This smell startup says its AI nose is as good as a human’s

This smell startup says its AI nose is as good as a human’s
In a lab near Newark Airport in New Jersey, a team of around 30 experienced human sniffers are slowly working their way through every odor on Earth. Their

In a lab near Newark Airport in New Jersey, a team of around 30 experienced human sniffers are slowly working their way through every odor on Earth. Their careful notes—“nutty,” “peachy,” “marine”—are cross-referenced with a chemical analysis to create a rich set of training data for what AI startup Osmo calls the world’s first foundation model for smell. The company’s goal, says founder Alex Wiltschko, is “to give computers a sense of smell.”

“Scent molecules are really the only language most species on this planet can speak,” he says. “Most species can’t communicate with sound or visually, so it truly is Earth’s original language. And it’s incredibly powerful for us, from a memory and an emotional standpoint. You can smell something once and remember it for the rest of your life. You can smell something on the street and be pulled back to high school.”

In the last few months, Wiltschko claims, Osmo’s model has achieved “human-level olfactory intelligence.” The company’s definition of this is that the model is about as good as the average person at detecting and discriminating between smells, and about as good as a junior perfume designer at creating them. It says its technology is already being used in the fragrance industry to help brands design and manufacture new scents. Last month, the company says, Osmo’s AI model outperformed in-house perfumers at one of the world’s largest consumer goods companies. Wiltschko believes it could be deployed eventually for threat detection at airports, for example, or diagnosing diseases. 

Wiltschko spun the startup out of his work at Google Brain, the search giant’s former AI research lab, in 2023, and he is running the same playbook as AI startups in other domains: Throw a lot of data at difficult problems, and see what happens. The challenge with scent, Wiltschko says, is that there was a dearth of good data. Before Osmo, there were fewer than 200,000 human descriptions of smells in existing databases, by his count. Plus, the chemical sensor devices needed to collect data are expensive, and the fragrance industry is secretive about its work on olfactory science. 

Osmo founder Alex Wiltschko says he wants to "give computers a sense of smell". OSMO AI

A lot of the company’s work has been in creating its own dataset of more than 10 million human labels and 800 million “scent tokens,” which describe the exact molecules and concentrations that go into creating specific odors. Osmo uses spectrometers and other chemical sensors, running around the clock, to break down the molecular makeup of everything from fruit and flowers to pantry  staples and organs. A team of around 70 trained data labelers in the US and overseas spend their days sniffing things and writing down what they smell like, using a standardized fragrance vocabulary.

Mapping out scents in this way is throwing out some unlikely parallels. In the same way that changing the order of words can drastically influence the meaning of a sentence, rearranging scent molecules in different proportions can drastically alter the smells humans perceive. “The scent tokens that define the smell of a newborn baby have massive overlap with the scent tokens in a fresh plate of cooked spaghetti,” Wiltschko says. (Specifically Barilla brand spaghetti, he says, which perhaps raises some questions for the people at the pasta factory.)

The company also creates its own scents, using a smell-printing robot the size of a school bus that mixes ingredients according to precise formulas to make samples for Osmo’s own research and fragrances for paying customers. It can design and print scents to order based on a ChatGPT-style prompt. Wiltschko flipped open his laptop to show me how the AI model worked. He opened a text box on screen, and I asked him to generate a fragrance that smelled like a Portuguese custard tart (we were in Lisbon for Web Summit), aimed at 18- to 23-year-old men. The system spent a few minutes cogitating, then generated three options for potential scent profiles that matched the brief (creamy, coconutty, fruity) and the molecules needed to create them. If I’d continued with the ordering process, my signature L’Eau de Custard would have been ready about a week later. 

Osmo’s technology has already been used to create retail fragrances, including for giants like Procter & Gamble. The Gates Foundation, which invested $5 million when Osmo launched, later gave the company a $3.5 million grant to assess molecules that repel disease-carrying insects, and Osmo identified eight substances that were more repellent than DEET. Osmo has also worked with StockX, the online clothing marketplace, on using smell to spot counterfeit sneakers by detecting differences in the chemical composition of the materials used in the shoes.

But not everyone is convinced by the technology or the company’s big claims. Avery Gilbert, a smell scientist and author of What the Nose Knows, has been sharply critical of Osmo on his Substack, arguing that some of its marketing claims are undocumented, and that its patent filings still rely on human sensory panels and perfumers, rather than AI. “This heavy reliance on human sensory judgments seems odd for a company premised on super-cool AI,” he wrote in March. A representative for Osmo did not respond in time for publication. 

Michael Schmuker, a former professor of neural computation at the University of Hertfordshire who has previously consulted for Osmo, says the company’s technology is an achievement. With colleagues, he created an open-source, reverse-engineered version of Osmo’s 2023 model, but it lacks the reams of data the company has collected since. 

But even with all this smell data, Osmo may be missing some of the factors that make human olfaction unique, Schmuker says. “Smell is super complex. Context is very important for smell. The same molecule may evoke a different association depending on context,” he says. “[The smell of] a French cheese like Camembert is partly the same molecule as in used nappies. You don’t confuse the two because of the context.”

In a controlled setting like a fragrance lab, this lack of context may not matter so much. But it could have an impact if Osmo’s technology becomes, as Wiltschko suggests, the “olfactive channel” for AI world models. Imagine a machine that could sniff out spoilage in food on a production line, or detect tumors through smell the way dogs have been trained to do in pilot studies.

For that to happen, sensor technology will need to get a lot better, Schmuker says. The models will need to keep improving. Medical diagnostics is the biggest prize. 

“There’s a lot of history of companies that have gone directly at that mountain peak and, unfortunately, not been successful, because it is very expensive and it is very difficult,” Wiltschko says. "So what we’re doing is building this foundation at the foothills of all these mountains. We’re sending expeditions up to the mountains, doing explorations, and basically smelling the whole world.”

Deep Dive

Artificial intelligence

Don’t be fooled—LLMs don’t reason

Ten years after AlphaGo’s match against Go champion Lee Sedol, today’s AI still isn’t tapping into the machinery that made that win possible.

By

“We were looking at it wrong”: The explosive claim that’s upended the AI world

An exclusive interview with Nick Cammarata, the researcher whose 26,000-word viral blog post has AI safety folks celebrating.

By

Don’t be fooled by this summer of AI hype 

Breathless claims about AGI and new capabilities fall apart pretty quickly under scrutiny.

By

The agentic web is here

With the rise of personal AI assistants, the internet isn't just for humans anymore.

By

Stay connected

Illustration by Rose Wong

Get the latest updates from
MIT Technology Review

Discover special offers, top stories, upcoming events, and more.