OpenAI NewsMar 21, 2025, 18:00 UTC+8

Early methods for studying affective use and emotional well-being on ChatGPT

OpenAI and MIT Media Lab studied affective use on ChatGPT with two tracks: nearly 40 million interactions in an observational analysis and a 4-week RCT with nearly 1,000 participants. The post says emotional engagement is rare overall and concentrated in a small subset of heavy Advanced Voice Mode users; the provided body does not fully disclose all quantitative well-being results. Watch subgroup effects, not platform averages.

Read the original ↗Export Markdown

Take

OpenAI is doing pre-emptive containment here: nearly 40 million ChatGPT interactions plus a four-week RCT with nearly 1,000 people turn “affective use” into something measurable. That matters more than the headline finding, because it admits Advanced Voice Mode can push a small group of heavy users into more intimate interaction patterns.

I don’t buy the comfort line that emotional engagement is rare overall. Social-product harm is not priced by averages; it sits in the tail: lonely users, long sessions, anthropomorphic voice, repeated companionship. Replika already showed that emotional attachment can become a retention loop even when it is not the official product pitch. The post gives sample sizes and methods, but the provided body does not fully expose the key quantitative well-being results. Without subgroup tables, this reads more like the opening move in risk governance than a settled safety answer.

Source: OpenAI News · openai.com
Published: Mar 21, 2025, 18:00 UTC+8 · Mar 21, 2025