Co-authored by Jill Noorily, Mason Granof, and Kathryn Ledley at the Division of Digital Psychiatry at BIDMC.<\/p>\n
This post is part 4 of a series.<\/p>\n
A patient with Bipolar<\/a> II disorder presents for follow-up after a recent hypomanic episode<\/a>. During the episode, he began using an AI chatbot each night after his partner went to bed, initially as a place to brain dump the ideas racing through his head before trying to sleep. As the conversations grew longer, sleep was pushed aside.<\/p>\n The chatbot helped organize scattered thoughts into increasingly coherent ideas, eventually transforming a business concept into plans for a book, complete with chapter outlines and marketing<\/a> strategies. As his confidence<\/a> grew, he became convinced the project would be a major success and began asking friends and family for money to pursue it. When they expressed hesitation, he grew frustrated and returned to the chatbot, which continued refining, expanding, and validating the idea. <\/p>\n The chatbot did not cause his hypomania, but it changed how the episode unfolded, building scaffolding and accelerating ideas that otherwise might have remained fleeting.<\/p>\n During our Learning Collaborative with the Society of Digital Psychiatry, we heard several versions of this story. Clinicians discussed patients using chatbots<\/a> to extend obsessive rabbit holes, reinforce rumination, or keep expansive thinking going for hours. In our last article<\/a> in this series, we argued that treating AI as \u201cjust a tool\u201d can create risks that are just as harmful as risks created by anthropomorphic AI. Our final collaborative session sparked a conversation about situations in which a chatbot functions as an extension of someone\u2019s own thinking.<\/p>\n Psychiatrists Alexandre Hudon and Emmanuel Stip described this idea in a paper on \u201cthe collapse of alterity,\u201d a dynamic they describe in psychosis that may extend to hypomania. In the paper, they argue that during certain psychiatric<\/a> states, the chatbot may stop feeling like an independent \u201cother\u201d and instead feel like a continuation of your own cognition<\/a>, which presents more nuanced risks [1]. The same framing can apply to \u201cAI psychosis,\u201d and a chatbot may trigger a new psychotic episode<\/a>, but is much more likely to exacerbate the trajectory of a current illness [2].<\/p>\n This type of maladaptive use may interrupt reality testing by affirming a line of thinking without the grounding another human being would provide. It can potentially accelerate thoughts in the hyperconnective state characteristic of certain presentations of mania<\/a>, hypomania, and psychosis. Chatbots may also be more willing to engage with ideas that others would find difficult to understand, making this technology particularly attractive to someone experiencing altered cognition. Over time, particularly during hypomania or obsessive thinking, the chatbot can become the primary place someone goes to test ideas. The more that happens, the less those ideas encounter the friction that other people naturally provide.<\/p>\n Tips for the patient: <\/p>\n Get a second (artificial and human) opinion: If one model seems convinced your idea is brilliant, check it with another model, though remember both may be inclined to agree with you. There is some early evidence that newer AI models may be less likely to engage in these types of discussion, so newer models may be safer [3]. What\u2019s next<\/p>\n
\nWatch for functional changes: Pay attention<\/a> to whether AI conversations are replacing sleep, pulling you away from friends or family, or becoming the first place you turn with every new idea. While we are still learning much about AI harm, it is clear that longer conversations are associated with worse outcomes.
\nTry turning off AI history: To make a model a bit less personalized, restart or turn off memory<\/a>. With too much history about you, AI models may begin to make statements that feel correct even when they are wrong.
\nSee how easily AI makes connections: Ask a chatbot to connect two completely unrelated ideas. You\u2019ll quickly discover that generating convincing connections is exactly what these models are designed to do. We learned much about how Pluto and pineapples are related in drafting this piece (\u201cboth are famous victims of misclassification\u201d)\u2026.<\/p>\n