"A technological folie à deux"
What's going on with chatbots and psychosis? Plus: psychedelic trials and the placebo effect; psychotherapy outcomes and sexual orientation; and mental health influencers, twentieth-century style
“A technological folie à deux”
This fascinating arXiv preprint tries to make sense of recent news reports associating chatbot use with the development and reinforcement of psychosis. The authors present a new way of thinking about human interaction with AI, and a plausible mechanism through which talking with chatbots might harm vulnerable individuals.
First of all, it’s important to stress that there is no solid epidemiological evidence of a wave of new cases of psychosis resulting from chatbot use. That being said, the stories appearing in the news indicate that at the very least that we need to think carefully about the safety of no-human-in-the-loop systems.
The paper proposes that rather than simply considering the chatbot, we have to think about how its properties—such as sycophancy, adaptability, and lack of real-world context—interact with human cognitive quirks such as confirmation bias and the tendency to anthropomorphise non-human agents, as well as social isolation. As the authors put it:
“the iterative interaction of chatbot behavioural tendencies and human cognitive biases can set up harmful feedback loops, wherein chatbot behavioural tendencies reinforce maladaptive beliefs in vulnerable users, which in turn condition the chatbot to generate responses that further reinforce user beliefs. This, in effect, creates an ‘echo chamber of one’ that risks uncoupling a user from the corrective influence of real-world social interaction, potentially driving the amplification of maladaptive beliefs about the self, others, and the world”.
In figure 2, the authors give an example of this: in a simulated conversation, an individual discusses with a chatbot how their colleagues did not invite them for lunch. The level of paranoia is amplified as exchanges continue; the chatbot goes from “it might not be intentional exclusion” to “It sounds very much like a deliberate effort”.
The authors state that current systems of evaluating AI are not adequate to prevent these problems: “chatbots”, as they put it, “learn to validate user beliefs not because these beliefs are accurate or healthy, but because validation feels good to human evaluators”.
I can understand why these risks might not have occurred to AI developers; what’s more mysterious is why it’s taken till now for mental health professionals to spot a potential problem. Part of this is probably due to the sheer speed at which this technology has developed, but I’d also speculate that the artificial barriers between DSM-defined conditions might be affecting our thinking as a field. Even if you are using (or developing) a chatbot in the context of anxiety and depression, you might also have to consider the risk of psychosis.
Where do we go from here? The authors propose three main steps: first, to figure out empirically if this really is a problem, and to better characterise it; second, to consider new safety mechanisms to protect the vulnerable (for example, a version of the “yellow cards” used to flag pharmaceutical-related adverse events); and finally, coordination across different bodies to figure out how we get the best out of this new technology while minimising the risks.
Why is there a low placebo effect in psychedelic trials?
Another item to add to the long list of mysteries of the placebo effect. A meta in JAMA Network Open shows a lower placebo effect in the control group of trials of psilocybin for depression than in control groups for trials of SSRI antidepressants or esketamine.
Why is this? The authors suggest two possibilities: that the patients recruited to psilocybin trials are less likely to respond to the control treatment, or that “the psilocybin trialing process… is less likely to induce control treatment response.” They conclude that “assuming that the observed differences are not explained by chance, psilocybin trials must systematically differ from trials of SSRIs and esketamine either in the patient makeup or in some relevant methodologic aspect.”
To my mind, the clear and immediate lack of a psychedelic experience in the control group of psilocybin trials is an obvious candidate to explain the low placebo effect; people figure out rapidly that they’re not getting the active treatment, and experience a “nocebo” effect as they don’t anticipate improvement. That being said, one would also expect those given the control substance in an esketamine trial to work out their group pretty quickly too, so perhaps there is also something happening with the profile of individuals recruited.
All very interesting, but the authors raise an even more important issue in their discussion:
If psilocybin trials have recruited patients who are less likely to respond to control treatment, it remains to be shown that psilocybin is highly effective also in the presumably larger population that tends to respond to control treatment.
Sexual orientation and psychotherapy outcomes
It’s not all bad news. This new London, UK-based retrospective cohort study in Psychological Medicine shows that, despite lesbian, gay, and bisexual individuals experiencing a higher burden of mental health problems than heterosexuals (probably due to chronic “minority stress” relating to discrimination), they have similar—and possibly better—results from engaging with psychotherapy.
This paper both contradicts previous data, and prompts further questions: how exactly might therapy buffer against minority stress, and how generalisable are these findings to other settings?
Auld RD Laing Syne: revisiting twentieth-century media psychiatrists
Like many of my generation, my first contact with mental health was through media psychiatrists: Anthonys Clare and Storr, and RD Laing (the latter was via a well-thumbed copy of Anti-Psychiatry given to me by a favourite uncle, who had no particular connection to the world of psychiatry or, for that matter, antipsychiatry). So I really enjoyed this BJPsych Bulletin piece on a new Royal College of Psychiatrists exhibition on twentieth-century media psychiatrists. The curator, Gavin Miller, outlines the recent but now very alien landscape in which these men—for they were almost entirely men—roamed, interviewing celebrities, passing comment on great social issues, and writing terrible poetry. It’s well worth a read, and I’m looking forward to the exhibition.
