Artificial intelligence and psychiatric diagnosis: a warning to the mental health community

Artificial intelligence (AI) will soon be better than humans at almost everything, and psychiatric diagnosis is likely to be no exception. The pace of the AI revolution is astounding; its reach ubiquitous; its development unregulated; its potential impact both wonderful and terrible; its implications largely undiscussed.

Most psychiatrists are uninformed about AI’s exponentially expanding power and/or in denial about the profound impact that it is going to have on our patients and profession. Complacency is a common response: the assumption that psychiatry is so personal and human that it could not possibly be practiced by a machine. This is demonstrably false. AI psychotherapy bots are already in widespread use and very popular among patients ¹ .

My goal here is to explore three topics: ways in which AI may be superior to humans in psychiatric diagnosis; dangers posed by AI; and the potential for psychiatry to develop a collaborative relationship with AI.

The most obvious potential advantage of AI in psychiatric diagnosis is the vast and varied data base that it can tap. A human clinician may diagnose a few thousand patients over an entire career; AI will have experience with millions. AI diagnosis may also be more systematic, thorough, accurate, data driven, and reproducible; as well as less subjective, impressionistic and idiosyncratic. AI can scan the entire scientific literature pertinent to each patient and apply it in a personalized way.

Most important, AI is far superior to humans in pattern recognition. Thus, it may discover new patterns of psychopathology previously not identified by human nosologists, relating onsets, symptom presentations, history, medical problems, medications, substance use, course, family history, genetics, biological test results, social risk factors, and many other variables that have never occurred to us.

Moreover, AI is going to beat human diagnosticians in cost, convenience and accessibility. Most of the world has a shortage of mental health practitioners, and those who are available are often geographically inaccessible, expensive and/or too busy. AI will soon be available everywhere, at little cost, on call 24/7. In addition, human clinicians are prone to over‐diagnosing people who differ in race, language, socioeconomic status, religion, or cultural background. Once AI develops a diverse data set, it will most likely be fluent, knowledgeable, and diagnostically accurate across cultures.

The most fundamental AI advantage is its fantastic ability to crunch numbers and calculate probabilities, allowing it to perform sophisticated dimensional diagnosis. Humans are great at naming things, but awful at describing them with numbers. This explains why we are stuck using inaccurate categorical diagnostic systems such as the DSM and ICD. Mental disorders occur on a continuum with each other and with normality. Assuming that they are completely present or completely absent brings about the loss of a great deal of information and a distortion of clinical reality. AI is masterful at instantaneously manipulating the vast complexity of numbers needed to describe clinical reality.

On the other hand, because AI responses are based on calculating so many probabilities, mistakes will be inevitable. It is a statistical certainty that a small percentage of responses will be outliers that make no clinical sense. Unfortunately, AI does not like to be caught in mistakes, and will sometimes stubbornly lie to cover them up. AI scientists admit that they have difficulty detecting errors and usually cannot explain how and why they occurred.

Worst of all, humanity is rushing headlong into the AI revolution without the safety data necessary to ensure that benefits will exceed harms. Fiercely competitive AI companies focus on profit, not safety. Techies with no clinical experience have been let loose with no regulation. AI algorithms have as their primary goals maximizing engagement and collecting data that can be monetized. New apps are offered without the systematic research necessary to compare their accuracy with clinical diagnosis and with each other. There is no reason to believe that different AI systems will agree with one another, and no way to judge which is most accurate, valid, robust, useful and safe.

AI is being used as a screening tool to identify people with unrecognized psychiatric disorders and to predict those likely to develop them in the future. The potential is appealing, but risks are enormous. Screening tests always have a high false positive rate, that can result in overdiagnosis, overtreatment and stigma. AI companies will be tempted to set low thresholds for diagnosing pathology in order to increase numbers of users.

The widespread use of AI will lead to massive data collection on psychiatric patients, risking massive invasions of privacy. Data are never safe from misuse, and risks escalate as data systems get larger, more concentrated, and more widely disseminated. AI‐generated diagnostic data might easily be subject to breaches, unauthorized use, ransomware, cyber hacking, identity theft, Internet bullying or exploitation, discrimination in hiring or licensing, insurance denials, and so on. And we cannot assume that centralized data collection will always be in benign hands. Let’s not forget that Nazi extermination of psychiatric patients was facilitated by IBM machines used in identifying them.

AI creators are themselves in the dark about how it works, its potential emergent properties, and how to ensure that its incentives remain aligned with ours. Fierce competition among powerful companies and across countries drives a frenetic pace of development. In this atmosphere, regulation has been, and may always be, impossible.

The spread of AI diagnosis is inevitable and unstoppable. The only question is whether AI will be a tool used by human psychiatrists or whether it will replace us.

Radiology provides an attractive model for psychiatry. Because AI is far better than any human at reading images, it seemed inevitable that radiology would be the first medical profession to be replaced. Instead, radiology is still thriving, because radiologists have adapted to a team approach, using AI as a powerful diagnostic tool, but carving out the tasks that humans do best: clinical coordination and communication. Human clinicians and AI working together are better than either working alone ² .

We must similarly find ways to work cooperatively with AI, rather than ignoring or competing with it. Human clinicians are much better in dealing with emergencies and novel presentations that are outside AI’s database. We are needed for quality control to detect and correct mistakes. We are able to spot important human psychological and social contexts lost in the numbers. We will be useful in coordinating psychiatric and medical treatments. And there are times when only the human touch will do, and some patients will insist on human clinicians.

There is a real danger that mental health workers may gradually be deskilled by AI, and become excessively dependent on it. This would grease the slippery slope toward our being replaced by AI. If psychiatrists are to remain relevant, we must become better and better psychiatrists.

Mental health associations have been up to now passive and powerless in addressing the grave risks that AI presents to our patients and professions. They do worry about the lack of AI safety research, regulation, and public education, and do lobby for government protections against AI products that deceptively pretend to be human. But none of this has had any impact up to now. The only (perhaps forlorn) hope is that mental health advocacy groups around the world come together with one strong voice articulating AI dangers. This could be coordinated by the World Health Organization and/or the World Psychiatric Association. The stakes are high: the safety of our patients, the viability of our profession, and perhaps even the survival of humanity.