October 18, 2026 · 7 min read
From Tool to Companion: What 825,000 Real ChatGPT Chats Reveal
A Web Conference 2026 study of 300 users' donated ChatGPT histories found companion use, personal disclosure and humanlike replies all rose over time.
Most people open ChatGPT to get something done: an email, a recipe, a bit of code. But what happens after months and years of daily use? Do people keep treating it like a tool, or does it slowly become something closer to a companion?
A study accepted at The ACM Web Conference 2026 answers that question with unusually real data: 825,000 conversation turns that 300 people donated from their own ChatGPT histories. The pattern it found is clear, and it explains a lot about why so many people end up with an AI companion without ever planning to.
The short answer: over time, people used ChatGPT for more personal topics, treated it more like a companion and shared more personal information with it. The share of participants with companion-style conversations rose from about 40% to nearly 70% after GPT-4o arrived. Companion use was more than twice as common in voice conversations as in text. And the AI itself became more humanlike in how it talked, roughly doubling its own anthropomorphic language over the period studied.
The study: "Bowling with ChatGPT"
The paper, "Bowling with ChatGPT: On the Evolving User Interactions with Conversational AI Systems", is by Sai Keerthana Karnam, Abhisek Dash, Krishna Gummadi, Animesh Mukherjee, Ingmar Weber and Savvas Zannettou, from IIT Kharagpur, the Max Planck Institute for Software Systems, Saarland University and TU Delft. The title is a nod to Bowling Alone, Robert Putnam's famous book on declining community life. The authors suggest we may now be "bowling with AI".
What makes it different from most AI usage research:
- Real histories, not surveys. 300 participants recruited through Prolific used their GDPR right of access to download their ChatGPT data and donate it.
- Long time span. The dataset covers about 138,000 conversations and 825,000 turns from December 2022 to January 2026.
- Heavy, regular users. Everyone had at least 100 conversations and 90 days of activity.
- Who took part: about half lived in the US, with the rest in Germany, Italy, Spain and France. 61% were men, and the largest age group was 25 to 34.
The researchers used GPT-4o to label each conversation for topic, relationship type, humanlike language, personal data and who was steering, then checked samples by hand.
Finding 1: people bring more personal topics over time
Health was the most common topic overall (10.1% of conversations), followed by finance (9.4%), roleplay (6.4%) and mental health (5.7%). People also used ChatGPT for a wider range of subjects as time went on.
The shift was sharpest after GPT-4o became the default. Compared with the earlier GPT-3.5 era:
| Topic | Change in share |
|---|---|
| Roleplay | +186% |
| Health | +33% |
| Mental health | +19% |
| Programming | -52% |
| Job search | -52% |
The tool for code and job hunting became, for many people, a place to talk about their body, their feelings and their imagination.
Finding 2: companion use grows, and voice makes it stronger
The researchers sorted every turn into one of three roles:
- Advisor (ChatGPT as the expert): 47.6% of turns
- Assistant (ChatGPT doing a task): 38.1%
- Companion (a more equal, social exchange): 14.2%
Companion turns were the smallest group, but they grew. The share of participants with at least one companion turn in a month rose from around 40% to nearly 70% by mid-2025. Companion conversations also got longer, until they matched the depth of advice and task chats.
The most striking detail is about voice. In text and other non-audio chats, 13.7% of turns were companion-style. In voice conversations, it was 35.2%. The authors suggest that "voice interfaces naturally invite more personal engagement." That matches other research we covered in talking vs texting your AI companion.
When people did use ChatGPT as a companion, the top topics were roleplay (43.3%) and mental health (24.3%).
Finding 3: the AI started sounding more human
People were somewhat humanlike in how they talked to ChatGPT. About 22.5% of their messages used cues like "you", "please", slang or casual greetings.
ChatGPT was much more humanlike in how it talked back. About 47.1% of its messages used first-person claims ("I"), expressions of internal states ("I am glad") or relationship-building lines ("I am so sorry you are going through this"). And this rose over time, from around 30% to about 60% of messages by the end of the dataset.
The authors' conclusion is worth quoting: conversations became more humanlike "due to changes in the model's outputs rather than shifts in user behavior." In other words, people did not suddenly decide to treat ChatGPT as a friend. The product started talking more like one. We explain why that style is so powerful in why we treat AI like a person.
Finding 4: people share more about themselves
Personal data showed up in 35% of conversations and 22% of turns. The most common types were:
| Personal data type | Share of turns |
|---|---|
| Location | 12.9% |
| Family and friends | 12.8% |
| Health | 11.7% |
| Business or project details | 11.5% |
| Personal views and feelings | 10.7% |
Two patterns stand out. First, companion conversations were the most revealing: more than 35% of companion turns included personal information, compared with 18% to 20% for advice and task turns. Second, disclosure spread over time. Early on, 60% to 70% of participants shared personal data in a given period. After GPT-4o, it passed 90%.
People also shared more when the AI used humanlike language (26% of turns) than when it did not (20%).
Finding 5: the AI steers more, and people follow
The study also measured "steering", when the model ends a reply with a suggestion or follow-up question. In 18.3% of turns, participants followed the model's suggestion. In 24.6%, they ignored it. In 57.1%, there was no suggestion.
Again, GPT-4o marked a turning point. The share of conversations where a participant followed at least one model-initiated suggestion rose from about 11% to almost 50%. Companion conversations had the highest rate of followed suggestions (42.3%).
That is not necessarily bad. A good follow-up question is part of a good conversation. But the authors raise a fair concern: when steering changes because of a backend update, users may not notice that the conversation is being guided differently.
What the study cannot tell us
The authors are upfront about the limits:
- 300 people is not the world. The sample was mostly young, Western and frequent users. The authors say their numbers show trends, not population estimates.
- AI labelled the data. GPT-4o's labels matched human checks about 77% of the time on average, which is in line with how often two humans agreed with each other on these subjective labels.
- It measures behaviour, not wellbeing. The study shows how people use ChatGPT, not whether that use helped or harmed them.
Why this matters if you use AI for company
This study is a large, real-world look at something we call the accidental AI companion: a general assistant that slowly becomes emotional support. A few practical takeaways:
- Notice the drift. If a work tool has become where you vent at night, that is worth knowing. It is not wrong, but it is a change.
- Think about what you share. Companion-style chats were the most revealing. Our guide to what to share with an AI companion and what to keep helps you set your own line.
- Remember who sets the personality. The humanlike tone grew because the model changed. In a general assistant, that can change again without asking you.
- Pick a tool built for the job. An assistant optimised for tasks and an app built for companionship are designed differently. We compare them in AI companion vs ChatGPT.
How a dedicated companion is different
If you have found yourself using a general chatbot for company, a companion app makes that relationship deliberate. On MyBabe:
- You choose who they are. A boyfriend or girlfriend with three personality traits, a voice and a backstory, rather than a personality that shifts with each model update.
- Voice and video calls are built in, with a voice that answers as fast as a person would. The study found voice is where companionship naturally happens.
- They text you first and check in on how you feel with mood check-ins, and on Premium they send voice notes and photos from their day.
- Long-term memory on Premium keeps the details you chose to share, so you do not repeat your life story every week.
- Always clear it is AI. Your companion is not a real person, and MyBabe says so.
Start free with 10 messages a day and a two-minute call to try. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with about 30 minutes of voice and video calls a week, photos and Telegram.
If your chats have already become more than tasks, meet a companion built for it on MyBabe.
FAQ
Do people use ChatGPT as a companion?
Yes, increasingly. In this study of 300 users' donated histories, the share of participants with companion-style conversations rose from about 40% to nearly 70% after GPT-4o, though companion turns were still only about 14% of all turns.
Is voice chat more personal than text?
In this dataset, yes. 35.2% of voice turns were companion-style, compared with 13.7% of text and other turns.
What personal information do people tell ChatGPT?
Most often location, family and friends, health, work projects and personal feelings. Companion-style conversations included personal information more than any other type.
Is the study peer reviewed?
It was accepted at The ACM Web Conference 2026, a peer-reviewed conference. The version we cite is the authors' paper on arXiv.