Anthropic found companionship use under half a percent of chats

A privacy-preserving analysis of 4.5 million Claude conversations found affective use rare, companionship or roleplay under 0.5 percent.

Historical event

June 27, 2025

First source published

June 27, 2025

Site publication

September 18, 2026

What happened

On 27 June 2025, Anthropic published 'How people use Claude for support, advice, and companionship', describing a study of Claude.ai usage. The company used Clio, its own automated, privacy-preserving classification tool, to analyse roughly 4.5 million conversations from Claude.ai Free and Pro accounts. After excluding content-creation tasks and requiring roleplay exchanges to include at least four human messages, the analysis retained 131,484 conversations classified as affective, meaning conversations motivated by emotional or psychological needs such as advice, coaching, counselling, companionship or romantic roleplay.

What the documents show

The post reports affective conversations made up 2.9 percent of sampled Claude.ai traffic, with companionship and roleplay combined under 0.5 percent and romantic or sexual roleplay specifically under 0.1 percent. Anthropic states this pattern aligns with separately published findings from OpenAI and the MIT Media Lab describing similarly low affective engagement with ChatGPT, though the post reports only its own figures. It also states Claude pushed back, meaning it refused or resisted a request, in under 10 percent of coaching or counselling conversations, usually for safety reasons, and that expressed sentiment tended to shift slightly more positive by a conversation's end.

The mechanism

The finding rests on an automated classifier applied after the fact to real conversations, not a survey of what users say they do or a controlled experiment. Clio's anonymisation and aggregation is built to protect any single conversation's content, but that design also means the analysis cannot follow one person over time or connect a conversation to that person's later wellbeing; it is a snapshot of language patterns, not of outcomes. Anthropic is explicit that Claude.ai was not built as a companionship product, unlike apps built specifically around a persistent persona, so a low measured rate here describes how a general assistant gets used, not a ceiling on tools designed for that purpose.

What it leaves open

Anthropic's stated limitations note some conversations were likely misclassified, that the analysis supports no causal claim about real-world emotional outcomes, that it lacks the longitudinal, user-level data needed to study emotional dependency, and that it covers only text-based use, before voice or video interaction became more prominent. The company notes OpenAI separately found affective topics more common in voice conversations, something this analysis did not examine for Claude. This is an editorial reading: a low prevalence figure for one general-purpose assistant does not establish a rate for the wider category of dedicated companion apps this record set is built around.

• Was the measured rate produced by a validated survey, an automated classifier, or a controlled trial?
• Does a low rate on a general assistant say anything about a product built specifically for companionship?
• What would the same organisation need to measure to study emotional dependency rather than one conversation's language?

The company's own figures support a narrow claim: affective use was uncommon in this snapshot of Claude.ai traffic, and companionship specifically was rarer still. They do not support a claim about how common comparable use is on apps engineered around persona and memory, a distinction the post itself draws.

Sources & reading trail

Full post: the Clio methodology, sample sizes, the 2.9 percent/0.5 percent/0.1 percent figures, pushback rate, sentiment finding, and the stated limitations.

Source published: 27 June 2025 · Retrieved: 16 September 2026

Independently confirms the comparable OpenAI/MIT Media Lab finding of low affective-use rates that Anthropic's post cites as consistent with its own result.

Source published: 21 March 2025 · Retrieved: 16 September 2026

Company documents, filings, studies and official records establish the record; the reading and the questions are Lovebot Journal editorial analysis. This retrospective draft does not imply the site published on the event date.

Continue reading

• A 2025 study pair found heavy chatbot use linked to worse outcomes
• A lab measured how models flatter user beliefs
• OpenAI withdrew a GPT-4o update for being sycophantic
• Browse the complete the archive

Published September 18, 2026, not on the date of the event described.