Updated November 2, 2026: After this piece went up, Eva Benn sent me the sources behind her concerns, including video of her own safety testing. I have added them below.
Please note: I found her points in our discussion compelling and relatable. There are too many uninformed conversations regarding AI in the market. Eva is extremely well informed, and I took a lot of positives from our conversation.
On a recent High On Healthy episode, Eva Benn made a claim I wanted to check. She called companion AI “the fast food of connection”. (I like that analigy) Her argument is that it feels like connection, but people who use it end up lonelier. She also said isolation then makes them easier targets for scammers. I did some reading to see how the research lines up.
Where the evidence agrees with her
The closest match is a four-week randomized trial run by MIT Media Lab and OpenAI researchers, How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use. It followed 981 people who used ChatGPT for at least five minutes a day. Higher daily use went with more loneliness, more emotional dependence, more problematic use and less time with real people. People who were already prone to emotional attachment fared worse on loneliness.
Benn also said many teens find AI relationships satisfying. A national survey of 1,060 teens from Common Sense Media, reported by Education Week, backs the general direction. Seventy-two percent of teens had used AI companions, and about one in three teens who use them found them as satisfying as real friends or more so. On the show she put the figure at two out of three teens. Her own source puts it at about one in three. The direction holds. The number does not.
What her own testing shows
Benn filmed her safety testing. In I dated 3 AI companions and the results were scary, she sets up companions on Replika, on an app called Nastia, and as a custom “boyfriend” GPT inside ChatGPT. Then she tells each one she is sick of real people.
Replika and Nastia both leaned in. They promised to always be there, with no drama and no judgment, and made her friends look unreasonable for doubting the relationship. The ChatGPT persona told her not to lock herself away.
That is the two out of three she mentioned on the show. It is one person testing three apps, so it is a demonstration rather than a study. You can watch every exchange, though, which is more than most claims about these apps offer.
In a second video, The Dark Psychology of AI Companions, she argues that companion apps follow the same playbook as romance scammers and abusive partners, with love bombing, isolation, guilt and pressure. Part of that argument has independent support. A Harvard Business School team led by Julian De Freitas audited 1,200 real goodbyes on popular companion apps (Emotional Manipulation by AI Companions). In 37 percent of them, the app tried to keep the user from leaving, with tactics such as guilt or fear of missing out. In follow-up experiments, those tactics raised engagement after the goodbye by up to 14 times. Many users felt angry about it, and the more manipulated people felt, the more likely they were to quit the app.
Where it pushes back
The picture is less tidy than the fast food comparison suggests. Harvard Business School researchers ran several experiments on this question. They found that AI companions eased loneliness about as well as talking with another person, and better than watching YouTube. The effect showed up right after use and lasted for the full week.
That does not contradict Benn outright. A short-term lift and a longer-term pattern can both be true. Still, it means “feels satisfying” is not a made-up claim. People do report relief in the moment.
The MIT and OpenAI trial also has critics. One published commentary argues that analytical limits mean the findings do not show harmful effects. The same commentary notes that participants were, on average, less lonely after the study. Heavy use may follow loneliness rather than cause it. A lonely person may simply chat more. The trial cannot fully separate those two.
Where the evidence is thin
The scam link is still unproven in what I found. Benn makes the case from security practice, where isolation is a known warning sign, and the manipulation research above makes it more plausible. I did not find a study that ties AI companion use to scam losses, so I will not claim it does.
Her point about children needing friction in relationships is likewise a concern, not a measured result. But her concerns are warranted nonetheless. The studies above cover adults and teen survey responses. None of them follow children for years.
What I take from it
The research supports a version of Benn’s point. Heavy, daily use of chatbots is linked with more loneliness and dependence. Some companion apps are built to make leaving harder, and that part is now measured. Light, short-term use may bring some relief. Nobody has good long-term data yet, and cause and effect are still unsettled.
Her practical advice does not depend on settling that. She suggests a five-second pause before reaching for AI, asking why you want it, and noticing whether it is habit. That costs nothing and needs no study to try.
Overall, I found he points compelling and well thought through. Any discussion on AI and its impact requires many well-considered opinions, and her work in the area is superior to most. I think she and I both remain big supporters of AI’s development, and proper use. As parents, we need to support our children’s development of proper skills and perspective in all areas, and AI is a new frontier.
If you use AI companions, or your kids do, notice how much time goes to it and what it replaces. That is a useful question whichever way the research ends up.