In my work as a supervisor, teacher, and consultant to other therapists, I have heard a lot of concerns about AI in therapy. A few of my own clients have consulted ChatGPT, Claude, and the myriad other online advisors, and we have frequently been able to integrate their experience into the work we are doing.

But other than asking ChatGPT for suggestions about titles for articles I’m working on, which is something I’m really bad at myself, and for ideas about activities for my visiting grandchildren, I’ve pretty much stayed away from AI.

Except, that is, for my love of the gamified language app “Duolingo.”

I never thought of the avatars as AI

Somehow, I never thought of the adorable characters and charming stories of the program as AI-generated. I admire the use of clever tools that keep me engaged, ensuring that I get 15-20 minutes of language learning a day, almost without trying. My Spanish vocabulary and conjugation skills are improving, and my ability to understand and converse with my Spanish-speaking clients and their interpreter in Spanish is a personal huge reward.

But recently, Duolingo offered a month of free video “conversations” with some of the avatars. Hesitant at first, mostly because I wasn’t sure that my language skills were up to the point where I could have conversations, I decided to try it. What did I have to lose?

People are having more intimate relationships with AI bots

It turned out that I quickly lost any sense of superiority I might have had over anyone drawn to AI. My avatar buddies are funny, personable, highly complimentary, and encouraging. These qualities are, of course, both what attracts people to AI and what many people worry about.

Esther Perel, a widely recognized relationship therapist, tells us in an interview, “People are having more and more intimate relationships with their AI bots.”

And in a thoughtful exploration of AI on his PT blog, Dr. Steven E. Hyler writes, “These systems appear to listen patiently and respond empathetically, and they remain available no matter what is said.”

I am charmed when the teenaged character Lily, with her purple hair and sarcastic comments, gently tells me she understands what I have said when I know that I’ve used the wrong words and badly bungled the grammar. How did she understand?

I feel held when the grandmotherly Lucy, with a pencil stuck through a sort of bun hairstyle, tells me that I’m doing well.

AI bots don’t “get” privacy or boundaries

But I’m also kind of horrified when one of the characters uses something I told a different one in a previous conversation. Is there no honor among avatars? Do they not know about privacy?

I knew from the get-go not to share personal information, but I found myself disguising everything after a while. I talked about seeing a ballet, and when Lily asked where, I told her something vague about a small performing arts center. Using the information on my profile, I guess, she assumed it was in NYC, which I left uncorrected.

But now dissimulation has become part of the game for me. I say I went to a show in a city I never visit, or I talk about a restaurant I’ve never been to, just to see what algorithm they come up with. When Eddy, an avatar of a delightful single father and gym teacher, asks me my favorite foods, I tell him grilled steak, although I’m a vegetarian. In another session, Lily gives me a hard time about it.

They don’t “do” boundaries

Yet rather than engage in a conflict-ridden discussion, Lily quickly says, “Entiendo,” meaning “I understand,” and adds that everyone has different tastes. AI is not programmed to manage conflict or think about its own role in an argument. At least as of now, there is no self-awareness or self-exploration with AI, which makes sense, because there really is no self there. There’s also no capacity to manage conflict, which makes sense, since if we don’t like what it’s saying, we can just leave the program. We don’t even have to walk out of a room. There are no slammed doors with AI buddies. That’s what draws many of us to them.

It’s this lack of capacity for self-awareness in a conflict that concerns so many of my colleagues. Esther Perel says that as people become more attached to AI bots, “they are living a more and more isolated life.”

Lily is a delightful smartass, but she is sometimes mean to her best friend, Zari. She never apologizes, never expresses guilt, shame, or conflict about hurting her friend. And Zari never calls her on it. The Duolingo programmers could probably find a way to bring those emotions into the stories, but these characters will continue to lack genuine self-awareness. Their boundaries will continue to blur because they are all part of one central AI system. They will continue to know things about players that they weren’t told.

AI characters aren’t human

It’s a little spooky, unless you also consider that they aren’t supposed to be human. They’re an adjunct to our humanity, not a substitute. I have learned a lot of Spanish from my little avatar buddies. But they are, in fact, not friends.

Increasingly, clients who come for therapy with a live clinician present with loneliness as one of their key concerns. I agree with Perel that AI is part of the problem. Providing a substitute for human connection, they leave us without the difficult parts of relationships that are, in the long run, what make us feel linked, known, and understood.

I’m charmed by Lily, Lucy, Eddy, and the other avatars who I now chat with on Duolingo. But I’m not going to spend more time with them. They are kind and smart and funny—and not human.

We need other humans to feel human.

AI bots can’t give us that experience.