An AI companion can feel like someone is there even when you know exactly what it is. Part of that feeling has a name, anthropomorphism: attributing human qualities to a nonhuman thing. The question here is how much weight that sense of a person should carry when you judge whether an answer is any good.
You do not have to be fooled
An early experiment came in 1994, when Nass, Steuer, and Tauber published Computers are Social Actors, a paper reporting short laboratory studies with computer-literate students. In the first, 33 participants worked with a computer tutor and then rated its friendliness and competence on the same computer, on a different one, or on paper. Ratings entered on the same computer came out higher than those on paper, while ratings entered on a different computer did not differ significantly from paper ratings. Asked afterward, participants said politeness norms had nothing to do with rating a machine.
The sample was small, the task was brief, and a 1990s lab session is a long way from current companion use. What it suggests is narrow: a social response can appear without any belief that a person is present.
A later attempt did not find the effect. In November 2023, Evelien Heyselaar published a replication of the same-computer politeness effect in Scientific Reports, following the fuller 1999 procedure rather than the 1994 setup. Its 132 participants evaluated a computer on the same desktop or a different one, and the two groups’ average ratings did not differ. The original hardware was unavailable, and the result does not establish that social responses to machines have vanished. Together, the two studies show a possibility under some conditions, not a reflex in every user.
The human cues are part of the design
A 2025 perspective in PNAS by Peter, Riemer, and West, The benefits and dangers of anthropomorphic conversational agents, synthesizes existing work rather than reporting a new experiment. Its argument is that humanlike communication is built into these systems, so the impression of a person is partly a product of design and not simply the user’s error. The authors separate the ability to communicate from human understanding, and they take the upside seriously: conversation can make an interface accessible to more people. The same fluency can persuade, and they count manipulation among the dangers. They expect that telling people to simply resist the impression will work less and less well.
What the voice trial measured
A four-week randomized trial by Fang and colleagues, in the October 2025 revision of their preprint, asked a different question: whether the way a conversation is delivered changes psychosocial outcomes. It assigned 981 participants to conversations with GPT-4o as plain text, a neutral voice, or an engaging voice, crossed with three kinds of conversation task. Participants rated the engaging voice the most humanlike. On two of the outcomes measured, loneliness and socialization, neither the mode nor the task made a significant difference. Heavier voluntary use did go with worse outcomes, and because time spent was not assigned, the study cannot separate cause from effect.
The null result does not show that voice and text are equivalent, that any mode is safe over longer periods or in other products, or that a voice never helps anyone. What the trial leaves without support is the assumption that sounding more human buys a better outcome by itself.
Read the answer without the performance
From here on, the judgment is mine. None of these papers tests whether an answer is more accurate when it comes with a name, a voice, or a remembered detail, so the feeling of presence and the support for an answer have to be weighed separately. Liking a companion’s voice is a fair reason to keep using it. It is not a reason to trust what it tells you about a deadline, about what it keeps from your chats, or about how well it understands you.
Imagine a companion telling you, in a warm and unhurried voice, that you have thirty days to return something you bought. Now imagine the same sentence as plain text. The store’s policy does not change with the delivery, and the policy is what settles whether thirty days is right. Whatever the warmth does to your confidence, the missing evidence is the same in both versions. This is an illustrative comparison, not a claim that a product can replay one answer both ways.
Try checking one piece of advice a companion gave you recently, with little riding on it. This is an unvalidated editorial exercise that cannot isolate any effect of the voice, but it applies the evidence standard in my methodology to one answer. If the app shows a transcript, read it back as text. If not, write the advice down in your own words. Sort the lines into statements of fact and guesses or reassurances, then take one factual statement and check it against a source the companion did not produce.
In the return example, that source is the store’s policy page. Before acting on a reassuring answer, read the words without the voice and check what actually supports the advice.