Osmo says its smell AI model matches human noses

MIT Technology Review profiles Osmo, an AI startup that calls its system the world's first foundation model for smell. Founder Alex Wiltschko says the goal is "to give computers a sense of smell." In the last few months, he claims, the model has reached "human-level olfactory intelligence." The company defines that as being 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.
The company says its technology is already used in the fragrance industry to help brands design and make new scents. It also says that last month its model outperformed in-house perfumers at one of the world's largest consumer goods companies. The article does not name the company or give test details. Wiltschko believes the technology could eventually be used for threat detection at airports or for diagnosing diseases.
Wiltschko spun Osmo out of his work at Google Brain in 2023 and is following the playbook of AI startups in other fields: throw a lot of data at a hard problem. His argument is that smell had too little good data. By his count, existing databases held fewer than 200,000 human descriptions of smells before Osmo, the chemical sensors needed to gather data are expensive, and the fragrance industry is secretive about its olfactory science.
Osmo built 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 specific odors. Spectrometers and other chemical sensors run around the clock to break down everything from fruit and flowers to pantry staples and organs. In a lab near Newark Airport in New Jersey, around 30 experienced human sniffers work through odors, and their notes ("nutty," "peachy," "marine") are cross-referenced with chemical analysis. A further team of around 70 trained data labelers in the US and overseas sniff things and write down what they smell like, using a standardized fragrance vocabulary.
Mapping scents this way produces odd parallels. Wiltschko compares it to word order in a sentence: rearranging molecules in different proportions changes what people perceive. He says the scent tokens for a newborn baby overlap massively with those of a fresh plate of cooked spaghetti (Barilla, specifically).
Osmo also makes its own scents with a smell-printing robot the size of a school bus. It mixes ingredients to precise formulas, producing samples for Osmo's research and fragrances for paying customers, and can design and print scents to order from a ChatGPT-style prompt. The reporter, in Lisbon for Web Summit, asked for a fragrance resembling a Portuguese custard tart, aimed at 18- to 23-year-old men. After a few minutes the system produced three scent-profile options (creamy, coconutty, fruity) with the molecules needed to make them. Had the order continued, the fragrance would have been ready about a week later.
The article says Osmo's technology has already been used to make retail fragrances, including for Procter & Gamble. The Gates Foundation invested $5 million when Osmo launched and later gave it a $3.5 million grant to assess molecules that repel disease-carrying insects; Osmo identified eight substances more repellent than DEET. Osmo has also worked with StockX, the online clothing marketplace, on spotting counterfeit sneakers by detecting differences in the chemical composition of the shoe materials.
Not everyone is convinced. Avery Gilbert, a smell scientist and author of What the Nose Knows, has criticised Osmo sharply on his Substack. He argues 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, calls the technology an achievement. With colleagues he built an open-source, reverse-engineered version of Osmo's 2023 model, which lacks the data the company has collected since. But he says Osmo may be missing some of the factors that make human olfaction unique. "Smell is super complex. Context is very important for smell," he says, noting that Camembert shares part of a molecule with used nappies, and people do not confuse the two because of context.
In a controlled setting like a fragrance lab, that lack of context may not matter much. It could matter if the technology becomes what Wiltschko calls the "olfactive channel" for AI world models, for example in a machine that sniffs out food spoilage on a production line or detects tumors the way dogs have been trained to in pilot studies. For that, Schmuker says, sensor technology will need to get a lot better and the models will need to keep improving. Medical diagnostics is the biggest prize. Wiltschko acknowledges that companies which went straight for that peak have failed because it is expensive and difficult, so Osmo is building a foundation at the foothills and sending expeditions up the mountains.
Key facts
- Osmo founder Alex Wiltschko claims the company's smell foundation model has reached "human-level olfactory intelligence": roughly average-person level at detecting and discriminating smells, and junior-perfume-designer level at creating them.
- The model is trained on a dataset of more than 10 million human labels and 800 million scent tokens, built with around 30 expert sniffers, around 70 data labelers and round-the-clock chemical sensors.
- Osmo says it is used in fragrance design, including retail fragrances for Procter & Gamble, and that last month it outperformed in-house perfumers at one of the world's largest consumer goods companies (unnamed).
- Other work: a $3.5 million Gates Foundation grant led to eight substances more repellent than DEET, and a StockX project on spotting counterfeit sneakers by smell.
- Critic Avery Gilbert calls some marketing claims undocumented, and Michael Schmuker says Osmo may lack the context that shapes human smell; medical diagnostics and airport threat detection remain future hopes.
Why it matters
Osmo is applying the data-at-scale approach that worked in other AI fields to smell, a sense where, in Wiltschko's account, good data was scarce (fewer than 200,000 human descriptions in existing databases, by his count). The company says the result already works commercially in fragrance. If the claim holds up, it would be a rare case of a foundation model for a sense other than sight, sound and text finding paying customers. The larger ambitions, medical diagnostics and threat detection, are still ambitions.
Who it affects
Fragrance brands and consumer goods companies are the current users; Procter & Gamble is named as a customer for retail fragrances. Perfumers are affected too, since Osmo compares its model to a junior perfume designer and says it outperformed in-house perfumers at one large consumer goods company. Other named work touches public health researchers (insect repellents, via the Gates Foundation grant) and online marketplaces such as StockX (counterfeit sneakers). Medical and security users are possible future audiences only.
How to use it
The article describes a service rather than a product anyone can download. Osmo designs scents from text prompts and makes them with its school-bus-sized smell-printing robot, supplying fragrances to paying customers. In the reporter's demo, a prompt for a Portuguese custard tart scent aimed at 18- to 23-year-old men produced three scent-profile options in a few minutes, and the fragrance would have been ready about a week after ordering. The article gives no pricing or access terms.
How solid is it
The central claim is the company's own. "Human-level olfactory intelligence" is Wiltschko's phrase and Osmo's definition, and the article describes no independent evaluation of it. The outperformed-perfumers result is attributed to the company, with the customer unnamed and no method or numbers given. The fragrance use for Procter & Gamble, the insect-repellent finding and the StockX work are stated by the article. Schmuker, who has consulted for Osmo, calls the technology an achievement, and his open-source reconstruction of the 2023 model shows part of it can be reproduced, though without the newer data. Osmo did not respond to Gilbert's criticism in time for publication.
Risks and caveats
Gilbert argues that some marketing claims are undocumented and that Osmo's patent filings still rely on human sensory panels and perfumers rather than AI. Schmuker warns that Osmo may be missing factors that make human olfaction unique, notably context: the same molecule can evoke different associations. That may matter little in a fragrance lab but more if the model becomes the olfactory channel for AI world models. Sensing spoilage or tumors would need much better sensor technology and continued model improvement, and no timescale is given. The tumor-sniffing comparison with dogs refers only to pilot studies. Wiltschko himself notes that earlier companies aiming straight at the hardest targets failed because it is very expensive and very difficult.
“This heavy reliance on human sensory judgments seems odd for a company premised on super-cool AI”
— Avery Gilbert, smell scientist and author of What the Nose Knows, on his Substack in March