A sentence can sound compassionate without sounding human. It can also sound unmistakably human while offering very little care.
That distinction is easy to lose in debates about whether a model writes like a person. We often bundle warmth, naturalness, sincerity, and understanding into one impression. A September 3 study in Nature Communications asks what happens when those qualities are examined separately.
Bennett Kleinberg and colleagues compared human and model writing across five studies involving relationship advice and descriptions. They found that prompting GPT-4 to appear human increased perceived humanness, while human writers changed less under similar instructions. Further experiments included GPT-4o. The authors also found that perceived empathy and perceived humanness could be separated: language judged more empathic did not automatically seem more human.
The paper examines judgments of text. It does not establish that a model experiences empathy, that a caring-sounding answer is beneficial, or that its advice improves a relationship. Its value is more precise: the cues that persuade us a person is present may differ from the cues that make words feel supportive.
The familiar voice is a set of clues
We infer a writer from small features. A contraction, a hesitation, an informal aside, or a slightly uneven sentence can make prose seem conversational. The paper's analysis connects model changes in perceived humanness with surface features of language.
That finding complicates a popular demand: make it sound more human. A system can respond to that instruction by changing its style. The resulting text may feel closer without gaining a biography, shared history, or a reason to care what happens next.
For a reader, the distinction matters because natural language invites familiar social habits. We respond to an apparent listener. We supply the person behind the voice even when the interface has already told us it is a machine.
That response is not foolish. Human communication depends on making inferences from language. The challenge is that the same cues can now be produced without the relationship we normally associate with them.
Warmth can still be useful
It would be a mistake to treat every supportive sentence as empty simply because a model generated it. Clear, considerate language can make a tool easier to use. A confusing error message does not become more honest by sounding cold. An explanation can acknowledge frustration without claiming to feel it.
The question is what the words ask the user to believe. “That sounds frustrating; here are the steps to recover the file” frames a practical response. A claim of personal attachment implies something else. Designers and writers can choose language that is kind while being accurate about the system's role.
Those examples are editorial illustrations, not recommendations validated by the study. They show why evaluating warmth alone misses part of the problem. A pleasant interaction can still obscure who can help, what the system knows, or where a person can take over.
When a service uses a conversational assistant, a clear path to a human can matter more than an impressive imitation of one. Readers need a way to resolve the issue that brought them there.
Stop judging sincerity by texture
The same research also cautions against judging people by whether their writing resembles our idea of AI. A precise, orderly message may come from someone who values clarity or works in another language. A casual, idiosyncratic paragraph may have been generated on request.
If “sounds human” becomes the sole test, people may be rewarded for performative informality while useful assisted writing is dismissed. The result could be a style contest with little connection to the qualities we actually want: truthfulness, care, responsibility, and the capacity to respond.
Ask more concrete questions. Does the statement reflect the writer's position? Are facts and promises accurate? Is the source of advice appropriate to the situation? Is someone available to hear a correction? These questions are harder than spotting a favorite punctuation mark, but they tell us more.
Name the quality you actually need
For a help article, the goal might be clarity and a usable next step. For a personal letter, it might be an honest expression of the sender's feelings. For a public explanation, it might be evidence that a reader can inspect. Each requires a different review.
The study does not settle the larger question of human-machine relationships. It gives that debate a valuable separation. Perceived empathy is one response to language. Perceived humanness is another. Neither, on its own, proves that someone understands or stands behind the words.
The next time a sentence feels especially human, pause over what that feeling is doing. It may be helping you read. It may also be answering a question the text cannot settle: who is here with me, and what can I reasonably expect from them?
Sources and limitations
- Kleinberg et al., “Perceived humanness and empathy are dissociated in relationship advice generated by a large language model”: checked October 15, 2026. Published September 3, 2026 in Nature Communications. Supports the study design, prompting effects, tested models, and distinction between perceived humanness and empathy. It does not measure emotional experience or establish the real-world benefit of advice. The broader design implications above are editorial analysis.