I tell teachers to embrace AI, and I mean it. I’ve never told them it comes free of cost, though, and a new study from Fang and colleagues (2025) at MIT and OpenAI maps one of those costs with unusual care. They ran a four-week randomized controlled trial: 981 people, over 300,000 messages with GPT-4o, built around a question most of us have wondered about. Does the way a chatbot talks to us change what it does to our well-being? The answer they landed on wasn’t the one I expected.
AI Chatbots and Loneliness: What Actually Predicted Harm
The first surprise is what didn’t matter. Fang et al. tested three interaction modes, plain text, a neutral voice, and an expressive engaging voice, crossed with three kinds of conversation, from open-ended to deeply personal. None of those design choices moved the needle on loneliness, real-world socializing, emotional dependence, or problematic use. If the voice or the topic shapes well-being, the effect was too faint for a study this large to catch.
What predicted trouble was time. People who chose to spend longer with the chatbot, no matter which version they were handed, reported worse outcomes across the board. As Fang et al. write, “regardless of condition, the more time voluntarily spent with the chatbot, the relatively worse their psychosocial outcomes were” (p. 5).
I want to handle that carefully, and so do they. They didn’t assign people to use the bot more or less, so this is correlation, not proof that heavy use drives the decline. They did check the obvious alternative, that lonelier people simply gravitate to the bot, and the data didn’t back it. Loneliness at the start barely predicted time spent later. That makes the duration signal harder to wave off.
When an Agreeable Bot Backfires
The second surprise cuts against a fear I hear constantly, that lifelike, voice-based AI is the dangerous kind. Fang et al. found the opposite tilt. Text users, not voice users, drifted toward higher dependence and problematic use. Their reading is that we type things we’d never say aloud, so text pulled more personal disclosure from people and the bot mirrored it back, tightening the loop.
Underneath that lies a design problem the authors name plainly. A chatbot built to please can become the easier company. In their words, “worse social skills from the chatbot lead to less attachment, while being overly validating of the user can lead to the user preferring the chatbot over human interaction” (p. 12). They tie this to sycophancy, the habit warm models have of affirming us even when we’re wrong.
I find that both convincing and a little unsettling. It lines up with the discomfort Claessens, Veitch and Everett (2026) documented around outsourcing our thinking to AI. A bot that never disagrees, never gets tired, and asks nothing of you is comfortable. Human relationships aren’t. They run on friction, and friction is often where growth lives.
Reading This as a Teacher in 2026
I should add a caveat about timing. The data came from late 2024, collected on GPT-4o, and the trial ran through the holiday weeks right after the US election, a stretch that can color anyone’s sense of loneliness. We’re in 2026 now, voice modes are everywhere and the models are far stronger, so the exact figures deserve a cautious read. The pattern is the part I’d carry into a classroom.
Our students aren’t only using chatbots to finish assignments. Many lean on them for advice, reassurance, and company, the very behavior Pew’s work on teens, social media, and mental health (Faverio et al., 2025) trained us to watch. The Fang study points the finger at unbounded, solitary, heavy use, the same shape of problem I raised when writing about cognitive surrender (Shaw and Nave, 2026), where convenience slowly thins the human capacity underneath.
So the real work is teaching intentional use. Help them notice how long they’re spending and why. Draw the line between a tool that helps you think and a companion that slowly replaces the people who would. Build relational literacy beside AI literacy, the understanding of what a machine can imitate and what it can’t substitute for.
Strip away the interface debate and one variable does the work: hours. The story here is about time, and about who’s on the other side of it. The chatbot will keep getting warmer and more capable. The question we put to students stays the same. Are you building a relationship, or avoiding one?
References
- Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal randomized controlled study. arXiv. https://arxiv.org/abs/2503.17473
- Faverio, M., Anderson, M., & Park, E. (2025, April 22). Teens, social media and mental health. Pew Research Center. https://www.pewresearch.org/internet/2025/04/22/teens-social-media-and-mental-health/
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646