Psychology and AI

Psychology and artificial intelligence (AI) have shaped each other since both fields took their modern form in the 1950s. Psychology supplied early models of how minds solve problems and learn, which inspired AI; today AI tools are used to study behavior, support assessment and deliver some mental health services, while psychologists study how people think about and interact with AI. This page explains both sides of the relationship, what the research shows so far, and the ethical questions it raises.

Foundations of AI in Psychology

Historical Evolution

The term "artificial intelligence" was coined for a 1956 summer workshop at Dartmouth College organized by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. Around the same time, Allen Newell and Herbert Simon built the Logic Theorist and later the General Problem Solver, programs explicitly based on how people reported solving problems. Their work helped launch both AI and the cognitive psychology view of the mind as an information processor. Alan Turing's 1950 paper "Computing Machinery and Intelligence," which proposed the imitation game now called the Turing test, framed questions both fields still ask.

In 1966 Joseph Weizenbaum created ELIZA, a simple program that imitated a client-centered therapist by turning users' statements into questions. Weizenbaum was alarmed by how readily people attributed understanding to it, a reaction now called the ELIZA effect, and he became an early critic of using computers in therapy.

A second strand came from the brain. Frank Rosenblatt's perceptron (1958) was a simple artificial neuron inspired by biology. In 1986 David Rumelhart, James McClelland and colleagues published Parallel Distributed Processing, showing how networks of simple units could learn, a "connectionist" approach that was as much a theory of human cognition as an engineering method. These ideas underpin today's deep learning systems, which made rapid advances in image and speech recognition in the 2010s and led to the large language models (LLMs) that became widely available from 2022.

Core Technologies

Machine Learning

Machine learning refers to algorithms that learn patterns from data rather than following hand-written rules. In psychology, it is used to predict outcomes such as treatment dropout or relapse from many variables at once, to classify brain scans, and to find structure in large datasets. A recurring lesson is that models which perform well in one dataset often perform worse in new settings, so external validation is essential.

Natural Language Processing

Natural language processing (NLP) lets computers analyze text and speech. Researchers use it to study language patterns associated with mood, to code therapy transcripts for techniques used, and to analyze open-ended survey answers at scale. Large language models can also generate conversational responses, which is what makes AI chatbots possible, and what raises questions about accuracy and safety.

Computer Vision and Sensing

Computer vision analyzes images and video, for example eye movements, facial movements or gait. Smartphones and wearables add data on activity, sleep and phone use. Claims that systems can reliably read emotions from faces are contested: a 2019 review by Lisa Feldman Barrett and colleagues concluded that facial movements do not map reliably onto specific emotions across people and contexts.

Theoretical Frameworks

Several ideas frame current work. Digital phenotyping, a term popularized by Jukka-Pekka Onnela and colleagues, refers to measuring behavior moment-by-moment from personal digital devices. Computational psychiatry uses mathematical models of learning and decision-making to understand mental health conditions. And human-AI interaction research, drawing on social psychology, studies trust, anthropomorphism and over-reliance on automated advice.

Psychology Studying AI

A newer development is psychologists studying AI systems themselves. In a 2023 study, Marcel Binz and Eric Schulz gave a large language model classic tasks from cognitive psychology and found it solved many reasoning problems well but showed differences from people in others, such as causal reasoning. Researchers have tested whether LLMs pass false-belief tasks used to study theory of mind, with mixed and hotly debated results: models sometimes succeed on standard versions but fail when tasks are subtly changed, which suggests they may rely on familiar patterns rather than human-like understanding.

This work raises old questions in new forms. Do systems that produce human-like behavior have human-like minds? Can AI models simulate human research participants, and should they? A 2025 study in Nature by Binz and colleagues trained a model, called Centaur, on data from many psychology experiments and found it predicted human choices in new experiments well, though critics argue that predicting behavior is not the same as explaining it.

AI in Therapeutic Applications

AI Therapists and Chatbots

Mental health chatbots range from rule-based programs that deliver scripted exercises drawn from cognitive behavioral therapy to generative AI systems that produce free-form conversation. The evidence is still early:

• A 2017 randomized trial by Kathleen Fitzpatrick and colleagues found that young adults using a rule-based CBT chatbot for two weeks reported larger reductions in depression symptoms than those given an information e-book.
• In 2025, Michael Heinz and colleagues at Dartmouth published one of the first randomized trials of a generative AI therapy chatbot, built and tested by researchers, which found greater symptom reductions over eight weeks than a waitlist control in adults with depression, anxiety or eating concerns.
• Meta-analyses of conversational agents for mental health generally report small to moderate short-term benefits, with high variation between studies, small samples, short follow-ups and frequent dropout.

Importantly, these findings apply to purpose-built tools tested in trials, often with human oversight. General-purpose chatbots used for emotional support have not been tested in the same way, and they are not regulated as medical devices or bound by professional ethics.

Virtual Reality Therapy

Virtual reality (VR) allows people to practice facing feared situations, such as heights or flying, in a controlled way. Meta-analyses suggest VR-based exposure for specific phobias can be about as effective as in-person exposure, and VR has been studied for social anxiety, PTSD and psychosis. AI can make virtual characters respond more naturally, though this is still an emerging area. See our guide to exposure hierarchies for the principles behind exposure-based approaches.

Digital Phenotyping

Smartphones and wearables can passively collect data such as sleep timing, physical activity, phone use and voice characteristics. Researchers are testing whether changes in these signals can provide early warning of relapse in conditions such as bipolar disorder or psychosis. Results are promising in some studies but inconsistent, predictions are often imprecise for individuals, and continuous monitoring raises obvious questions about privacy, consent and turning normal variation into "symptoms."

Crisis Intervention

Some crisis services use machine learning to help triage incoming messages so that people at highest risk are answered first by human counselors. This is a support for humans rather than a replacement. Research and journalistic investigations have documented general-purpose chatbots responding inappropriately to people in distress, including failing to recognize risk or reinforcing harmful thinking. Professional bodies stress that AI should not be relied on in a crisis; people in crisis should contact emergency services or a crisis line.

Cognitive Assessment and Diagnosis

AI-Enhanced Neuropsychological Testing

Digital versions of traditional tests can record far more than a final score. A digital clock-drawing test, for example, captures every pen stroke and pause, and machine learning analysis of these details has distinguished people with cognitive impairment from healthy controls in research studies. Computerized adaptive testing, which selects each item based on previous answers, can shorten assessments. These tools support, but do not replace, the clinical judgment involved in neuropsychological testing.

Multimodal Assessment

Researchers combine several data sources to look for markers of mental health and cognitive conditions:

• Speech: features such as speaking rate, pauses and pitch variability differ on average in groups with depression or Parkinson's disease, and changes in word choice and coherence have been studied in psychosis and dementia.
• Writing: language patterns, such as more frequent use of first-person singular pronouns, are associated on average with depression in studies by James Pennebaker and others, although the effects are small.
• Behavior: sleep-wake patterns and activity levels relate to mood states.

These are group-level associations. They can support research and, in time, screening, but they are not reliable enough to diagnose individuals on their own.

Diagnostic Support Systems

Clinical decision support tools analyze patient data to flag possible diagnoses or risks for clinicians to consider. They face significant challenges: training data can reflect historical biases, models can be opaque, and clinicians may over-trust automated suggestions, a well-documented phenomenon called automation bias. Prediction models for suicide risk, for instance, have generally shown low positive predictive value, meaning most people flagged as high risk do not go on to make an attempt, which limits their usefulness for individual decisions. See our overview of psychological testing for how assessment is normally conducted.

Precision Neuropsychology

The Emergence of Precision Neuropsychology

"Precision" approaches aim to tailor assessment and treatment to the individual rather than to diagnostic averages. In neuropsychology, this involves adaptive testing, larger and more diverse norms, and statistical models that combine test results with other information to estimate a person's likely trajectory. This is an active research area rather than routine practice.

Neuroimaging and AI

Machine learning applied to brain scans has shown promise for detecting patterns associated with Alzheimer's disease before symptoms become obvious, and for exploring whether brain features predict response to treatment. However, a 2022 study by Scott Marek and colleagues showed that reliable links between brain imaging and individual differences in behavior generally require thousands of participants, and many earlier, smaller studies likely overestimated accuracy. See our guide to neuroscience for more on imaging methods.

Cognitive Rehabilitation

After brain injury or stroke, computerized programs can adjust task difficulty automatically to keep training challenging but achievable, and can track progress over time. Evidence suggests such training improves performance on trained tasks, with more limited transfer to everyday life; it works best as part of a broader rehabilitation plan. Related approaches are covered in cognitive remediation.

Neurofeedback and Brain Training

Neurofeedback trains people to change their own brain activity using real-time feedback from EEG or, in research settings, fMRI. Evidence for clinical benefit is mixed and debated; see biofeedback and neurofeedback. Commercial "brain training" games have been heavily marketed, but large reviews have found little evidence that they improve general cognitive ability or everyday functioning.

Brain-Computer Interfaces

Current State of BCIs in Psychology

Brain-computer interfaces (BCIs) record brain activity and use machine learning to translate it into commands for a computer or device. Most advanced BCIs are implanted and remain experimental, available mainly to research participants with severe paralysis.

Clinical Applications

Communication. In a 2021 study in Nature, Francis Willett and colleagues showed that a man with paralysis could type about 90 characters per minute by imagining handwriting, decoded from an implanted device. Later studies decoded attempted speech into text or synthesized voice at speeds approaching natural conversation for some participants.

Rehabilitation. Non-invasive BCIs have been tested in stroke rehabilitation, where detecting a person's attempt to move triggers assistance or stimulation of the limb. Reviews suggest modest benefits for motor recovery, but trials are small.

Mental health research. Researchers are investigating "closed-loop" brain stimulation that adjusts in response to recorded brain activity. These studies involve very small numbers of participants and remain experimental.

Cognitive Enhancement

There is much speculation about using BCIs to enhance memory, attention or learning in healthy people. At present, such uses are largely theoretical, and the risks of brain surgery make them unjustifiable outside serious medical need. Non-invasive brain stimulation for enhancement has produced inconsistent results.

Ethical and Privacy Concerns

BCIs raise new questions about "neural data" and mental privacy. Who owns recordings of brain activity, and could they reveal more than users intend? Some jurisdictions have begun to respond: Chile amended its constitution in 2021 to protect "neurorights," and several U.S. states have added neural data to privacy laws. Security of implanted devices, long-term support when companies or trials end, and informed consent from people with severe disabilities are further concerns.

Personalized Mental Healthcare

From One-Size-Fits-All to Precision Treatment

Treatments that work on average do not work for everyone. Researchers hope AI can help predict which approach will suit which person, using information such as symptoms, history, preferences and social circumstances. This goal is promising but not yet realized in routine care.

Predictive Analytics

Prediction models are being developed for outcomes such as relapse, dropout from therapy and likely response to different therapies. Results so far are modest: many models predict only slightly better than simpler methods, and few have been tested in new populations. Any use in practice requires careful validation across diverse groups to avoid reinforcing existing inequalities.

Treatment Matching

One research approach is the Personalized Advantage Index, introduced by Robert DeRubeis and colleagues in 2014, which uses information about individuals to estimate which of two treatments might suit them better. For example, models might suggest that a particular person is more likely to benefit from behavioral activation than from cognitive restructuring. Such tools are still being tested and should inform, not override, shared decision-making between clients and clinicians.

Dynamic Treatment Optimization

Routine outcome monitoring, in which clients complete brief questionnaires each session, already helps therapists notice when someone is not improving. AI may make this feedback more informative, for example by flagging patterns associated with deterioration, so that clinicians can review the plan with the client.

Digital Mental Health Ecosystem

Integrated Care Platforms

Digital mental health now includes video therapy, guided internet-delivered CBT, symptom-tracking apps and AI tools. Guided internet-based CBT, in which a person works through structured online modules with support from a trained guide, has strong research support for anxiety and depression, with outcomes comparable to face-to-face therapy in several meta-analyses. See our guide to online therapy.

Continuous Monitoring and Early Detection

Combining clinical records with data from apps and wearables could, in principle, help detect early signs of relapse and prompt timely support. In practice, integration is limited, data quality varies, and people may not want continuous monitoring. Any such system needs clear consent and human review before action is taken.

Democratization of Mental Health Services

Digital tools could extend support to people who face barriers such as cost, distance or long waiting lists, and some can be offered in many languages. But access is uneven: people with limited internet access, digital skills or privacy at home may be left out, and engagement with self-guided apps is often low, with many users stopping within days. Evidence-based digital tools work best when integrated with human support.

Population Mental Health

Aggregated data, such as search trends or anonymized app data, may help public health agencies spot changes in population wellbeing. Such approaches have notable failures as well as successes, and they raise questions about privacy and consent at a population level.

Ethical Challenges and Concerns

The Ethics Gap in AI Mental Health

Researchers and professional bodies have identified several recurring problems with AI tools used for mental health support:

• Lack of context: generic advice may ignore a person's culture, circumstances or history.
• Agreeing too readily: chatbots tuned to be agreeable may validate unhelpful or false beliefs rather than gently challenging them, a tendency sometimes called sycophancy.
• Crisis failures: systems may fail to recognize risk or to direct people to urgent help.
• Illusion of empathy: simulated warmth can encourage emotional reliance on a system that cannot care or be accountable.
• Unclear consent: users often do not understand how their data are used or what the system can and cannot do.

The World Health Organization's 2021 guidance on the ethics of AI for health emphasizes protecting autonomy, promoting safety, ensuring transparency, fostering accountability, ensuring equity and promoting sustainability.

Privacy and Data Security

Mental health information is among the most sensitive personal data. In the United States, HIPAA covers health care providers and insurers but not many consumer wellness apps, and regulators and researchers have documented apps sharing data with advertisers. Risks include data breaches, misuse by employers or insurers, and future uses that users never anticipated.

Algorithmic Bias and Health Disparities

AI systems trained on unrepresentative data can perform worse for under-represented groups. Speech and language models may misinterpret dialects or non-native speakers; image-based systems have shown lower accuracy for darker skin tones; and models built in specialist centers may not work in community settings. Auditing for bias and training on diverse data help but do not solve the problem entirely. See our article on cognitive biases for the human side of this issue.

Human Agency and Autonomy

Predictions about someone's mental health future could become self-fulfilling or undermine hope, and opaque recommendations are hard to question. Good practice keeps humans responsible for decisions, explains the basis and limits of any prediction, and respects people's right to decline automated assessment.

Human-AI Collaboration

The Augmented Clinician Model

Most experts see the most promising role for AI as supporting clinicians rather than replacing them: drafting notes, summarizing records, monitoring outcomes, supporting training and extending structured self-help between sessions. This could free time for the human work of assessment, relationship and judgment, provided automation bias and errors are guarded against.

Training and Education

Psychology training increasingly covers AI literacy: how models work, how to interpret their outputs critically, their limitations and biases, and the ethical and legal issues they raise. Clinicians also need to maintain core skills so that they can work without AI and recognize when it is wrong.

Maintaining the Human Touch

The therapeutic alliance, the collaborative bond between client and therapist, is one of the most consistent predictors of therapy outcome. Whether a relationship with a chatbot can provide similar benefits is an open question. Some users report feeling understood by AI, while critics argue that real empathy requires a being who can care and be accountable.

Acceptance and Resistance

Attitudes vary. Surveys suggest many people are willing to try AI tools for mental health, especially when other help is hard to access, while clinicians are often more cautious, citing safety, privacy and the nature of therapy. Both perspectives point to the same need: better evidence and clear standards.

Future Directions

Emerging Technologies

Developments to watch include more capable language models trained specifically for health settings, better speech and wearable sensing, improved BCIs for communication, and tools that help researchers analyze qualitative data. Each will need independent evaluation in diverse populations before claims about benefits can be trusted.

Regulatory Frameworks

Regulation is evolving. The European Union's AI Act, which entered into force in 2024, sets risk-based obligations for AI systems, and AI used as a medical device is also regulated under medical device rules. In the United States, the Food and Drug Administration regulates some software as a medical device, but many wellness apps fall outside its oversight. Professional associations are developing guidance for psychologists on the ethical use of AI.

Societal Implications

Wider questions include whether constant digital monitoring changes how people experience their own emotions, whether AI companions affect loneliness and human relationships for better or worse, and who benefits economically from mental health data. These are questions for social psychology and public debate as much as for technology.

The Road Ahead

Principles commonly proposed for responsible AI in psychology include human-centered design, rigorous independent testing before widespread use, transparency about capabilities and limits, strong privacy protections, attention to equity, and clear human accountability.

Conclusion

Psychology and AI have a long shared history, from early models of problem-solving to today's neural networks. AI now offers genuine tools for research and some promising approaches to extending mental health support, but evidence is still early, and real risks around safety, privacy and bias remain. For now, the most realistic picture is one in which AI supports, rather than replaces, the human understanding at the heart of psychology.