Even the smartest AI still can’t smell the roses. New research shows that artificial intelligence fundamentally misunderstands the world — especially the parts we sense and feel.
Key Points at a Glance
- AI struggles to understand sensory-rich human concepts like “flower”
- Language models lack the physical experiences that ground human knowledge
- Words tied to touch, smell, and body movement expose AI’s conceptual blind spots
- Future AI may improve by integrating sensors and robotic interaction
Artificial intelligence has mastered text and begun to crack images — but it still doesn’t know what a flower truly is. That’s the striking conclusion of a new study from The Ohio State University, which found that AI systems fail to represent the sensory and emotional dimensions of many human concepts.
“A large language model can’t smell a rose, touch the petals of a daisy, or walk through a field of wildflowers,” said Qihui Xu, lead author of the study published in Nature Human Behaviour. “Without those experiences, it can’t fully understand what a flower is.”
Researchers compared how humans and several advanced AI models — including GPT-4 and Google’s Gemini — understood over 4,400 words, ranging from abstract terms like “humorous” to sensory-rich words like “hoof” and “flower.”
Using two well-established linguistic measures, the Glasgow Norms and Lancaster Norms, the study explored how words are rated in terms of emotional impact, visual imageability, sensory experience, and motor interaction. The results revealed a clear pattern: AI performs well on abstract or language-based concepts, but stumbles when the meaning of a word is tied to touch, smell, or bodily action.
“For humans, the concept of a flower includes its fragrance, the feel of petals, the sight of color, and even the emotional reactions it stirs,” the researchers wrote. “AI, trained only on language or images, misses this multisensory richness.”
The implications are far-reaching. While current AI models excel at summarizing text or completing sentences, they may falter in contexts that require embodied understanding — such as caregiving robots, education, or human-like conversation.
Even though some AI models trained on images outperformed those trained only on text, the gap remained for concepts linked to non-visual senses and physical interaction. Xu emphasizes that this limitation isn’t simply a matter of more training data. “AI consumes orders of magnitude more text than a human sees in a lifetime,” she said, “and still can’t match the human concept of a flower.”
The study suggests a potential path forward: integrating sensor data and robotics into AI systems to give them real-world sensory experiences. “Only by touching, smelling, and moving in the world can AI begin to build the kind of deep, embodied knowledge humans take for granted,” Xu said.
While large language models have dazzled with their fluency and power, this research is a reminder of their limits. As AI continues to evolve, bridging the sensory gap between humans and machines may be essential to making truly intelligent systems — ones that don’t just talk about the world, but live in it too.
Source: Ohio State University
Enjoying our articles?
We don’t show ads — so you can focus entirely on the story, without pop-ups or distractions. We don’t do sponsored content either, because we want to stay objective and only write about what truly fascinates us. If you’d like to help us keep going — buy us a coffee. It’s a small gesture that means a lot. Click here – Thank You!
