You type "the smell of rain on dry earth" into an image generator. What comes back is a visual: dark soil, grey skies, perhaps a mist rising. It's beautiful. But does it capture the smell? The petrichor, the ozone, the particular sweetness of water hitting dust? The AI has read about this smell thousands of times. It knows what words we use to describe it. But can it represent it in any meaningful way?
This is the challenge of olfactory prompting. Smell is the most elusive sense, the one with the poorest vocabulary, the one most tightly bound to memory and emotion. And we're asking AI to generate representations of it without ever having smelled anything.
Let's explore this sensory frontier. By the end, you'll understand why smell is uniquely difficult, how models try to bridge the gap, and techniques for prompting that evoke the ineffable.
The Poverty of Olfactory Vocabulary
Of all our senses, smell has the weakest linguistic toolkit.
The Numbers:
We have thousands of words for colors.
Hundreds for sounds and textures.
Dozens for tastes.
And for smells? A handful. "Sweet," "sour," "musky," "pungent," "floral," "earthy." That's about it.
Why?
Smell is processed in the brain's emotional and memory centers, not its language centers.
Most odors don't have dedicated names. We describe them by analogy: "smells like a rose," "like coffee," "like the ocean."
Smell is intensely personal and context-dependent. The same odor can evoke different associations for different people.
The Consequence for AI:
The model has read millions of smell descriptions. But those descriptions are almost always analogical: "smells like X." The model knows what X is (a rose, coffee, the ocean). But it doesn't know what the smell is. It knows the comparison, not the sensation.
A Contrarian Take: The Poverty of Olfactory Language Isn't a Bug. It's the Whole Point.
We tend to see the lack of direct smell words as a limitation. But what if it's actually the key to how AI can represent smell?
Because we can't name smells directly, we're forced to use context, association, and emotion. A smell isn't just a smell; it's the memory of your grandmother's kitchen, the feeling of a coming storm, the comfort of a familiar place.
When AI generates a representation of "the smell of rain on dry earth," it's not trying to simulate a chemical compound. It's trying to evoke the world of that smell the dark clouds, the thirsty ground, the anticipation. And that's exactly how humans experience smell. We don't smell molecules; we smell memories.
The AI's olfactory poverty is a mirror of our own. We can't name smells either. We can only describe their contexts, their associations, their emotional weight. In this, the AI is doing exactly what we do: using language to point toward what cannot be said directly.
How Models Handle Smell
Given these limitations, how do models generate olfactory content?
Direct Association
The model knows that certain objects have characteristic smells. "Coffee" → bitter, rich, morning. "Rose" → sweet, floral, romantic. It can use these associations to generate descriptions.Contextual Reconstruction
The model can build a scene that implies a smell. A description of a bakery at dawn implies the smell of bread. A forest after rain implies petrichor. The smell is evoked through setting, not naming.Metaphorical Extension
The model can use metaphor to suggest smell qualities. "The air was thick with the sweetness of decay." "A sharp, metallic tang hung in the air." These phrases give texture to smell without naming it.Emotional Association
Smell is tied to emotion. The model can evoke smell through emotional context. "The smell of her perfume brought back a decade of longing." The smell isn't described; its effect is.
The Prompt Engineer's Toolkit for Smell
- Use Objects as Smell Proxies Instead of naming the smell, name the source.
Instead of "a sweet, floral scent," try "the scent of jasmine climbing the garden wall."
Instead of "a smoky aroma," try "the smell of a campfire still clinging to his jacket."
- Build Scenes That Imply Smell Create the context that carries olfactory associations.
"A bakery at dawn, ovens warming, yeast and sugar in the air."
"A pine forest after rain, needles soft underfoot, the sharp scent of resin."
- Use Metaphor for Texture Describe the quality of the smell through comparison.
"The smell was sharp as cut metal."
"The air was thick with sweetness, like honey left too long in the sun."
- Connect Smell to Memory and Emotion Anchor the smell in human experience.
"The smell of her grandmother's kitchen: cinnamon and butter and the particular warmth of a wood stove."
"He caught a whiff of salt and brine and was instantly a child again, standing on a dock at dawn."
- Use Sensory Layering Combine smell with other senses to build a fuller picture.
"The air was cool and damp, carrying the smell of wet earth and the distant sound of thunder."
Prompt Templates for Olfactory Scenes
For Direct Smell:
"Describe the smell of [object/scene] using [metaphor] and focusing on [memory/emotion]."
For Implied Smell:
"Create a scene that evokes the smell of [place/experience] without directly naming it. Use sensory details: temperature, humidity, light, sound."
For Olfactory Memory:
"Write about a character who encounters a smell that triggers a vivid memory. Describe both the smell and the memory in sensory detail."
Reading Olfactory Outputs
When you generate smell descriptions, evaluate them.
Signs of Success:
You feel a ghost of the smell, however faint.
The description evokes a specific memory or place.
You can almost imagine being there.
Signs of Failure:
Generic descriptions ("it smelled nice").
Over-reliance on a few adjectives ("sweet," "floral").
No sensory grounding, just abstraction.
Your Olfactory Practice
Step 1: Start with What You Know
Choose a smell you know intimately. Describe it to the AI in your own words, using metaphor, memory, and context. Then compare the AI's output to your actual experience.
Step 2: Collect Olfactory Language
Gather descriptions that work for you. Build a personal lexicon of smell words, phrases, and metaphors. Share with others.
Step 3: Experiment with Implication
Try to evoke a smell without naming it. Describe a scene that implies the smell. See what the model produces.
Step 4: Push into the Abstract
Try to describe smells that don't exist. "The smell of a forgotten dream." "The scent of a color you've never seen." See what emerges.
Step 5: Document Your Findings
Keep a journal of successful olfactory prompts. Over time, you'll develop an intuition for what works.
The Smell of Words
We will never know if AI "understands" smell. It has no nose, no memories, no embodied experience. But it has our language our metaphors, our memories, our attempts to capture the uncapturable.
When you prompt for a smell and get a response that moves you, you're not experiencing AI's sense of smell. You're experiencing your own, reflected through the lens of everything humans have ever written about scent.
The AI doesn't smell. But it remembers everything we've said about smelling. And in that archive, there's a kind of memory of smell a ghost of sensation that can, sometimes, evoke the real thing.
What's the most powerful smell memory you have? Can you describe it in words that make someone else almost smell it? What's the closest you've come?
Top comments (0)