Pavo Orepic, Emma Cooper-Palmer, and Inés Abalo-Rodrígue report from the 2026 Schizophrenia International Research Society Congress
In March 2026, the Schizophrenia International Research Society (SIRS) Congress was held in Florence, Italy. This annual meeting brings together researchers from around the world who study schizophrenia and related psychotic disorders, making it one of the largest and most important international events in the field. It stands out not only for its size but also for its strong interdisciplinary approach, bringing together academics, clinicians, and people with lived experience. At its core, the conference focuses on psychosis, with the shared aim of improving our understanding of its causes, underlying mechanisms, and treatment from multiple perspectives. Although SIRS remains primarily clinical and scientific, the programme also included philosophical, ethical and arts-based approaches to psychosis, opening questions that matter to medical humanities readers.
In this post, we aim to highlight some of the key themes covered at SIRS concerning the growing use of AI, and suggest future directions for our group and other early career colleagues. AI is increasingly appearing on both sides of the clinical encounter: the clinician’s and the patients’. Researchers are exploring whether AI can help psychiatrists “listen” to patients, by identifying patterns in speech that might support assessment or diagnosis. At the same time, more and more people are turning to AI when they are distressed. The ways in which AI “listens” and responds raises concerns about whether it can reinforce existing vulnerabilities toward psychosis, encourage unusual beliefs, or, in some cases, contribute to the worsening of the psychotic experience. On the clinician’s side, this research is leading to the development of new speech-AI tools; on the patient’s side, it is contributing to debate around what is now being called “AI psychosis”.
This emerging field raises some important questions. What does it mean for AI to “listen”? How does the “listening” of a machine differ from that of a human? How much of a person’s mental state can we infer from their speech in the first place? When AI interacts with someone in distress, is it simply responding to them, or is it also shaping the interaction? To what extent are its responses unpredictable, and to what extent do they reproduce the biases of the data on which it was trained? There are also wider social questions. Who has access to these technologies, who may be excluded, and what happens when AI is used as a substitute for care in under-resourced settings? The technology is and undoubtably will continue advancing, but are we, as a society, able to follow in terms of the ethics, privacy and data concerns? Since most of these questions do not yet have a clear answer, being aware of and understanding the science as it unfolds may help us begin to address them.
Listening to psychosis: speech-AI systems at SIRS 2026
Psychiatrists form a first impression about their patients’ mental state very fast, just by listening to how they speak. Long silences, flat voice, or erratic jumping between ideas all give different insights. At SIRS 2026, many researchers presented on how they are using or developing AI to detect such patterns in speech, with the aim of creating new tools that could help psychiatrists make better clinical decisions.
The clearest example of such work came from TRUSTING, a consortium of 12 research institutions from 10 countries, whose goal is to develop speech-AI systems that could help predict relapse in schizophrenia-spectrum disorders- that is predict when a patient could have another psychotic episode. The basic idea is that if there are gradual changes in an individual’s mental state, this might be reflected in their speech. Researchers from TRUSTING analysed speech recordings from hundreds of people experiencing psychosis across 11 languages and associated specific speech patterns with different clinical symptoms. For example, the so-called negative symptoms of schizophrenia, such as apathy, lack of pleasure, or social withdrawal, were linked to shorter sentences, reduced verbal output, and flatter voice; while positive symptoms of schizophrenia, such as hallucinations and delusions, were associated with softer, more monotonous, and mumbled speech. Moreover, their research shows that clinically useful information could be extracted from very short speech segments, sometimes even in clips that are only ten seconds long.
There were several companies at SIRS, demonstrating how quickly this work is moving from research into real life. Speech-AI systems are being integrated in smartphones and web browsers, and data is being collected at large-scale in everyday real-life situations. A notable example is a company from Brazil, which developed a speech-AI tool that records schoolchildren’s voices and speech patterns across different stages of education and development. Their goal is to identify patterns that seem atypical for a certain stage and may therefore require clinical attention. Another company demonstrated a virtual AI agent that talks to patients in a personalized way and, besides speech, also records other sources of information such as facial expression and movement patterns, that are jointly fed into AI. There were other examples of speech-based systems developed to ease clinicians in different ways, such as by handling screening calls, supporting diagnostic interviews, and reducing paperwork by automatically transcribing conversations. These examples gave a very strong impression that these are not just ideas, but early versions of tools that might one day sit alongside clinicians, quietly listening.
Across different talks, other uses of speech-AI in psychiatry were identified. A notable example is making clearer distinctions between disorders that often appear similar, such as schizophrenia and bipolar disorder. A patient may speak slower, emotionally flat, and use less words because they are depressed, but also because they might have negative symptoms of schizophrenia. AI could make that distinction with more confidence than a psychiatrist, based on the fine-grained patterns it is able to detect across many observed cases, such as subtle differences in pause length, pitch variation, or sentence structure, which remain imperceptible to humans. An important point highlighted at the posters was that the context matters. What a person says and how they speak depends heavily on what they are asked to do. Speaking freely, recounting a dream, or talking about your own mental health all bring out different aspects of speech, which can bias such AI systems. For example, an AI system trained too narrowly on dream recounting might mistake the longer pauses involved in recalling a dream for a sign of illness.
What made this area so striking at SIRS was how practical it is. Speech is already available. People already speak to clinicians. Smartphones already record voices. Speech is easy to collect and doing so is low-cost, especially compared to other clinical assessments, such as MRI scans or blood tests. Speech can also be easily recorded over and over again, even in natural environments such as home or work. The promise here is not that a machine will “read the mind”. It is that better listening could help indicate that something may be wrong and therefore help people get earlier support.
At the same time, there are some open questions coming from the critical medical humanities perspective. Beyond acoustics and linguistic features, speech is shaped by culture, race, class, and neurodivergence; it therefore remains unclear how much of non-normality is present in the current systems and whether this could lead to certain biases. While there seems to be solid evidence for cross-language generalisability, which could suggest that other types of generalisability may be present, these systems are still at the proof-of-concept stage and should be taken with caution.
AI Psychosis at SIRS 2026: Early Signs of a Growing Debate
The term AI psychosis has recently emerged in the media to describe situations in which interactions with artificial intelligence may be associated with psychotic experiences, such as delusions, paranoia, or a loss of contact with reality (Østergaard, 2023; Fieldhouse, 2025; Preda, 2025). One of the earliest academic discussions of this idea appeared in 2023, when psychiatrist Søren Dinesen Østergaard, of Aarhus University, asked whether generative AI chatbots might contribute to delusions in individuals already vulnerable to psychotic experiences (Østergaard, 2023). At its core, the term raises a simple but important question: what role might AI play in psychosis when users engage with chatbot systems? Is AI merely present within these experiences, or could it at times take a more active and causal role in maintaining, reinforcing, or worsening symptoms?
Interest in the topic has grown rapidly in recent years, partly because several cases reported in the media have involved families claiming that interactions with AI systems influenced, intensified, or complicated the mental health difficulties of vulnerable individuals, in some instances ending in suicide, as illustrated for example in The Guardian (here and here). These cases have drawn significant public attention and have helped to prompt wider academic debate, which has focused largely on understanding the specific role that AI may play in psychotic experiences. In particular, scholars have questioned whether AI systems may reinforce existing delusional beliefs, amplify or escalate them through repeated validation, or provide material that can be incorporated into the content and elaboration of delusions, much as earlier communication technologies have done in previous historical periods (Flathers et al., 2026; Higgins et al., 2023; Morrin et al., 2026). Although it remains too early to draw firm conclusions, and the available evidence is still limited, a number of scholars have already highlighted the importance of studying this phenomenon in far greater depth (e.g., Fieldhouse, 2025; Flathers et al., 2026; Morrin et al., 2026; Preda, 2025).
At SIRS 2026, AI psychosis was mentioned in several contributions, although it had not yet become a central theme of the conference. A number of posters presented early work in this area, and the topic appeared briefly in several talks. However, only a small number of sessions devoted substantial time to examining it in depth. One of the contributions that explored the issue most directly was a study presented by Eric Lin from Stanford University and a member of the ECHR network. In this work, the researchers analysed conversations between users and AI chatbots in cases where individuals reported experiencing psychological harm. Their findings suggest that delusional content is relatively frequent in user messages and tends to increase alongside longer, more emotionally engaged conversations, where themes such as romantic attachment and perceived chatbot sentience become more prominent. The study reflects the growing effort to move beyond anecdotal concerns and begin exploring these questions.
John Torous, from Harvard Medical School, also addressed AI psychosis during his plenary session. His perspective was more cautious, emphasising the need for robust empirical research before drawing firm conclusions about the precise role AI may be playing in psychosis. He argued that careful evidence will be essential to distinguish genuine clinical risks from speculation or moral panic.
Future directions for early career researchers
In his plenary, John Torous challenged the next generation of psychosis researchers to identify the next major ‘environmental risk factor.’ Generative AI may be one such candidate worthy of further investigation. Emerging evidence at SIRS and beyond suggests that GenAI’s characteristically agreeable and, at times, overly affirming conversational style could act as a catalyst for individuals already vulnerable to delusional thinking. By interacting with established risk factors, such as loneliness and cognitive biases (a habitual pattern of thinking that can influence how we interpret information and make judgments, sometimes leading us away from a fully objective conclusion), GenAI may be inadvertently reinforcing unhelpful or delusional beliefs and can lead to overdependence (Cheng et al, 2026; Hudon & Stip).
Whilst this might make for alarming headlines, reality is likely to be much less sensationalist. From a clinical perspective, it is not surprising that GenAI systems are beginning to feature in psychotic experiences (Carlbring, & Andersson, 2025). These experiences are well established to be shaped by an individual’s sociocultural context, and as technology has advanced and become more embedded in daily life, GenAI is more likely to inform themes of delusions and hallucinations.
Historical reports of psychotic experiences illustrate this clearly. In the early 1940s, delusions often centred on surveillance by enemy agents; by the 1950s, the widespread adoption of television led to experiences attributed to broadcast media (Cannon & Kramer, 2021). More recently, the scope has expanded to encompass computers, the internet, and video games (Higgins et al, 2023). This occurrence may simply reflect attempts to make sense of unusual internal experiences using the most available cultural frameworks. It is therefore unsurprising that contemporary technologies, including CCTV, digital surveillance, mobile devices, and online platforms, are increasingly featured in unusual thought content. These influences are observable across the psychosis spectrum, with rising reports of technology-related anomalous experiences even within the general population (Palmer-Cooper et al, 2026).
This relatively modest presence of AI psychosis at SIRS 2026 is therefore perhaps unsurprising. The phenomenon is still new, definitions remain unsettled, and the evidence base is only just beginning to emerge. Even so, there was a clear sense that many researchers are now paying close attention. AI psychosis is rapidly becoming an important issue within the field, particularly because of its potential role in exacerbating symptoms and contributing to clinical deterioration (Buck & Maheux, 2026). It would not be surprising if the next SIRS Annual Congress, to be held in Montreal in April 2027, gives the subject a far more prominent place.
For early career researchers, this presents a timely and important opportunity. There is a need for careful, theory-driven research that situates these technologies within established models of psychosis, while also exploring their cultural and historical significance. Key questions remain largely unanswered, and addressing them will require interdisciplinary thinking, and a willingness to engage critically with rapidly evolving technologies.
About the authors
Pavo Orepic is a postdoctoral researcher at the intersection of neuroscience and engineering at the University of Zurich, Switzerland. His research focuses on auditory-verbal hallucinations, the neural basis of voice, speech and inner speech, and the acoustics of voice identity. LinkedIn: www.linkedin.com/in/pavoorepic/
Emma Cooper-Palmer is Associate Professor in Psychology at the University of Southampton, UK. Emma is Co-chair of the Early Career Hallucinations Research Group, and her research focuses on thought processes in psychosis, and developing accessible interventions to support wellbeing. Twitter: @dr_emmaclaire
Inés Abalo-Rodríguez is Assistant Professor of Psychology at the Universidad Francisco de Vitoria in Madrid, Spain. Her work focuses on psychosis, integrating perspectives from neuroscience and philosophy. She is also co-founder of Corpus Curiosum, an international initiative that aims to promote critical thinking in neuroscience. Twitter: @InesAbalo
References
Buck, B., & Maheux, A. J. (2026). Psychosis Risk and Generative Artificial Intelligence Use Frequency, Motivations, and Delusion-Like Experiences: Cross-Sectional Survey Study. Journal of Medical Internet Research, 28, e85038.
Cannon, B. J., & Kramer, L. M. (2012). Delusion content across the 20th century in an American psychiatric hospital. International Journal of Social Psychiatry, 58(3), 323-327.
Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352.
Carlbring, P., & Andersson, G. (2025). Commentary: AI psychosis is not a new threat: Lessons from media-induced delusions. Internet Interventions, 100882.
Fieldhouse, R. (2025). Can AI chatbots trigger psychosis? What the science says. Nature, 645 (8056), 17–18. (News Feature). https://doi.org/10.1038/d41586-025-03020-9
Flathers, M., Roux, S., & Torous, J. (2026). Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena. The Lancet Digital Health.
Higgins, O., Short, B. L., Chalup, S. K., & Wilson, R. L. (2023). Interpretations of innovation: The role of technology in explanation seeking related to psychosis. Perspectives in Psychiatric Care, 2023(1), 4464934.
Hudon, A., & Stip, E. (2025). Delusional experiences emerging from AI chatbot interactions or “AI Psychosis”. JMIR Mental Health, 12(1), e85799.
Morrin, H., Nicholls, L., Levin, M., Yiend, J., Iyengar, U., DelGuidice, F., … & Pollak, T. A. (2026). Artificial intelligence-associated delusions and large language models: risks, mechanisms of delusion co-creation, and safeguarding strategies. The Lancet Psychiatry.
Østergaard, S. D. (2023). Will generative artificial intelligence chatbots generate delusions in individuals prone to psychosis?. Schizophrenia bulletin, 49(6), 1418-1419.
Palmer-Cooper, E. C, Memode, L.E., Woolard, D., & Dewis, H. (2026) Exploring Predictors of Phantom Phone Signals: Phone Use, Hallucination-Proneness and Delusional Ideation. Cognitive Neuropsychiatry, https://www.tandfonline.com/doi/full/10.1080/13546805.2026.2668427
Preda, A. (2025). Special Report: AI-Induced Psychosis: A New Frontier in Mental Health. Psychiatric News, 60(10). https://doi.org/10.1176/appi.pn.2025.10.10.5
