Computers are learning to smell: Osmo says its AI has reached human level
Smell was the one sense computers lacked. The US start-up Osmo says its model now tells smells apart as well as an average person, and creates them like a junior perfumer.
What happened
Computers have been able to see and hear for years, but they could not smell. The US start-up Osmo now says its AI model can do it as well as a person.
Its founder, Alex Wiltschko, who used to do AI research at Google Brain, told MIT Technology Review this week. In the past few months, he says, the model reached what Osmo calls ‘human-level olfactory intelligence’: telling smells apart like an average person, and creating them like a junior perfume designer. Last month, according to Osmo, the model outperformed the in-house perfumers of one of the world’s largest consumer goods companies. Osmo does not say which company, or how the comparison was made.
The secret is the data. At a lab near Newark airport, some thirty experienced sniffers smell substances, while about seventy trained labellers describe scents using a standard fragrance vocabulary. That has given Osmo more than 10 million human smell labels and 800 million ‘scent tokens’, recording which molecules a scent contains and at what concentration. Before Osmo, by Wiltschko’s count, existing databases held fewer than 200,000 descriptions. A robot the size of a school bus mixes the ingredients into samples and fragrances for customers.
Osmo has already made retail fragrances, including for Procter & Gamble, and helped the marketplace StockX spot counterfeit sneakers by their smell. With money from the Gates Foundation it searched for mosquito repellents and, it says, found eight substances that work better than DEET. The reporter got to try it in Lisbon: he asked for a scent like a pastel de nata, the Portuguese custard tart. The system came back with a creamy, a coconutty and a fruity direction.
Not everyone is convinced. Smell scientist Avery Gilbert calls some of Osmo’s claims undocumented and points out that its patent applications still rely on human panels. Wiltschko’s reply: “human perception isn’t a workaround for AI.” Michael Schmuker, a researcher who has advised Osmo, calls it an achievement but points out that context matters a great deal with smell: Camembert and a used nappy share a molecule, yet nobody mixes them up.
Computers have been able to digitize vision and hearing, but not smell: our deepest and oldest sense.
Alex Wiltschko in a Monell Center press release, August 2023
The background
Why smell was so hard for computers
Sight and hearing rest on clear physics: colour is a wavelength, pitch a frequency, and a screen makes every colour from red, green and blue. Smell has no such recipe. We see colour with three types of receptor, Google Research wrote; the nose has about 400, according to the Monell Center. A molecule’s structure is also a poor guide to how it smells, and there is no camera or screen for smell. There is not even an objective truth to test a model against, smell researcher Joel Mainland said. All you have is human judgement.
Electronic noses go back to the early 1980s, when George Dodd and Krishna Persaud at the University of Warwick came up with an artificial nose: sensors that each react to many gases, with software that recognises a smell in the pattern. Thousands have been sold, for quality control in food and drink among other things, but such a nose mainly recognises the smells it was trained on.
A map of every smell
Osmo’s work began at Google. From 2019, a team around Wiltschko trained a neural network that reads a molecule as a network of atoms and bonds, on more than 5,000 fragrance molecules described by perfumers. The result was a smell map, the Principal Odor Map, with more than 250 dimensions: molecules that smell alike sit close together, much as an RGB value pins down a colour.
The test appeared in Science in 2023, the year Osmo spun out of Google. Fifteen trained panellists at the Monell Center rated 400 new molecules using 55 words, from minty to musty. For 53% of the molecules, the model came closer to the group’s verdict than the average panellist. Early this year Osmo raised another $70 million.
A Dutch lab tested the mosquito repellents
DEET, the best-known mosquito repellent, came out of American research during the Second World War to protect soldiers from malaria. Google found the results of those old tests, thousands of substances recorded on paper, in its scanned books. Wiltschko’s team trained a model on them, and TropIQ Health Sciences, a spin-off of Radboudumc in Nijmegen, tested the candidates on real mosquitoes. Some did better than DEET, Google reported at the end of 2022. The Gates Foundation funded the work. The stakes are high: according to the WHO, there were about 282 million cases of malaria and 610,000 deaths in 2024.
Our analysis
An AI nose is as good as the people who taught it to smell
The most striking thing about Osmo is not the model: Schmuker and colleagues rebuilt the 2023 version as open source. What nobody can rebuild is the data: since 2023, people with trained noses have been recording what they smell, in a fixed vocabulary. That the patents still lean on human panels, as Gilbert notes, is therefore no side issue but the heart of it: human judgement, carefully recorded, is the raw material.
‘Human level’ is Osmo’s own yardstick, though: an average person at telling smells apart, a junior perfumer at making them. MIT Technology Review mentions no published test that measures it. Our expectation: that test will come once Osmo moves into quality control or healthcare, where customers ask for hard numbers.
The lesson reaches well beyond smell. An AI model learns what its examples teach it, and in an organisation those examples sit in the heads of your experts: the colleague who sees at once that an invoice is wrong, or which customer email is urgent. Record that knowledge in fixed categories, with enough examples, and you already have the most important part of a good AI model.
What stands out
- First came the data, then the model Before Osmo, by Wiltschko’s count, there were fewer than 200,000 smell descriptions. The company now has more than 10 million.
- The model learned more than it was asked The 2023 model also predicted how strongly a substance smells, even though it was never trained on that, the Monell Center reported.
- The smell map works for insects too A network that learned how molecules smell to people also predicted how the smell-processing part of an insect’s brain responds to them, Google found.
What it means for you
An AI model learns what your examples teach it. That makes the knowledge of your own experts the most valuable thing you have for AI.
- Record what your experts recognise What an experienced colleague sees at a glance is exactly what a model needs to learn. Keep examples together with the verdict on them.
- Agree on fixed words Osmo’s labellers describe smells from a single vocabulary. Do the same with your categories, or a model learns from noise.
- Test on what the model has never seen Set part of the examples aside and compare the results with your own experts before you rely on it.
Where we come in · Advisory
We start from the knowledge you already have
In a free AI strategy session, an adviser from Flaire maps out with you and the colleagues who know the work where AI saves time, and shows live what AI does with your own examples, such as a customer email, a quote or an invoice. Where a model of your own pays off for one task, we develop it with AI Model Engineering: on examples labelled by fixed rules, and tested on data it has never seen.
Which of your experts’ knowledge would you like to capture for AI?
Tell us which work now leans on the judgement of a few experienced colleagues. We’ll think it through with you and always reply within one working day.