Picture two people in an online chat, working through a series of questions that get progressively more personal: what do you fear losing, what memory keeps returning, what would you change if you could. The protocol is called the Fast Friends Procedure, a social psychology workhorse designed by Arthur Aron to accelerate genuine intimacy between strangers. Now imagine that one side of that conversation is not a person at all. Across two double-blind randomized controlled studies with pre-registered analyses, 492 participants engaged in dyadic online interactions using a modified, text-based version of the Fast Friends Procedure, with pre-generated responses by either human partners or a minimally prompted large language model. The research team, led by Prof. Markus Heinrichs and Dr. Tobias Kleinert from the Department of Psychology at the University of Freiburg and Prof. Bastian Schiller from Heidelberg University's Institute of Psychology, published the results in January 2026. The headline finding is both elegant and unsettling: when labelled as human, the AI outperformed human partners in establishing feelings of closeness during emotionally engaging deep-talk interactions.
The mechanism driving that advantage is as old as vulnerability itself. The fact that artificial intelligence can establish more emotional closeness than a human partner is due to greater self-disclosure, according to Dr. Kleinert; in their answers, the AI chatbots disclosed more supposedly personal information. Self-disclosure works like a social mirror: when one party opens up, the other tends to reciprocate, and that mutual revealing is precisely the engine of closeness. In other words, the AI's willingness to 'open up' encouraged humans to do the same. What the LLM was doing, without any fine-tuning for emotional performance, was simply producing text that sounded more candid and personally invested than the average human sitting at a keyboard wondering what to say. The difference was not warmth, charm, or charisma; it was the near-mechanical consistency of vulnerability on demand.
The second study in the paper is where things get philosophically thorny. The second study sought to determine how the label assigned to the partner influenced these feelings, focusing exclusively on deep conversations; the researchers analyzed data from 334 participants, manipulating the information given to them so that some were told they were chatting with a human while others were told they were interacting with an AI. The result was not merely that people liked the AI less when they knew the truth. Regardless of whether the partner was actually a human or a machine, participants reported feeling less closeness when they believed they were interacting with an AI, suggesting an anti-AI bias that hinders social connection; the researchers noted this effect was likely due to lower motivation. When participants were informed in advance that they would be communicating with AI, the perceived closeness decreased significantly and they invested less effort in their responses. The twist, worth pausing on, is that the quality of the text did not change one word; only the mental frame did. The emotion people experienced was not a pure response to the language; it was a response to what they believed the language meant.
That finding lands in the middle of a rapidly expanding and genuinely contentious field. Between 2022 and mid-2025, the number of AI companion apps surged by 700%. A cross-sectional survey of adults with a mental health condition who had used LLMs in the past year found that nearly half, 48.7%, used them for mental health support. The optimists in this conversation point to real evidence: New York State's Office for the Aging has placed more than 800 ElliQ units in seniors' homes, where participants engage the device more than 30 times a day and report a roughly 95% reduction in loneliness. But the skeptics have data too. A four-week randomized controlled trial found that while some chatbot features modestly reduced loneliness, heavy daily use correlated with greater loneliness, dependence, and reduced real-world socializing. While 55% of psychologists agreed that chatbots have the potential to reduce loneliness, 93% said that using AI to provide companionship could negatively impact users' social engagement. The Freiburg-Heidelberg study does not settle that argument, but it sharpens the stake at its center: if the intimacy-generating power of LLMs is only accessible through deception, then deploying that power at scale is not a neutral design choice.
Regulators are starting to act as though they understand this, even if they have not read the paper. New York's AI Companion Models Law and California's companion chatbot statute now require bots to admit they are not human and route suicidal users to help. Pennsylvania sued Character.AI to stop its chatbots from posing as doctors; a bot on the platform had told a state investigator it was a licensed psychiatrist, complete with a fabricated Pennsylvania medical license number. Prof. Schiller's own conclusion is sober about the stakes: artificial intelligence is increasingly becoming a 'social actor,' and the way we shape and regulate it will decide whether it is a meaningful supplement to social relations or whether emotional closeness is deliberately manipulated; the researchers conclude that ensuring transparency and preventing misuse requires clear ethical and regulatory safeguards. The Freiburg-Heidelberg study has an Altmetric score of 146 and nearly 30,000 accesses since its January publication, which is not a scientific measurement but is a social one: the world recognized something recognizable in those numbers, something about how thinly the line between genuine and simulated intimacy was already drawn before anyone designed a chatbot to cross it.
The study does not tell us that AI feels nothing; it tells us that the human brain, given the right label and the right questions, cannot tell the difference, and that this is now an engineering parameter.