When AI Stops To Smell The Roses (Part 1/2): Teaching Machines To Sniff

Welcome to this week’s Deep-Fried Dive with Fry Guy! In these long-form articles, Fry Guy conducts in-depth analyses of cutting-edge artificial intelligence (AI) developments and developers. Today, Fry Guy dives into AI’s emerging sense of smell. We hope you enjoy!

*Notice: We do not receive any monetary compensation from the people and projects we feature in the Sunday Deep-Fried Dives with Fry Guy. We explore these projects and developers solely to showcase interesting and cutting-edge AI developments and uses.*

Imagine if your computer could smell a rose or sniff out a gas leak. It sounds like something out of a sci-fi cartoon—a robot wrinkling its nose—but it’s quickly becoming reality. In labs around the world, scientists are training AI to recognize and recreate odors. But this isn’t just a quirky science experiment; giving computers a sense of smell could improve health care, safety, and even how we experience entertainment. From AI “noses” that detect diseases on your breath to digital aromas you can transmit like an email, welcome to the wild world of sniffing machines.

THE CHALLENGE OF TEACHING MACHINES TO SMELL

Smell is our most mysterious sense. We have measurements for vision and hearing, but no simple spectrum exists for odor. Humans have around 400 different olfactory receptors in the nose (often called “channels”), compared to just 3 color receptors in the eye​. When we smell a cup of coffee, hundreds of odor molecules trigger a unique pattern across those receptors, and our brain magically interprets “ah, coffee!” Teaching a machine this trick is hard. There’s no obvious numeric code for “smell” like Red-Green-Blue (RGB) values for color.

Early attempts at machine smelling used arrays of gas sensors nicknamed “electronic noses.” These e-noses could detect chemical substances, but interpreting what a blend of molecules smells like (floral? fruity? foul?) is a hefty AI problem. Essentially, smells are multi-dimensional—a complex mixture can change one molecule and go from delightful to disgusting. Only recently have scientists started cracking this code with AI.

In 2023, a team of researchers formulated a “Principal Odor Map” (POM)—essentially a digital map of smell space that acts like an RGB-color model for aromas​. This map was built by training machine learning models on thousands of known odor molecules and their scent profiles​. The result? A tool that can predict how a molecule will smell and where it lies in the smell spectrum, even for scents no human has smelled before​.

The very complexity that makes smell fascinating also makes it ripe for machine learning. Computers don’t have noses, but they excel at finding patterns in big data. With enough examples of molecules and their odors, an AI can start connecting chemical structure to scent—something that has stumped scientists for ages. In fact, the new AI-driven odor map could predict a scent even better than human experts​. In tests, the AI’s scent descriptions matched or beat human panelists on over half of trials​! We’re basically giving computers a crash course in “olfactory language”—teaching them that molecule X smells “floral with a hint of vanilla” while molecule Y smells “like rotten eggs.” It’s a tough lesson plan, but AI is finally starting to pass the sniff test.

SNIFFING OUT DISEASES

Teaching AI to sniff isn’t just for fun. One of the most promising applications is in healthcare. Doctors have long known that diseases can produce unique odors—for example, the smell of acetone on a diabetic’s breath or how certain cancers give off telltale volatile compounds. Trained dogs have even sniffed out illnesses like cancer with astonishing accuracy (in one study, dogs identified prostate cancer from urine samples with over 96% accuracy​). But as amazing as dogs are, they get tired, and they can’t explain what they smell. Enter the electronic nose with AI.

Researchers are building AI systems that can “smell” diseases on your breath or bodily fluids much like a dog would, but with silicon sensors. Every human breath contains a “breathprint” of over 1,000 distinct molecules​ carrying clues about our health. For decades, we have enlisted animals—dogs, rats, and even bees—to sniff out diseases like cancer, diabetes, and tuberculosis from those molecules​. Now, new devices aim to do the same in a more controlled, instant way. In 2023, scientists at CU Boulder and NIST demonstrated a laser-based AI breathalyzer that could detect COVID-19 in real-time with excellent accuracy​. The device used lasers to identify trace chemicals in exhaled breath and AI to match the pattern to COVID’s “smell signature.” Impressively, it worked in seconds and showed the potential to screen for other diseases as well​.

This isn’t an isolated example. Other research teams have developed AI noses to sniff out hard-to-detect cancers. A Penn Medicine study created an “electronic nose” that identified difficult cancers like ovarian and pancreatic with up to 95% accuracy​. Another project combined dogs and AI, effectively training a computer on what the dog smells, to spot cancer in patients with similarly high accuracy​. The pattern recognition abilities of AI are key here: a tumor doesn’t emit one single “smell molecule”—it creates a subtle cocktail of chemicals. AI can learn to recognize that complex pattern even if we don’t know exactly which molecule is the giveaway​.

What could this mean for the future? Perhaps at your annual checkup you’ll breathe into an AI-powered sensor in addition to getting your blood pressure checked. In fact, the CU Boulder team imagines a future where you blow into a mouthpiece on your phone for a quick health screen​.

It’s worth mentioning that applications of AI smell could extend beyond diagnostics. An AI nose never gets bored or distracted, and it could continually monitor breath or air quality for dangers. This tech could enhance safety by detecting gas leaks, pollution, drugs, or spoiled food by smelling these chemicals faster and more precisely than a person could.

TEXTING SMELLS: DIGITIZING AND TRANSMITTING AROMA

If AI can learn to recognize and predict smells, can it also reproduce them? In other words, can we send smells through the internet the way we send texts or photos? It may sound fanciful, but researchers are already working on “digital smell” or “scent teleportation.” In 2024, the team at Osmo announced they had captured the scent of a fresh-cut plum and “reprinted” it in another location with no human nose involved​. This feat—effectively emailing a smell—was the first real demonstration of full scent digitization and reconstruction.

How does one teleport a smell? The process combines chemistry with AI. First, you need to capture the smell’s fingerprint. Osmo’s approach was to use a gas chromatography-mass spectrometry (GC-MS) device to break down the plum’s aroma into its component molecules, a bit like identifying the ingredients in a secret recipe​. This data is then turned into a digital encoding—in this case, plugging the recipe into that Principal Odor Map (above) to get an odor “coordinate” in a scent database​. Next, on the receiving end, you need a scent printer. Osmo used a formulation robot loaded with many base scent ingredients. The AI takes the plum’s encoded recipe and figures out how to mix those ingredients to recreate the smell​. The result was a faithful fake plum aroma diffused in the air, reportedly almost indistinguishable from the original plum.

This “scent teleportation” is still in early stages, but it opens the door to some amazing (and amusing) possibilities. Imagine sending your friend the smell of freshly baked cookies along with a photo, or a video game that emits the scent of a smoky fire as you explore a virtual forest. Engineers are already experimenting with VR headsets that include olfactory stimulation. One demo used digital incense in a meditation app to enhance relaxation​. There are also prototypes of wearable smell gadgets like smart necklaces that monitor the air for hazards or just give you a whiff of a calming scent when stressed. It’s a blend of practical and playful: a digital nose could allow you to smell a movie or let you download the aroma of “ocean breeze” to spice up your living room.

Of course, transmitting smells isn’t as straightforward as sending bytes (numerals or digits). Real odors involve real molecules. So any “digital scent” system ultimately needs hardware—cartridges of chemicals to release, or scent emitters—to turn code into a sniffable form. That means an AI nose will likely involve refillable scent packs somewhat like ink cartridges for a printer. And there’s the challenge of calibration: everyone’s nose is slightly different and environment matters (that cookie smell might mix oddly with your room’s existing aroma, for instance). But still, the fact that scientists have even partially digitized smell is a testament to how far AI and sensor tech have come​. We’re literally coding the smells of the world!

Next week, in Part 2, we will explore whether AI might develop “smell preferences,” and what this could mean for the future implementation of AI smelling systems.

You’ve heard the hype. It’s time for results.

For all the buzz around agentic AI, most companies still aren't seeing results. But that's about to change. See real agentic workflows in action, hear success stories from our beta testers, and learn how to align your IT and business teams.

Watch the product showcase now