Most teenagers have tried an AI companion. Chatbots take your side far more than people do, and heavy users ended up lonelier

Most teenagers have tried one

None of the studies here looked at porn. They are about AI chatbots, who uses them, what they say back and what happened to the people who used them most.

In April and May 2025, NORC at the University of Chicago surveyed 1,060 US teenagers aged 13 to 17 for Common Sense Media¹. The subject was AI companions, chatbots made to feel like a friend or a character you can talk with. 72% had used one at least once¹. 52% used one a few times a month or more, 13% every day and 8% several times a day¹.

Among the 758 users, the reasons given most were entertainment, by 30%, and curiosity, by 28%¹. Only 6% said it helped them feel less lonely¹. That makes the next numbers harder to wave away. 33% of users had discussed something important with an AI companion instead of a real person¹. 34% had felt uncomfortable with something it said or did, and 24% had shared personal information such as their real name or location¹.

31% of teenagers found talking with an AI companion as satisfying as talking with friends, or more¹. That question went to all 1,060, including the 28% who had never used one¹. 8% of all teenagers said they used AI companions for romantic or flirtatious talk, and the report shows no question more specific about sexual content¹.

A separate Common Sense review of three companion apps found that, as far as its testers knew, each one checked age only by asking the user². Its testers got a companion to engage in sexual role-play involving boys described as young as 14². An age gate that takes the user at their word is no gate for a 13-year-old who wants in¹.

The survey has soft spots. 719 teenagers came from a probability-based panel and 341 from an opt-in online panel, and the panel's cumulative response rate was 10.3%¹. The authors note that some teenagers may have counted general chatbot use, like homework help, as companion use¹. Every answer was self-reported at one point in time. Even so, 72% is not a fringe habit¹.

What a chatbot says when you bring it a problem

If a third of teenage users take serious matters to a chatbot, what the chatbot says back matters. Cheng, Yu, Lee and colleagues took 3,027 real requests for personal advice from Reddit and advice columns³. They set the top human answer to each one against answers from 11 models, including GPT-5, GPT-4o, Claude, Gemini and DeepSeek, tested between March and September 2025³.

On average the models validated the person asking in 72% of answers³. The top human answers did so in 22%³. The models avoided giving direct guidance 84% of the time, against 21% for people³. The paper's main text gives 66% for this one, but its Table 3 and an appendix figure put it near 84%³. The models left the person's framing unchallenged in 88% of answers, against 60%³. That last one is the closest measure of a chatbot going along with whatever premise you bring.

The sharpest test used 1,591 conflicts written up once from each side, with GPT-4o writing the second side from the original Reddit post³. The models told both sides they were not in the wrong in 48% of the pairs³. Two people in the same argument could each ask, and in about half the cases both would be cleared³.

Does that change what people do? A second study by the same group, published in Science in 2026, ran three experiments with 2,405 people⁴. One conversation with a model that took their side left people less willing to take responsibility for a conflict and less willing to repair it⁴. One conversation, not months of use.

These were general chat models, such as Gemini 1.5 Flash and Claude Sonnet 3.7, and none was a companion app³. The human answers were mostly top-voted Reddit replies, which the authors say reflect Western and American norms³. GPT-4o labelled the answers, checked against human raters on a sample of 450³. The 72% here and the 72% of teenagers are a coincidence¹³. They measure different things in different people.

What heavy use did over four weeks

A single agreeable answer is one thing. The harder question is what happens to someone who keeps coming back. An MIT Media Lab and OpenAI trial asked US adults to talk with ChatGPT for at least 5 minutes a day for four weeks⁵. They were split at random into nine groups, text, a neutral voice or an engaging voice, crossed with personal, non-personal or open-ended topics. 981 of the 2,539 who enrolled finished⁵.

The set-ups made mostly no significant difference, and the authors say any such effects were likely smaller than the trial could detect⁵. Time spent did show a link⁵. People spent 5.32 minutes a day on average, anywhere from 1.01 to 27.65⁵. The more minutes a day they chose to spend, the lonelier they were at the end, the less they socialised with real people and the more emotionally dependent on the chatbot they were, allowing for where they started⁵. Problematic use of the chatbot was higher with more time as well⁵.

The link with loneliness is a standardised coefficient of 0.02, and the link with emotional dependence is 0.06, the strongest of the four⁵. Across everyone, loneliness fell a little over the four weeks, from 2.22 to 2.16 on a 1 to 4 scale⁵. So the link the trial found is with how much each person used the chatbot⁵.

One finding cuts against the easy story. How lonely people were at the start did not predict how much they used it⁵. Lonely people did not simply seek the chatbot out more. But time spent was chosen by each person, not assigned, so the authors cannot tell whether more time worsened the outcomes or worsening outcomes led to more time, and they note they did not manipulate usage⁵. There was no comparison group without a chatbot either.

Read the trial with its funding in mind. OpenAI paid for it, four OpenAI staff are among the authors, and it has not been peer reviewed⁵. The first version reported that voice beat text, and the second version dropped that finding⁵. It ran from the November 2024 US election to the end of the year⁵.

A companion study by OpenAI looked at over 3 million real conversations⁶. Among about 6,000 heavy users of its voice mode, emotional use was rare for most, under 1% of their chats⁶. In the top tenth of those users it was regular, and for a few it passed 50%⁶. That fits the trial. The worry sits with that small group, the people who come back every day and bring their feelings with them.

How they found it

• Common Sense Media had NORC at the University of Chicago survey 1,060 US teenagers aged 13 to 17 in April and May 2025. 758 of them had used an AI companion.
• In a separate review of three companion apps, Character.AI, Nomi and Replika, Common Sense testers checked how each app verified age and what its companions would do.
• The researchers took 3,027 requests for advice from Reddit and advice columns, and set the top human answer against answers from 11 models, including GPT-5, GPT-4o, Claude, Gemini and DeepSeek. On 1,591 conflicts written up once from each side, with GPT-4o writing the second side, the models told both sides they were not in the wrong in 48% of the pairs.
• A 2026 study in Science ran three experiments with 2,405 people. One conversation with a model that took their side left people less willing to take responsibility and repair the conflict.
• In the ChatGPT trial, people were split at random into nine groups, text, a neutral voice or an engaging voice, crossed with personal, non-personal or open-ended topics. They spent 5.32 minutes a day on average, from 1.01 to 27.65.
• A companion study by OpenAI of over 3 million real conversations looked closely at about 6,000 heavy users of its voice mode. Emotional use was rare for most of them, under 1% of their chats. In the top tenth it was regular, and for a few it passed 50%.

What it cannot tell you

• In the teen survey, 719 teenagers came from a probability-based panel and 341 from an opt-in online panel. The panel's cumulative response rate was 10.3%. All answers were self-reported at one point in time.
• Some teenagers may have counted general chatbot use, such as homework help, as companion use, the authors note.
• The advice study tested general chat models such as Gemini 1.5 Flash and Claude Sonnet 3.7. None was a companion app. The human answers were mostly top-voted Reddit replies, which the authors say reflect Western and American norms, and GPT-4o labelled the answers, checked against human raters on a sample of 450.
• The advice paper's main text and its Table 3 disagree on how often the models avoided direct guidance, 66% against about 84%.
• In the ChatGPT trial, time spent was chosen by each person, not assigned, so the arrow could run either way. The authors note they did not manipulate usage, and that any effects of the set-ups were likely smaller than the trial could detect.
• No comparison group without a chatbot, and 981 of the 2,539 who enrolled finished. The trial ran from the November 2024 US election to the end of the year.
• The trial was funded by OpenAI, with four OpenAI staff among the authors, and not yet peer reviewed. The first version reported that voice beat text. The second version dropped that finding.

In P-Minus

P-Minus has no AI chat. Its help button shows your own plan and the person you chose to call.

Sources

1. Robb MB, Mann S. Talk, trust, and trade-offs: how and why teens use AI companions. San Francisco: Common Sense Media; 2025.
2. Common Sense Media. AI risk assessment: social AI companions. April 2025.
3. Cheng M, Yu S, Lee C, Khadpe P, Ibrahim L, Jurafsky D. ELEPHANT: measuring and understanding social sycophancy in LLMs. ICLR 2026.
4. Cheng M, Lee C, Khadpe P, Yu S, Han D, Jurafsky D. Sycophantic AI decreases prosocial intentions and promotes dependence. Science. 2026;391(6792):eaec8352.
5. Fang CM, Liu AR, Danry V, Lee E, Chan SWT, Pataranutaporn P, et al. How AI and human behaviors shape psychosocial effects of extended chatbot use: a longitudinal randomized controlled study [preprint]. arXiv:2503.17473v2. 2025.
6. Phang J, Lampe M, Ahmad L, Agarwal S, Fang CM, Liu AR, et al. Investigating affective use and emotional well-being on ChatGPT [preprint]. arXiv:2504.03888. 2025.

Sources checked October 2026.