People often treat metaphors as little more than linguistic icing. But what if they’re actually shaping how we think, feel & even vote?
That’s the premise of an interesting, yet somewhat depressing, new study by Meng, Li & Sun (2025), who combined embodied cognitive linguistics (i.e. the idea that our thoughts & metaphors are grounded in physical experiences like movement, force, or bodily sensation) with cutting-edge prompt engineering in ChatGPT-4 to analyse political metaphors in four of Donald Trump’s key campaign speeches.
Their aim? To see if large language models (LLMs) could systematically identify the deep, often subconscious metaphorical structures that influence public opinion-something traditional methods like manual annotation or corpus tools (e.g. Wmatrix) have long struggled with.
What did they find?
Using a critical metaphor analysis framework & a “chain-of-thought” prompt design, the researchers asked ChatGPT-4 to identify metaphors, map them to source domains (like Force, Movement, or Health), & interpret their social impact.
ChatGPT identified 119 metaphorical sentences across 28,000+ words of speech data with an impressive 86.2% accuracy, according to expert raters. The model tagged phrases such as:
“We rise together or fall apart” → Movement & Direction
“A tide of change is sweeping the country” → Force
“Help our country heal” → Health & Illness
These metaphors helped frame Trump’s agenda not just as policy, but as emotionally resonant narrative: recovery, conquest, resurgence.
Even more interesting: the metaphor “tariffs protect the soul of our country” was mapped to Human Body imagery, linking economics to cellular-level defence. For example, when Trump says, “tariffs protect the soul of our country”, the metaphor maps economic policy onto biological defence- as if the nation were a living body with a protective immune system. It’s the kind of imagery voters (& learners) might not consciously notice, but still emotionally respond to.
When the researchers compared ChatGPT’s results to more traditional, manual methods, they found a very strong overlap. The model was especially good at spotting metaphors based on the human body or health -like “ambition is the lifeblood”- but it struggled with less common ones, such as plant-based metaphors. In other words, AI is getting impressively good at metaphor detection- but human judgement is still essential, especially for subtler or culturally specific language.
Crucially, the study doesn’t claim these metaphors changed votes- but it does show how they frame complex political ideas in ways that feel emotionally natural. These metaphors don’t just describe- they mobilise, persuade & shape perception. And that’s what makes them so powerful.
Teacher Takeaways?
- Metaphors teach more than vocabulary. Maybe students could analyse metaphors for what they reveal about values, attitudes, or cultural scripts– especially in persuasive genres like ads, political speeches, or news headlines.
- Try a metaphor-spotting challenge. Ask learners to identify metaphors in political discourse from different countries, then compare the source domains. Are British politicians more fond of sporting metaphors than Americans?
- AI as a research tool. This study models how LLMs like ChatGPT can help students investigate language more critically- e.g., using prompts to reveal metaphor use in news reports, business English, or TED Talks.
Do you ever explore how metaphors differ across languages or cultures?
Leave a Reply