September's Paper of the Month comes from Tilburg University, UCL, University of Amsterdam and Scuola IMT Alti Studi Lucca, published in Nature Communications.
Paper: Perceived humanness and empathy are dissociated in relationship advice generated by a large language model Authors: Bennett Kleinberg | Jari Zegers | Jonas Festor | Stefana Vida | Julian Präsent | Riccardo Loconte | Sanne Peereboom
What the research is about
What makes us believe a message was written by a person? Across five studies, the researchers compared relationship advice written by humans with texts from GPT-4 and GPT-4o. Other participants judged whether each text came from a human or AI and, in one study, how empathetic it was. Why this matters As AI-generated text becomes more common, it gets harder to tell whether we are dealing with a person or a machine.
What they found
Normally, people were fairly good at spotting the AI. Its advice stood out through long, less common words and stiff phrasing such as "Dear friend". When GPT-4 was told to appear as human as possible, it switched to a casual "hey", words like "yeah" and "haha", more "I" and simpler vocabulary. People then became much less accurate, although human advice was still judged as somewhat more human. So it was style, not empathy, that made a message feel human. When asked to show empathy, the AI could do it: its advice was rated as more understanding and caring. Yet that same advice was rated as less human. The authors call this "stochastic empathy": the AI reproduces the language of empathy, but that alone does not make a text seem human.
What it could mean
If AI can sound human on demand, it could become harder to know who we are talking to online, and trust in digital communication may suffer.
What's next
The study covered only English and two GPT models, so the authors suggest testing other languages and more advanced models.
