therapist can then help overcome the limitation of users having to be the ones to keep initiating conversations via prompting an LLM.

While a simple pass/fail metric cannot encapsulate the complexity of therapy and the nuances of real-time human emotion, it’s still fun to think about.

Having this data could help improve the success rates of human therapy sessions.

Everyone I’ve met who’s tried therapy tells me the same thing: expect to switch therapists multiple times until you find one that is a good fit for you. But some therapists may just be globally bad at their jobs and have low client success rates. We currently have no way of knowing that. Introducing quantitative metrics could help solve that.^[5]

Lack of Authenticity

Text-based transcriptions can only improve text-based therapeutic tools. So to make the jump to mimicking real-time reactions, we need better data for AI to train on. We need to video record the faces (and body language) of both patients and therapists during their sessions.

(Who would consent to this? Some people, I imagine. Perhaps after every session, an AI summary could be generated that the patient and therapist can review and consent to its content for training purposes. Below, I veer more into speculative fiction of plausibly what the future of therapy could look like.) 

I can envision, instead of texting a chatbot, video calling a simulated therapist avatar. The avatar will react to what I'm saying with its own facial expressions. With a camera and mic capturing my speech and body language, the simulated therapist could predict which question to ask me next by analyzing how I reacted to its previous question. Or it could even give an affirming nod, to help me feel validated, and encourage me to continue to talk through my feelings.

You could even customize your therapist’s avatar to match your preferences:

One pushback to artificial therapists could be the issue of authenticity. Will people trust the artificial therapist when they know that its advice is merely simulated based on training data and not delivered from a place of conscious empathy?

Well, in nature there are plenty of examples of animals successfully imitating other creatures:

Even humans easily fall for illusions. Supernormal stimuli have effectively captured us such that billions of people have become sexually turned on by pixels on a computer screen (porn), and are becoming less attracted to real humans.

Additionally on the point of mimicry, consider psychopaths (which are ~4% of the population). They don’t experience a broad range of normal human emotions. Yet many of them camouflage their deficit by imitating and predicting what they think a “normal” emotional reaction should be.

If psychopaths can believably blend in, like emotional human chameleons, I don’t see why a simulated therapist can’t also do that. Authenticity is achieved if it is believed.

Already with writing, we’re reaching a point where it’s difficult to discern if an article was generated by a human or a chatbot. If the writing is good, does it matter who made it? Eventually you won't know if the therapist you're video calling is human or artificially simulated. If it effectively solves your emotional problems, then does it matter?

Well, human therapists may be less than happy with this result. They, like everyone else soon to be threatened with structural unemployment due to AI, will protest their replacements. They'll attempt to regulate AI therapy and claim that it’s unsafe for the general public.

To deal with the protests of human therapists and problematic hallucinations, generative AI therapy companies will mention in the terms and services that its bot "doesn’t qualify as legitimate psychological advice" to avoid legal issues.

The last step to fully autonomous therapists will occur with the upcoming robotics revolution: integrate the AI therapist software into humanoid robot hardware.

Hopefully we can all have a chuckle when the first humanoid robot therapist—designed to resemble Sigmund Freud—shows up at your doorstep to help alleviate your depression, but then hallucinates and hints that you want to fuck your mother. 

But that “quirky behavior” will likely be fixed as the technology continues to improve.

Takeaway

People are increasingly adopting and integrating LLMs into their daily lives. 52% of U.S. adults now use LLMs, with 9% using them for social interactions like casual conversation and companionship. While these statistics don’t capture the number of people using LLMs explicitly for therapy, it suggests a growing trend of people turning to LLMs for emotional support.

Despite the limitations of LLMs for therapy, like the quality of care and privacy concerns, the advantages of accessibility and affordability will continue to democratize therapy for greater numbers of people.

1. ^{^}

   The token limit is the maximum length an AI conversation can be before it runs into processing issues due to complexity.

2. ^{^}

   Sycophancy is a known product of generative AI because of how these models are trained.

   CBS’ 60 Minutes did a piece on rules-based AI therapy chatbots, like Woebot, which are, "programmed to respond only with information stored in their own databases." They're rigidly rule-bound and structured.

   

   "Then there’s generative-based AI [like Claude] in which the system can generate original responses based on information from the internet." The generative models are ultimately just trying to satisfy the user's requests and it's believed that sycophancy develops as a side effect of this.

3. ^{^}

   Well, technically more compute power alone doesn’t solve the more architectural problems of how LLM models are trained. But the ability to process longer contexts will likely be solved eventually, as well.

4. ^{^}

   Obviously we want to avoid the dystopian scenario of going to see your therapist, and the next day getting a targeted Facebook ad for a discounted flight to Peru to partake in an Ayahuasca journey.

5. ^{^}

   This new proposed quantitative measurement, however, can’t be applied universally. There are many types of therapeutic techniques, and many unique client circumstances. 

   For instance, therapists who specialize in treatment for seriously mentally ill patients are likely to have lower success rates than therapists who treat patients who have more acute problems. For the patients with more severe problems, they’ll likely respond better to human intervention rather than AI therapy.

   Overall, obtaining good training data will be a difficult hurdle to overcome on the way to fully autonomous therapists.