Earlier this year, we published a Viewpoint on uses of and challenges for generative AI in psychology1. All six experts who contributed to that article were sceptical about the possibility for generative AI systems such as large language models (LLMs) to ‘think’ like humans or to replace humans in most tasks. However, they agreed that this new technology can be useful to advance research and clinical practice. We now close the year as we started it, by thinking about AI: our December issue includes three articles in which psychologists describe how people interact with new AI tools such as LLM chatbots and how they can be used to enhance and support human behaviour rather than undermine or replace it.

In a Perspective, Gonzalez and Heidari discuss the concept of human–AI complementarity, or how the unique strengths of humans and AI can be productively combined. In broad terms, humans are skilled at navigating uncertainty and novel situations and understanding other humans’ minds, emotions and intentions. By contrast, AI can process large amounts of data quickly and perform detailed tasks repeatedly. Gonzalez and Heidari review how these distinct skill sets can be drawn together for better decision-making outcomes than could be achieved by humans or AI alone. For instance, in a disaster scenario such as a hurricane, AI systems can rapidly process incoming weather and flood data and make recommendations for evacuation and shelter locations, while humans can track the ethical and societal risks of impactful decisions. Overall, understanding how to productively and safely integrate AI with human decision-making can ensure the best outcomes in both extreme and mundane situations.

In another Perspective, Timmons and colleagues lay out a related framework for human and AI collaboration in the realm of mental health interventions. They point out that bias in AI tools used in mental health contexts can lead to underdiagnosis or overdiagnosis and to unequal treatment accessibility and efficacy. To mitigate these risks, they advocate for co-creation processes in which community members are involved across each stage of AI development, testing and implementation to ensure that AI systems are grounded in psychology evidence and the needs of the individuals who will receive the interventions. Through such an iterative process, equitable AI-based mental health interventions can adapt to societal and individual needs.

Finally, in a Review, Brady and colleagues acknowledge that LLMs and chatbots are already being used to make real-world decisions, and reason that researchers need to understand what underlies these decisions and how they relate to decisions made by humans alone. Brady et al. highlight a framework developed for human decision-making in which humans are thought to use a fast ‘system 1’ mode and a slower, more deliberate ‘system 2’ mode to make decisions2. The research they review suggests that LLM outputs can have similarities to each of these human thinking modes. A combination of fast, automatic responses and slower, more energy-intensive processing might lead to the most robust and human-like responses.

“Studying these applications can prompt theoretical advances... as researchers grapple with foundational questions about what it means to be a thinking being”

These articles provide greater clarity about what to expect when offloading a decision or a mental health intervention to a chatbot as opposed to another human being. To use generative AI responsibly, researchers and the general public need to understand how these systems respond in specific situations and what the risks are. Generative AI outputs can repeat or exacerbate biases and stereotypes in their training data3 and generally are not built to respond according to human ethical or moral codes4. Thus, the use of generative AI for decision-making can reduce fairness5 and have unintended consequences6 that — given the complexity of these systems — are not easy to predict. Furthermore, the use of generative AI for decision-making has the potential to reduce human self-determination and motivation7. As a consequence, there is considerable risk to wholly replacing therapists, lawyers, health practitioners or scientists with generative AI tools. However, studying these applications can prompt theoretical advances in psychology as researchers grapple with foundational questions about what it means to be a thinking being.

The articles in this issue are joined by other pieces we’ve published on AI and its applications in psychology in a Collection that will be periodically updated. We look forward to seeing how the study of AI intersects with psychology research in the coming years.