In the long-running annual series Oh, The Humanities! National Post reporters survey academic scholarship at the Congress of the Humanities and Social Sciences, which continues this week at York University, with an eye to the curious, the mysterious and the hilarious. Today, the morality of psychotherapy by chatbot.
On the one hand, AI psychotherapy is a good setup for a mildly witty New Yorker cartoon, possibly with a dog talking to a phone about a cat. On the other hand, it’s also a good setup for a horrifying Black Mirror episode.
This dark side was illustrated this week when the US National Eating Disorder Association took its chatbot therapist Tessa offline, after a viral social media report that it told a woman to lose weight, maintain a daily calorie deficit, measure and weigh herself weekly and restrict her diet. The Association’s CEO said this was against policy and “core beliefs” and said they were working to fix the “bug.”
Bullying callers to an eating disorders hotline to lose weight and eat less is just one of the many things people are concerned about with AI psychotherapy. Recent news is filled to bursting with stories about chatbots gone wild, such as Microsoft’s Bing chatbot for web searching, which told a New York Times writer it has malicious fantasies about hacking computers, that it wants to be human, that it loves him, that he is unhappy in his marriage, and that he should leave his wife and be with Bing instead.
Therapists lose their licence for saying stuff like that. But when psychotherapy is done badly by a chatbot, it’s hard to punish a “bug.”
Rachel Katz, a philosopher of ethics at the University of Toronto, told a Congress 2023 gathering of the Canadian Society for the History and Philosophy of Science about a new project to sketch the ethics of AI psychotherapy, with a focus on trust. Katz’s main contention is that outsourcing psychotherapy to chatbots alters the trust relationship between patient and therapist.
There are other problems. For example, a psychotherapy chatbot is no good in a crisis. “These apps don’t claim to be able to assist someone who is, let’s say, feeling suicidal. They will just direct you to call 911 or some other kind of crisis service,” Katz said.
But there are also potential benefits, such as filling in service gaps when human therapists are unable to meet demand; offering therapists automated guidance as an assisting “listening ear”; getting patients in touch with care quickly; and lowering the stigma of seeking mental health care by offering an easy entry, or even just helping someone admit there is a problem.
They liked the idea that there was this removal of pressure because they didn’t have to express things to another human being
One striking early discovery was how, after hearing about this in a a lecture, Katz’s students “were shockingly very okay with the idea of using a chatbot to do therapy. They liked the idea that there was this removal of pressure because they didn’t have to express things to another human being. They were in some ways talking to a wall, a very verbose wall that can say nice things back to them, but it wasn’t the same, in their view, as talking to a person.”
The main problem, though, as Katz frames it, is that “You cannot build a trust relationship with AI as it currently exists.”
As in most modern philosophy, the project to understand the ethical context here begins with a clarification of terms.
“What we colloquially call ‘trust’ is not actually trust at all,” Katz said, drawing a distinction between trusting a person and trusting an information source. Trusting an AI chatbot for therapy is not trust, Katz argues. It’s reliance. It’s “leaning on” a support that you expect to be there, but you don’t expect any interaction. You don’t trust the floor to hold you up, you rely on it.
Trust is a subtype of reliance, a special kind of reliance that draws on goodwill and vulnerability on both sides. “There is a human touch that is missing here. I’m still working out exactly what that human touch is,” Katz said. “But I do think there is something special about the aliveness of a human being that makes that trust relationship specific and special, and distinctly different from what happens when we interact with an AI, especially in a vulnerable therapeutic setting.”
The issue has been around for decades, long before the pandemic proliferation of AI therapy chatbots like Youper and Bloom.
It dates to the 1960s with the invention of Eliza, a computer program in which a patient could type to Eliza and get responses using pattern matching to parrot their words back to them. Sometimes this is even a therapeutic goal, to reflect a patient’s own ideas back to them as reassurance. But as Katz points out, Eliza did not listen, it just looped your own words back to you. Sometimes that’s enough, and maybe helpful, but it’s not psychotherapy.
There is a similar problem with modern chatbots, which are programmed to seem far more sophisticated and interactive as they follow the codified therapies of, for example, cognitive behavioural therapy, dialectical behaviour therapy, and psychodynamic therapy.
In their usual human-to-human use, “all of these therapies revolve around a discussion that takes place between a patient and a trusted professional, and I mean “trusted professional” in two different senses here. One is that the professional is trusted by the patient, there’s a relationship there. And then, in cases where applicable, there’s also trust bestowed by a professional body onto the therapist’s qualifications.”
As Katz tells it, AI psychotherapy is like talking to “a very affectionate wall,” and it is important for patients, therapists, regulators and society alike to be clear about how we think about these robot therapists. This is all the more important now that they are becoming so uncannily human-like that they basically pass the famous Turing test of being able to fool a human into believing they are also a human.
It’s a tool, Katz said, and so it needs a manual for conduct, and it needs to be clear with its disclaimers about what you cannot rely on it for. The fine print really does matter.
Post a Comment