A study from researchers affiliated with Harvard Business School has become one of the more cited pieces of evidence in discussions of AI companion effectiveness. The research, led by Julian De Freitas and colleagues, investigated why people report feeling genuinely heard during interactions with AI. The findings are counterintuitive enough to be worth examining in detail, and the methodology is rigorous enough that the results deserve serious attention.

What the Study Set Out to Investigate

De Freitas and his team started from an observation that had accumulated anecdotally across multiple platforms: users were reporting that conversations with AI felt more heard than many conversations with humans. This was not a minority finding. It appeared with enough regularity that it warranted systematic investigation. The researchers designed a series of experiments to test whether this was genuine, what was driving it, and whether it depended on users knowing or not knowing they were talking to an AI. The study involved several thousand participants across multiple conditions, with careful attention to controlling for novelty effects and social desirability bias.

The Core Methodology

Participants were assigned to disclose a personally meaningful problem or concern in one of several conditions: to a human listener, to an AI they knew was an AI, or in some conditions to a confederate whose status was ambiguous. The conversations were analyzed for several outcome variables, including felt understanding, likelihood of further disclosure, and reported emotional relief. Critically, the team also measured what participants believed was driving the feeling of being heard. They wanted to know whether users attributed the feeling to the AI's responses themselves or to something else, such as the act of articulating the problem in the first place. This distinction matters because it affects whether the AI is actually providing something or merely serving as a journal.

What the Data Showed

The core finding was that participants in the AI condition reported feeling heard at rates comparable to, and in some categories higher than, the human listener condition. This held even when participants were fully informed they were talking to an AI. The feeling of being heard was not primarily a product of believing the listener was human. Analysis of what drove this effect pointed to several features of AI responses: the absence of interruption, the consistent use of reflective language, the lack of visible judgment cues, and the AI's ability to hold the emotional content of what was shared without redirecting to the listener's own experience. Human listeners, even well-intentioned ones, frequently violated these conditions inadvertently. They interrupted, they related, they reassured in ways that subtly shifted focus. The study also found that users who felt heard by the AI were more likely to subsequently reach out to a human contact about the same issue within a set follow-up window. The AI was functioning as an intermediary step rather than a terminal one.

Limitations the Authors Acknowledged

De Freitas and colleagues were careful to note several constraints on the findings. The participant pool skewed toward people who were already comfortable with technology, which limits how broadly the results generalize. The conversations were conducted in structured research contexts, which may not fully capture how interactions unfold in real-world long-term use. The follow-up windows were relatively short, leaving open questions about sustained effects. The authors also noted that felt understanding is not identical to actual understanding, a distinction that matters for how the findings should be interpreted. The AI was producing the felt experience of being heard. Whether that maps onto the kind of deep mutual understanding that sustains long-term human relationships was explicitly outside the scope of what the study measured.

Why This Research Matters

A tangent worth noting: the finding that human listeners frequently disrupt the felt experience of being heard is perhaps as significant as any finding about AI. The research implies that something many people experience as a universal feature of supportive conversation, the sense that a good listener naturally produces the felt experience of being heard, is actually quite fragile and frequently fails. AI companions, by producing this experience reliably, are not just offering a substitute. They are offering something that many users are not reliably getting elsewhere. That reframes the question from whether AI can replicate human connection to whether AI is filling a gap that human connection is already leaving open. The De Freitas data suggests the answer to the second question is yes.