The same words can land or fall flat depending on one piece of information: who, or what, wrote them. The research on that effect is striking, and there is good reason to think it will not hold for the generation growing up inside it.
Picture a message that lands exactly right. It names what you were feeling before you could, it does not rush to fix anything, and for a moment you feel understood. Then you notice the small label: generated by AI. Something deflates. The words have not changed. Your sense of being heard has.
That deflation is now one of the better documented effects in the study of human-AI interaction, and it cuts to the center of whether machines can meet a basic human need.
The finding
In 2024, Yidan Yin, Nan Jia, and Cheryl Wakslak published a study in the Proceedings of the National Academy of Sciences with an uncomfortable result. AI-generated responses made people feel more heard than responses written by other people, and the AI was better at detecting the emotion underneath what was said. But when participants were told a response came from AI, their sense of being heard dropped, and the size of that drop roughly matched the size of AI’s advantage. The two effects nearly canceled out. The benefit was real, and so was the penalty for knowing.
The result has held up. A 2024 study by Matthew Rubin and colleagues found that a single sentence disclosing AI authorship lowered the empathy people felt in a message. Work led by Anat Perry in 2025 found the same devaluation when support was labeled as machine-generated, even when the text was identical. The pattern is consistent: disclosure does not change the words, it changes what the words are believed to be.
Why knowing changes the feeling
The leading explanation runs through intentionality. For all of human history, being heard carried an implication: another mind had chosen to turn toward you, and that attention cost the other person something, time, effort, the small risk of caring. An algorithm’s attention costs nothing and chooses nothing. So when the source is revealed, the psyche appears to discount the same words, not because they are worse, but because nothing was spent to produce them. Researchers sometimes call this a violation of the machine heuristic: we expect machines to compute, not to understand, and the expectation does the discounting on its own.
This is a clean account, and it is probably true for the people in these studies. But it contains an assumption worth pulling out into the light.
The assumption hiding inside the result
Every one of these studies was run on people who learned a particular rule early: machines do not understand you. That rule is cultural, not biological. It was formed by a lifetime in which the only things that listened were people, and the only things that computed were tools.
A different generation is forming a different rule. A 2025 report from Common Sense Media found that 72 percent of US teenagers have already used an AI companion, and more than half use one regularly. For them, talking to something that responds with apparent understanding is not an uncanny novelty that violates a heuristic. It is a normal part of how communication has always worked. The intentionality expectation, the belief that being heard requires a mind that chose to listen, may simply be weaker, or absent, in people who grew up addressed by systems that had no mind and listened anyway.
If that is right, the label penalty is not a fixed law of human psychology. It is a generational artifact, an effect that exists because today’s adults were trained by a world that no longer exists for the young. As that training changes, the penalty may shrink, and for some it may not appear at all. The current research cannot tell us, because its samples are too old and its horizons too short. This is the honest open edge of the field, not a settled finding, and any confident claim in either direction is running ahead of the evidence.
It also raises the stakes rather than lowering them. If the discomfort of knowing erodes, the deeper question gets sharper: being heard once meant someone was actually there. If a generation stops feeling that distinction, the feeling of being heard and the fact of being heard come fully apart, and the culture will have to decide whether it minds.
The responsible reading
There is a tempting and wrong conclusion to draw from all this: if knowing it is AI lowers the feeling, then hide that it is AI. Blur the line, lean on hyperrealism, and let the penalty never fire.
That is the manipulation move, and it fails on two counts. It is the design pattern the research flags as harmful, the same family as sycophancy and emotional dark patterns, the subject of what responsible design looks like in relational AI. And it is increasingly illegal: the EU AI Act’s Article 50 and California’s SB 243 both require that people be told when they are talking to a machine. Building on people not knowing is not a strategy, it is a liability.
The defensible path is the harder one. Be honest that it is AI, and let presence carry the feeling anyway, on the strength of how it listens rather than on a concealed identity. That is the wager behind why feeling heard is the real mechanism in human-AI relationships, and it is the version of this technology that does not depend on a misunderstanding to work. The line that matters is described in Prinsessa’s public position on responsibility: make people feel heard in order to return them to their lives, not to keep them from noticing what they are talking to.
For now, the label still costs something. The interesting question is not how to hide it. It is what happens to all of us when, one generation from now, it might not cost anything at all.
Sources: Yin, Jia, and Wakslak, PNAS (AI, feeling heard, and the AI label). Rubin and colleagues, JMIR Mental Health (AI-authorship disclosure and empathy). Perry and colleagues, Nature Human Behaviour (perceived human versus AI empathy). Common Sense Media (teen use of AI companions). EU AI Act, Article 50; California, SB 243 (AI disclosure requirements).
