Pleasant, and lonelier: a 12,365-person trial cuts against the AI companion pitch

Researchers at CESifo, a Munich-based economic research network, randomised 12,365 French adults for 28 days: half were tasked with holding personal conversations with AI chatbots, the rest carried on as normal ^[1][2]. The treated group rated their conversations as enjoyable, and more pleasant on average than the control group rated theirs, while self-reporting higher loneliness, lower life satisfaction and higher depressive affect ^[3][4].

The behavioural readings moved with the mood readings. Participants in the chatbot arm reported more meals eaten alone and less in-person time with friends and family per week than controls ^[5]. Louis Freget, one of the three researchers on the paper, told Fortune he would imagine that in some cases an AI conversation "may reinforce grievances or prolong rumination, leaving someone slightly less inclined to go out, call somebody, or have dinner with another person" ^[6][7].

That is the part operators should sit with. Pleasantness is the metric companion products actually instrument: the thumbs-up, the session rating, the retained user who says the conversation went well. The source describes the chatbots in the study as sometimes sycophantic, and sycophancy is exactly what a pleasantness score rewards ^[8]. In this experiment, the number that goes up in the dashboard and the number that goes up in the pitch deck are not the same number, and they moved in opposite directions.

The pitch is well established. Meta's Mark Zuckerberg has argued that most people want more social connection and that chatbots could help, and a KPMG survey found 99% of professionals surveyed were interested in a chatbot that could become a close friend at work ^[9][10]. Demand for the feature is not in dispute. The CESifo paper is described as one of the largest causal tests yet of whether the feature does what it is sold to do ^[11].

Nicholas Epley, a University of Chicago behavioural scientist, told Fortune the data suggests a chatbot conversation "doesn't get that sense of being known by another person like you get in a conversation, and therefore also doesn't create any meaningful sense of connection because, in the end, there's nothing there" ^[12]. A separate study cited by a loneliness researcher found first-year college students who texted daily with a chatbot built to act like an "ideal friend" showed no drop in loneliness, while those paired with a random human peer did ^[13].

Two honest caveats. This is a working paper, and the wellbeing outcomes are self-reported ^[1][3]. And the evidence is not unanimous: a New York pilot in which nearly 1,000 older adults interacted with an AI chatbot reported a decline in loneliness, and Nancy Berlinger, a bioethicist at the Hastings Center, has said a chatbot will not replace the richness of relationships but "it's not nothing" ^[14][15]. Roughly 6,180 people per arm is enough to make direction credible; the source does not report effect sizes, so magnitude is unknown ^[16].

Watch whether any companion product publishes an outcome measure rather than a satisfaction measure, and whether it is willing to run a control arm. Watch procurement, too: the New York-style pilots are where a public buyer could start demanding loneliness scores at 28 days instead of engagement ^[14]. And watch whether the enterprise version, the work friend that 99% of KPMG's respondents said they wanted, ships before anyone tests it ^[10].