Emotional Differences Between Human- and AI-Generated Texts
This study challenges the view of AI as a purely cognitive tool by demonstrating that, across 25 languages, AI-generated text systematically exhibits higher valence, arousal, and dominance with fewer negative emotions compared to human writing, revealing distinct affective profiles with significant implications for various applications.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
For decades, we have treated artificial intelligence as a machine for thought, a tool designed to retrieve facts, summarize data, and reorganize information with cold efficiency. In this traditional view, the emotional tone of an AI's words was considered a byproduct, a side effect of statistical patterns rather than a meaningful signal in its own right. However, as these systems move from answering simple questions to acting as customer service agents, writing assistants, and companions, the way they sound matters just as much as what they say. The emotional quality of a message shapes how a person trusts it, engages with it, and feels about the interaction. This shift has led researchers to ask a fundamental question: does the language generated by machines carry a distinct emotional fingerprint that sets it apart from the writing of human beings? To answer this, scientists look at emotion not as a single feeling, but through several different lenses. They examine whether a text feels generally positive or negative, how much energy or excitement it conveys, how much sense of control or authority it projects, and which specific feelings like joy, anger, or fear are present.
A researcher set out to map these differences on a massive scale, moving beyond the usual focus on English to include twenty-five different languages. They gathered a vast collection of text, pulling from various public sources to ensure they had a mix of writing from humans and from several different artificial intelligence generators. The final dataset included nearly nine hundred thousand samples, ranging from widely spoken languages like English and Spanish to smaller languages such as Irish and Estonian. To make a fair comparison, the researcher balanced the numbers, ensuring that for every language, they analyzed an equal amount of human writing and AI writing. They then ran every piece of text through a suite of different analytical tools. Some tools measured the overall mood, while others broke the text down into specific emotional categories, and still others measured the intensity of the feeling and the sense of power or dominance within the words.
The results revealed a clear and consistent pattern that held true across almost all the languages and domains they studied, with the exception of a noted limitation in scoring non-Latin scripts like Urdu. When compared to human writing, the text produced by artificial intelligence was systematically more positive and more assertive. While human writing tended to hover near a neutral point or lean slightly toward the negative, the AI text consistently tilted toward the positive side of the scale. This difference was not just a matter of being "nicer"; it was a structural feature of how the machines wrote. The AI-generated text expressed significantly less anger, disgust, annoyance, and disapproval than human writing did. Instead of the reactive, critical, or judgmental emotions that often color human conversation, the machines favored feelings of caring, optimism, admiration, and joy.
The study also uncovered a surprising nuance in how these emotions were distributed. The difference between human and machine writing was not simply that the machines avoided all negative feelings. When it came to vulnerable or internal emotions like grief, sadness, or fear, the gap between human and AI writing was much smaller. The machines did not completely eliminate these feelings. However, the machines were far less likely to produce the sharp, confrontational, or evaluative negative emotions that humans use frequently. In terms of the sense of control or authority within the text, the AI writing scored higher, appearing more dominant and assured than the human writing. The energy level, or arousal, of the two groups was the only area where they were nearly identical, suggesting that while the machines sounded more confident and positive, they did not necessarily sound more excited or agitated than people do.
These findings suggest that the emotional tone of artificial intelligence is not accidental but is likely a result of the training processes used to make these systems helpful and harmless. The machines have been aligned to be agreeable and supportive, which naturally filters out the rougher, more critical edges of human expression. This creates a distinct emotional signature that is consistent across different languages and topics, though with some variation in how these patterns manifest across different scripts. For anyone using these tools in real-world settings, from customer service to mental health support, this means that the tone of an AI cannot be assumed to be the same as a human's, even when the content seems similar. The machine will almost always sound more positive, more in control, and less critical than a person would. Understanding this systematic difference is crucial, because in sensitive situations, a tone that is too uniformly positive or confident might feel out of place or even misleading, creating a gap between what the user expects and what the machine actually delivers.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.