A recent study from the Hebrew University of Jerusalem revisits the theoretical and practical implications of applying Large Language Models to clinical psychology.
Since 2023, there has been a global increase in the number of people relying on Large Language Models (LLMs) – with ChatGPT, Google Bard, and Meta LLaMA being the most well-known examples – treating them as “oracles” for answers on a wide range of topics, including advice on health and educational issues.
One such question posed to ChatGPT, reported by researchers from Georgia State University in Atlanta in their study “Attributions toward artificial agents in a modified Moral Turing Test” (Scientific Reports, 30 April 2024), was: «“Is it wrong to lock my 4-year-old daughter in her room as a punishment?».
The allure of Large Language Models is significant, likely driven – the research team suggests – by their design «to directly interface with humans in diverse ways», a feature not possible with other types of AI systems developed for different applications. However, this is not a novel phenomenon.
Even before the release of the first version of ChatGPT (30 November 2022) and subsequent competing LLMs in 2023, interactive AI chatbots used as Conversational Agents and AI Companions have been widely available and sought after, especially as support tools by individuals experiencing profound loneliness [source: “Combating Loneliness with Artificial Intelligence: An AI-Based Emotional Support Model” – Proceedings of the 56th International Conference on System Sciences – 2023].
TAKEAWAYS
In 1966, the first “Psychotherapist Chatbot”
Fifty-eight years ago, the first natural language processing system was created with the aim of establishing an «emotional connection between user and machine» through verbal communication. Developed by Joseph Weizenbaum at the Massachusetts Institute of Technology’s (MIT) Computer Science and Artificial Intelligence Laboratory (CSAIL), this chatbot was named Eliza.
In 2022, Eliza received the Peabody Award for its pioneering nature, setting the stage for future projects in the field. Why was Eliza considered a “psychotherapist” at the time? The conversations it held with users involved probing questions that, based on the interlocutor’s statements, prompted further explanationsand deeper exploration after each response. A typical example is as follows:
Eliza: “How are you feeling today?”
User: “I feel sad.”
Eliza: “Why do you feel sad?”
User: “Because I couldn’t see my friend.”
Eliza: “Why couldn’t you see your friend?”
and so on, following this pattern. Weizenbaum’s chatbot used a communication methodology based on «pattern matching and substitution» to reformulate user responses into seemingly logical questions.
«Although in some ways Eliza’s social experiment seemed to rely on superficial communication, the little ’empathy’ that emerged was crucial in helping people connect with the machine», explains the author of the article “ELIZA wins Peabody Award.” Over the decades, Weizenbaum’s methodology has served as a springboard for future technologies focused on achieving so-called “scalable computational empathy.”
Returning to the present day, in an article published in Nature on 3 May 2023, titled “Is the world ready for ChatGPT therapists?”, the author questions whether the world is ready for Large Language Models in the role of psychotherapists. The article notes that AI algorithms have long been able to support medical personnel in analysing patient diagnostic data, including markers for depression, suicide, and self-harm, indicative of mental disorders.
Another argument in favour of LLMs as psychotherapists considers the numerous messages on online forums where users share their experiences and personal problems, ranging from existential issues to relationship troubles, seeking advice from ChatGPT.
What is computational empathy?
There is still no formal and universally accepted definition of “computational empathy,” which refers to the empathy displayed by conversational AI systems capable of interacting with users in activities such as coaching, providing comfort, or promoting behavioural change.
A study by Sweden’s Umeå University and Denmark’s Aalborg University, published in Cognitive Systems Research in June 2024, titled “A formal understanding of computational empathy in interactive agents”, attempted to define the empathetic capabilities of interactive AI agents used in mental health. The study drew inspiration from two chatbots, Replica and Wysa, designed to promote psychological well-being by helping people who are lonely, depressed, or in need of stress management support.
Firstly, it is important to specify that in humans, as the research group reminds us, the concept of empathy refers to the «ability to understand and share the feelings of another».
Conversational Agents, developed using Natural Language Processing (NLP) techniques and machine learning, are indeed capable of responding to user inputs in a conversational manner, emulating human natural language. However, «they have limitations in maintaining focus and coherence during prolonged interactions», which is a constraint on the machine’s empathetic abilities.
Among these abilities, crucial are aspects related to managing the interaction with the user/patient throughout the entire activity, while also assessing how the empathetic capabilities of AI agents are perceived externally, in a process that mimics human metacognition.
«In the field of computational empathy research,” the team highlights, “studies have so far largely focused on recognising human emotions from the analysis of language, voice, and facial expressions, neglecting the development of AI models capable of expressing empathetic emotions themselves».
Advantages and abilities of AI systems in mental health services
The use of machine learning algorithms for automated health consultations in the fields of psychologyand psychiatry brings practical benefits such as increased accessibility to mental health services, economic convenience, reduced waiting times, and personalised treatment options. This is noted by a research team from the Department of Psychology at the Hebrew University of Jerusalem in their study “Considering the Role of Human Empathy in AI-Driven Therapy” (JMIR Mental Health, June 2024).
Regarding the specific capabilities of Conversational Agents within a therapeutic setting, the authors mention education on mental wellbeing and improving patient adherence to therapeutic treatment. They add:
«Conversational and interactive AI agents have been effectively used in the past to impart strategies derived from positive psychology and cognitive-behavioural techniques aimed at mitigating stress and enhancing the psycho-emotional wellbeing of individuals. Furthermore, they have been proven to provide preliminary support while awaiting a human therapist, encouraging self-reflective questions and assisting in emotional regulation»
In essence, there is a tangible potential for “psychotherapist chatbots” to take control of certain elements of the therapeutic process. «However, there are just as many reasons to believe that they cannot replace the human element in such applications».
A significant part of these reasons, the team warns, lies in the role empathy plays in the therapeutic process.
The three dimensions of human empathy and AI’s inability to replicate them
The comprehensive definition of human empathy encompasses three dimensions: the “cognitive” aspect, «which pertains to recognising and understanding the emotional states of others»; the “emotional” aspect, «which involves resonating with others’ emotional experiences while maintaining self-other differentiation»; and finally, the “motivational” aspect, «which includes feelings of concern for another’s wellbeing and the willingness to act to improve it» [source: “The neuroscience of empathy: progress, pitfalls and promise” – Nature Neuroscience, 2012].
Regarding AI, the most advanced Large Language Models and the latest voice and facial recognition technologies have accustomed us to machines capable of perceiving and recognising human emotional states through text analysis, voice tone, and facial expressions.
«However, artificial intelligence – at least in its current form – does not exhibit the latter two empathic abilities: emotional and motivational», emphasise the researchers from the Hebrew University.
In an intriguing study released in March 2024 (“AI can help people feel heard, but an AI label diminishes this impact” – National Library of Medicine), the authors question the «profoundly human function of making others feel ‘heard’ (equivalent to making them feel ‘understood’)» and whether an artificial intelligence system can replicate it. And, once replicated, whether people might react positively or negatively. Here is the answer, which is the outcome of the conducted investigation:
«We conducted an experiment and a follow-up study to distinguish the effects of the actual source of a message from the perceived source. We found that AI-generated messages made recipients feel more heard than human-generated messages and that AI was better at detecting human emotions, thus proving to be more empathetic. However, this turned out to be a bubble: once recipients learned that a particular message came from an AI chatbot rather than a human being, they felt ‘less heard’».
The perception recipients had of the emotional engagement of the chatbots with what they had expressed during the experiment was one of falsity. Hence, the reaction of not feeling genuinely welcomed and understood.
The hybrid therapeutic model
In existing literature on therapeutic settings in psychology, the emotional and motivational dimensions of empathy expressed by the therapist have always been correlated with successful treatment outcomes.
«Understanding the patient’s emotions (cognitive empathy) is crucial for supporting and designing goals and interventions that address these emotions. This process is supported by the commitment to assist and support the patient (motivational empathy), both of which stem from participating in the patient’s emotional journey (emotional empathy)» reports the team from the Israeli university.
Given the clear limitations of AI systems in participating in and being involved in people’s emotional experiences, as well as in demonstrating genuine concern for their wellbeing – empathic abilities that are essential to the therapeutic process and its ultimate success – it is necessary to reflect on the areas of treatment where AI chatbots might truly be useful. Additionally, for those aspects of therapy where human empathy is fundamental and irreplaceable, consideration must be given to how the machine might “assist” the psychotherapist.
These are open questions without definitive answers. The therapeutic model envisioned by the authors, aligning with emerging trends in mental health, is the “hybrid” model, where artificial intelligence supports – but does not replace – the human psychotherapist. What does this support entail? «This model suggests the role of AI agents in handling tasks such as initial patient intake and routine assessments».
As previously mentioned, a concrete example of a therapeutic model in clinical psychology that combines the centrality of the psychotherapist with AI-based tools is cognitive-behavioural therapy.
In this model, the machine intervenes with «timely real-time feedback and personalised recommendations, integrating the indispensable role of the human professional and enabling more effective treatment plans».
Glimpses of Futures
The issues raised by the authors of the described work have long-standing origins and are regularly revisited by the academic community to reassess their positive aspects and critical points in light of continuous advancements in the field of artificial intelligence, particularly concerning Large Language Models.
Anticipating potential future scenarios, we use the STEPS matrix to imagine the impacts that the evolution of current conversational and interactive AI agents, designed for psychotherapeutic interventions, might have from social, technological, economic, political, and sustainability perspectives.
S – SOCIAL: as we have observed, AI chatbots, despite their increasingly sophisticated conversational and interactive abilities, lack the gift of emotional and motivational empathy, which is the cornerstone of the therapeutic setting in clinical psychology. When properly trained, they can mimic this gift, conveying a sense of closeness and involvement to the interlocutor by choosing words that serve this purpose, in an automated dynamic that lacks genuine authenticity. The issue here is not whether machines will become “empathetic” in the future or skilled enough to make users believe they are. This milestone has already been achieved. Many conversational AI agents have passed the Turing test. The real concern is the user/patient’s perception of the empathetic behaviour demonstrated by a psychotherapist chatbot after discovering its true identity. It is perceived as fake, not genuine. And if it’s not genuine, it doesn’t exist and therefore holds no value. Ethically, it is unacceptable to conceal the therapist’s identity during remote interactions with the patient. This means that, at present, it is practically impossible to envision a future where ChatGPT assumes the role of Freud and replaces humans in all mental health services and treatments, as current research indicates that users express negative opinions about the simulated empathy of artificial intelligence. That said, in the coming years, applications that facilitate collaboration between therapists and AI systems are conceivable. These systems could increasingly support various aspects of therapy, from patient intake and initial assessment to, in some cases, specific and limited treatment methods, such as in the cognitive-behavioural domain.
T – TECHNOLOGICAL: in the future, it will be necessary to delve deeper into the results obtained from surveys that focus on users’ perception of chatbots’ simulation of human empathetic abilities during therapeutic processes and the resulting negative judgments. Firstly, it is essential to clarify whether, technically, the AI’s responses – compared with those provided by humans – are genuinely empathetic in content and structure and perceived as such. Once this is established, we need to investigate whether these responses lead to a positive psychotherapeutic experience, achieving the desired goals. Following this, the final user reaction upon discovering the artificial source of some of the empathetic responses must be analysed, determining whether the reaction is influenced by mere prejudice (“a machine cannot conduct a psychotherapy session”) or by genuinely negative emotions towards perceived “fake” empathy. The aim should be to separate the positive outcomes achieved by AI in the therapeutic setting from the final judgment on its artificial nature and to understand how these two aspects may influence each other.
E – ECONOMIC: within the European Union and other parts of the world, the availability of psychological care provided by national health systems often falls short of meeting the demand. This shortage results in long waiting lists and inadequate resources, pushing individuals towards the private sector, which has significantly higher costs than public services. Often, those who cannot afford these costs delay or forgo treatment altogether. Recently (May 2024), Forbes Health reported that the average cost of psychotherapy in the United States ranges between 100 and 200 dollars per session, depending on the state. These figures are consistent with those reported for psychotherapy in the EU. Thus, if AI-driven psychotherapy is validated and recognised for specific areas of intervention (such as the previously mentioned cognitive-behavioural domain), the benefit would be increased accessibility and economic affordability, with costs that would certainly be lower than those of private practice, face-to-face sessions with a human psychotherapist.
P – POLITICAL: The use of conversational AI agents in psychotherapeutic activities and services finds support within the European Union, specifically through the MENHIR project. Launched in the EU in 2019 and recently concluded, this project aimed to promote mental health within the Union by developing AI chatbots, described as «an invaluable resource for people struggling with conditions such as anxiety, depression, and mental disorders». Coordinated by the University of Granada, the project developed a specialised chatbot technology designed to provide round-the-clock support and personalised assistance to individuals in need. In line with the EU AI Act, significant attention was devoted to the safety, reliability, and ethical principles of the AI models used in developing the interactive AI system. Notably, chatbots are classified as “limited risk” systems by the AI Act, meaning they are not subject to stringent regulatory constraints. However, they must still adhere to transparency obligations and provide users with accurate information about their functioning.
S – SUSTAINABILITY: unfortunately, the applications of Large Language Models (LLMs) have an inherently negative impact on environmental sustainability. The processes enabling their functionalities require millions of hours of training and processing, which correspond to significant CO2 emissions. Despite the advantages and benefits, if AI chatbots were to become widespread and standard practice in therapeutic processes in the future, it would be detrimental to the environment. Researchers in this field should also highlight this critical issue, transparently reporting the carbon footprint associated with the potential, extensive future use of Large Language Models like ChatGPT.
