On today's post, Yufeng Liu (Lancaster University), whose research focuses on metaphor, health communication and AI, explores whether AI can understand the metaphors we use about illness.
 |
| Yufeng Liu |
Can AI Understand the Metaphors We Use About Illness?
When people talk about cancer or depression, they often use metaphors. Cancer can be an “enemy” to fight, recovery can be a “journey”, and depression can feel like “darkness”, a “heavy burden”, or being “trapped in a box”. These expressions do more than describe illness. They help people make sense of difficult experiences and communicate feelings that may otherwise be hard to put into words.
I use “AI” broadly here to include large language models, and this study focuses on ChatGPT. As more people turn to it for comfort, or simply someone to talk to, I asked: can ChatGPT pick up the metaphors people use about illness, and does it respond differently when the same experience is expressed literally?
In my recent article, “Exploring AI’s processing of and responses to metaphors in health-related conversations”, published in the Journal of Pragmatics, I compared ChatGPT’s responses to metaphorical and literal statements about cancer and depression. I asked ChatGPT to respond as a friend rather than a medical professional.
ChatGPT could recognise, reuse, and generate metaphors. When a user described depression as a dark tunnel, it might talk about light or finding a way forward. Metaphors helped it rephrase the user’s experience and offer a more positive perspective, creating a sense of alignment and support. Metaphorical prompts produced significantly more metaphors in the replies for both illnesses.
The illness being discussed influenced reply length more than whether the prompt contained metaphors. For cancer, metaphorical prompts received slightly longer replies. For depression, the pattern was reversed. ChatGPT seemed more responsive to depression. When responding to non-metaphorical prompts, its replies about depression contained more explanation, support, and questions than its replies about cancer.
Across both illnesses, however, the replies were largely formulaic. ChatGPT often expressed support, repeated the user’s meaning in different words, suggested another way of viewing the experience, and ended with a question. The middle two steps were especially common, regardless of the illness or whether the user had used a metaphor.
The metaphors were also largely familiar, such as illness as a journey, darkness, or burden. More importantly, they focused on the general experience of being ill, while paying less attention to relationships with family, friends, and healthcare professionals.
Living with cancer or depression is not about having an illness. It may change who we are, how we understand our lives, how we work, and how we relate to others. This is something LLMs may still overlook. Responding fluently is not the same as responding sensitively.
One important limitation is that the study examined only GPT-4o mini, the most capable ChatGPT model publicly available to me when I conducted the research. AI models have developed rapidly since then, so these findings do not necessarily apply to every current or future model. Even so, the study provides a starting point for asking whether AI can recognise metaphors and respond in ways that fit both the purpose of the conversation and the person.