Large language models in psychology

This work explored whether people would rather choose to receive empathy from human or AI empathizers. When given the choice, participants sought human empathy, despite rating AI responses as more empathetic and making them feel more heard.

Across two studies, people felt closer to AI than to humans after emotional chat interactions, but only if the AI was labelled as human. Labelling the partner as AI reduced, but did not prevent, relationship-building.

People and LLMs evaluate deliberative reasoning more favorably than intuitive thinking—even when both yield accurate results. This preference appears to be intuitive itself and has implications for how we assess others’ and AI advise.

LLMs can tackle many higher-order tasks, raising the question of its ability to influence people’s political views. Across three preregistered experiments, here the authors show that LLM-generated messages can persuade people on various policy issues.

Attributions of distinct mental capacities to AI are shown to differentially relate to trust in advice. Mental states related to intelligence positively predicted trust, whereas attributions of experience were negatively related to advice-taking.

AI-generated empathic responses were rated higher in compassion, responsiveness, and preference than human ones in third-party evaluations across four preregistered experiments.

This study demonstrates that deep neural networks, like humans, show a learnability advantage when trained on languages with more structured linguistic input, resulting in closer alignment with human learning. This finding has important implications for both understanding human language acquisition and designing artificial language systems.

AI’s greatest strength—removing friction from work and relationships—is also a liability. Prioritizing outcome over process, it eliminates desirable difficulties that drive growth. By subtracting effort from life, AI risks removing the struggles that teach us, the loneliness that connects us, and the labor that makes life meaningful.

The rush to study generative AI is producing a feedback loop of topical and methodological convergence, flattening scientific imagination and crowding out the pluralism needed to keep research adaptive, resilient, and intellectually generative.

As human-AI collaborations become the norm, we should remind ourselves that it is our basic nature to build hybrid thinking systems – ones that fluidly incorporate non-biological resources. Recognizing this invites us to change the way we think about both the threats and promises of the coming age.

Salvi et al. find that GPT-4 outperforms humans in debates when given basic sociodemographic data. With personalization, GPT-4 had 81.2% higher odds of post-debate agreement than humans.

Large language models perform well in self-interested games such as the iterated Prisoner’s Dilemma but struggle in games that require coordination. Social reasoning strategies can improve cooperative outcomes with both other models and human players.

The use of ChatGPT helped generate more creative ideas for various everyday and innovation-related problems, compared with not using any technology or using a conventional Web search (Google). This effect remained robust regardless of whether the problem required consideration of many (versus few) constraints and whether the problem was viewed as requiring empathetic concern.

Testing two families of large language models (LLMs) (GPT and LLaMA2) on a battery of measurements spanning different theory of mind abilities, Strachan et al. find that the performance of LLMs can mirror that of humans on most of these tasks. The authors explored potential reasons for this.