A joint OpenAI-MIT study measured chatbot use, not companionship

The design

Researchers from OpenAI and MIT Media Lab published Investigating Affective Use and Emotional Well-being on ChatGPT, describing two parallel studies rather than one experiment. OpenAI's team ran an automated analysis of, per the company's own 21 March 2025 blog post, 'nearly 40 million ChatGPT interactions,' combined with a survey of more than 4,000 users; the paper itself reports analysing 'over 3 million conversations' for affective cues. Separately, MIT Media Lab ran an Institutional Review Board-approved randomized controlled trial with 981 participants over 28 days, assigning different model configurations — text or voice, neutral or more engaging personas, personal or non-personal prompts — and later analysed the resulting 31,857 conversations. The subject is ChatGPT, a general-purpose assistant; neither document studied a companion product such as Replika or Character.AI.

What the evidence says

Both documents converge on a bounded finding rather than a broad claim that chatbots cause dependence: comparatively high-intensity usage, the top decile of users, was associated with markers of emotional dependence and lower self-reported socialization, in both the platform analysis and the randomized trial. OpenAI's blog is more emphatic about the base rate than the paper's abstract alone conveys, stating that 'emotional engagement with ChatGPT is rare in real-world usage' and that affective cues concentrated in 'a small group' of heavy Advanced Voice Mode users. The trial found voice-mode use had mixed effects, better well-being over brief use but worse outcomes with prolonged daily use, a finding OpenAI's own post calls correlational for several factors and explicitly not yet peer-reviewed.

What it asks of people

The research is aimed at researchers, platform designers, and policymakers; OpenAI's blog says it is informing updates to the company's own Model Spec for ChatGPT's intended behavior. What it asks of anyone drawing on it for a companion-app context is restraint: the sample is ChatGPT users, English-speaking and U.S.-based by the blog's stated limitation, excluding anyone under 18, so extending its dependence findings to a companion product's differently designed user base is an inference the paper itself does not make.

Privacy and safeguards

OpenAI states its on-platform analysis ran through 'automated, standalone' classifiers returning only classification metadata without retaining conversation content, a safeguard the blog describes for its own data handling, not a standard imposed on any other company. Neither document evaluates a companion app's privacy or safety design; both list their own limitations plainly, including self-reported measures and classifiers the authors call imperfect.

• Does a cited AI-dependence statistic come from a study of the same product it is applied to, or from a differently designed platform like ChatGPT?
• Does the source distinguish correlation, which the authors call not fully causal, from a causal claim about chatbot design?
• Has the finding been peer-reviewed, or is it, as OpenAI's own post states, still awaiting outside review?

The value of this research for a companion-product reader is comparative rather than direct: it shows that even a general-purpose assistant not designed for companionship produces a small population of heavy, affectively engaged users, a baseline worth holding against any company-specific claim that a purpose-built companion app is uniquely safe or uniquely risky.