The 1972 Parry Chatbot: An Entry in AI and Mental Health

During the 1970s, around the time the artificial intelligence (AI) industry began to take root, the Parry chatbot was created, representing an important step in AI by impersonating a human suffering from psychotic disorders and mental health therapy. Parry was designed to replace the voice of a paranoid schizophrenic individual, engaging in inviting conversations with many users and practitioners curious about the possibility of divorcing psychotherapy from AI.

The Development of Parry: An Adventure Beyond Computation

Parry was developed in 1972 by psychiatrist Kenneth Colby, debuting after the creation of the 1966 ELIZA chatbot. Colby’s aim extended beyond making basic replies. Parry was designed to reflect the paranoia of a mentally ill individual. Parry operated on rules to respond emotionally, interpreting questions through suspicion and mistrust. Users were welcomed into a new experience with Parry, who had nearly five hundred heuristics governing her responses, offering a unique interaction that felt almost human.

Real World Reactions and Insights

One of PARRY’s most discussed experiments explored whether psychiatrists could distinguish its conversational responses from those of real patients.

In a study published in 1973, psychiatrist Kenneth Colby and his colleagues presented psychiatrists with transcripts of conversations involving PARRY and people diagnosed with paranoid schizophrenia. The psychiatrists were asked to identify which transcripts came from the computer and which came from human patients.

The results attracted considerable attention because the psychiatrists’ judgements were not consistently reliable. PARRY’s responses could resemble the conversational patterns associated with paranoid thinking closely enough to challenge professional assumptions about how such behaviour could be recognised through dialogue alone.

However, the experiment requires careful interpretation. It did not demonstrate that PARRY understood paranoia, possessed human emotions or could provide effective psychotherapy. It demonstrated something narrower, but equally significant: a computer program could reproduce recognisable patterns of human communication convincingly enough to complicate the task of distinguishing simulated behaviour from human behaviour.

This distinction remains important today. A system does not need consciousness, empathy or an understanding of subjective experience to produce responses that people perceive as emotionally meaningful.

PARRY therefore raised a question that extends beyond the capabilities of early computing: when behaviour appears convincing, how do we know what lies behind it?

Consequences of Parry for Modern AI and Psychotherapy

PARRY operated through carefully designed rules and conversational strategies. Contemporary generative AI systems work differently, using large-scale statistical models to generate responses based on patterns learned from extensive datasets. Their capabilities are far broader, but the fundamental question PARRY raised remains relevant: does convincing communication necessarily indicate genuine understanding?

Modern mental-health technologies bring this question into particularly sensitive territory.

The promise of personalised support.

AI systems can provide accessible conversational support, help users reflect on their experiences and, in appropriately designed applications, guide structured exercises. Their availability may make certain forms of support easier to access, although their effectiveness depends on the application, the intervention and the evidence supporting it.

The challenge of trust.

People may interpret fluent, empathetic language as evidence that a system understands their feelings or can reliably assess their circumstances. Yet a response that sounds compassionate is not, by itself, evidence of empathy, clinical competence or sound judgement. This gap between perceived understanding and demonstrated capability has implications for both users and the organisations deploying these systems.

The importance of evidence and safeguards.

Mental-health applications must be assessed on more than conversational quality. Relevant considerations include clinical outcomes, privacy, transparency, the handling of high-risk situations and access to appropriate human support. A promising research result is a reason to investigate further, not a guarantee that an application is safe or effective for every user.

The evolution from PARRY to generative AI represents a major technological advance, but it does not eliminate the need to distinguish appearance from capability. The more natural an interaction becomes, the more important it is to understand what the system can reliably do, where its limitations lie and when human judgement must take precedence.

These questions are no longer confined to mental-health research. They also matter wherever people use AI to communicate, make decisions or interpret human behaviour.

Vision for the Future of AI in Mental Health

Parry’s development marked a crucial moment in exploring AI’s role in understanding human emotion and psychology. This early experiment forged a path that contemporary AI applications continue to navigate and expand upon today. By comprehending the historical progression from Parry to modern AI solutions, we see a clear lineage of technological growth and refinement.

Widening Scope and Customization

The growth of AI technology holds enormous potential for broadening the reach of mental health resources. Making interactions more personal and developing finely tuned, appropriate responses can significantly improve user experiences and outcomes.

Innovations to Look For

New technologies such as deep learning and augmented reality have the potential to further integrate the mental health space. These innovations could provide a more engaging and interactive experience that enables users to relate better.

The Global Impact of Parry

Parry’s translational report brought a new perspective and a sense of hope for the use of AI in mental health. From the early attempts to create chatbots capable of human like conversation to today’s ever-evolving intelligent digital bots, AI has played an important role in the evolution of mental healthcare. As the field advances, integrating ethical approaches and actively considering empathy and collaboration will help ensure these technologies benefit society appropriately. Respecting past challenges and achievements can pave the way for a future mental health landscape that is both AI integrated and ethically satisfying.

Where things stand in 2026

The questions PARRY raised have returned in a more pressing form.

The scripted therapy bot. Woebot, a chatbot based on cognitive behavioural therapy, was launched in 2017. Its responses were pre-scripted, closer to ELIZA than to today’s generative systems. Its consumer app was retired on 30 June 2025. The company’s founder attributed the decision largely to the cost of meeting FDA requirements, and to the rise of language models it wanted to use but could not yet bring through regulatory review.

The first trial of a generative therapy bot. In March 2025, Dartmouth researchers published a randomised trial of Therabot, a generative AI chatbot, in NEJM AI. The trial involved 210 adults with depression, anxiety or high-risk eating-disorder symptoms, and the chatbot group showed statistically significant symptom reductions compared with a waitlist control at four and eight weeks. The authors were clear that larger and longer studies are needed, and that remains the right position.

Federal caution. In November 2025 the FDA’s Digital Health Advisory Committee discussed generative AI mental health devices, including risks such as hallucination and symptom worsening without human oversight. At the time, the agency had authorised more than 1,200 AI-enabled medical devices, but none for mental health.

State action. Several states have moved faster. Illinois banned AI-delivered therapy to the public in August 2025, and Nevada adopted similar restrictions. Other states, including Maine, Rhode Island, Vermont and Colorado, enacted restrictions in 2026, and the list is changing, so check the current position before relying on it. California’s companion-chatbot law, in force from January 2026, requires crisis protocols, and New York requires companion chatbots to remind users periodically that they are not human.

These developments share a common thread. The technology is accessible, the people using it are often vulnerable, and the evidence and rules are still catching up.

What this means for leaders

PARRY’s significance extends beyond artificial intelligence. It offers a useful starting point for examining how people interpret behaviour, form judgements and decide whether to trust what they observe.

In organisations, these processes influence recruitment, leadership, collaboration, conflict resolution and the management of culturally diverse teams. People rarely encounter behaviour without interpreting it. The challenge is recognising that an interpretation is not necessarily an accurate explanation.

Three lessons are particularly relevant.

1. Context shapes judgement

The same behaviour can carry different meanings depending on the situation, cultural background, professional expectations and relationship between the people involved. Silence might indicate reflection, uncertainty, disagreement or respect for hierarchy. Direct communication might be interpreted as confidence in one context and discourtesy in another.

Effective leaders resist the temptation to treat their first interpretation as the only plausible explanation. Cultural intelligence helps them recognise these differences, examine their assumptions and adapt their approach when working across cultural contexts.

At London Intercultural Academy, our Certified in People & Cultural Intelligence (PCI™) and Working Across Cultures programmes address these capabilities.

2. Ask before assuming

Human beings routinely infer intentions and emotions from limited information. A colleague who challenges a proposal may be perceived as difficult, while someone who remains quiet may be judged as disengaged. In both cases, the observed behaviour is real, but the explanation may be wrong.

This is where emotional intelligence becomes important. Recognising emotional cues, managing our own reactions and exploring alternative explanations can help prevent premature judgements from becoming interpersonal conflict.

Our Emotional Intelligence Training for Leaders explores how these capabilities can be developed and applied in professional settings.

3. Candour depends on trust

People are more likely to share concerns, admit mistakes and challenge assumptions when they believe they can do so without disproportionate negative consequences. When that confidence is absent, silence can be mistaken for agreement and apparent consensus can conceal unresolved problems.

Psychological safety helps create the conditions for more open communication, constructive disagreement and learning. It does not mean removing accountability or avoiding difficult conversations. It means making it possible to have those conversations honestly.

The same principle matters when organisations introduce AI-based tools for employee support or feedback. People should understand whether they are interacting with a human or an AI system, what happens to the information they share, and when a human can intervene.

Our Psychological Safety Training explores how leaders can build environments in which people feel able to speak up and contribute.

From Insight to Practice

These three capabilities share a common foundation: better judgement begins with recognising the limits of our own assumptions.

AI can produce increasingly convincing representations of human communication. People, meanwhile, continue to interpret one another through the filters of experience, expectation and culture. Developing cultural intelligence, emotional intelligence and psychological safety helps leaders navigate that reality more thoughtfully.

The objective is not to eliminate interpretation, but to make it more informed, reflective and open to correction.

A closing thought

Colby’s program was a small, theory-driven experiment. It was never a therapist, and it was never meant to replace one. Its legacy is a set of questions about how we test whether a machine represents a mind accurately, and how we decide when an impression of understanding is earned. Those questions now shape how medicine, regulators and employers approach AI, and they are questions for leaders as well.

If you would like to explore how cultural intelligence, emotional intelligence or psychological safety can be developed in your organisation, our team can help you choose the right programme.

Sources

• Weizenbaum, J. (1966). ELIZA: A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45.
• Colby, K. M. (1971). Artificial paranoia. Artificial Intelligence, 2(1), 1–25.
• Colby, K. M., Hilf, F. D., Weber, S., & Kraemer, H. C. (1972). Turing-like indistinguishability tests for the validation of a computer simulation of paranoid processes. Artificial Intelligence, 3, 199–221.
• Heinz, M. V., Mackin, D. M., Trudeau, B. M., et al. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI, 2(4). https://doi.org/10.1056/AIoa2400802
• MIT Technology Review (2025, 28 March). The first trial of generative AI therapy shows it might help with depression.
• STAT News (2025, 2 July). Woebot mental health chatbot shuts down; founder says AI moving faster than regulators.
• Psychiatric Times (2025, November). FDA committee meets on generative AI digital mental health devices.
• Stateline (2026, 15 January). AI therapy chatbots draw new oversight as suicides raise alarm.
• Salovey, P., & Mayer, J. D. (1990). Emotional intelligence. Imagination, Cognition and Personality, 9(3), 185–211.