Osmo says its AI outperformed an unnamed consumer-goods company’s in-house perfumers in October. The claim, reported in a November 18 MIT Technology Review interview with founder Alex Wiltschko, comes alongside his assertion of human-level smell intelligence. But that label has a specific company definition, and scientists raise questions about both documentation and real-world sensing.
Speaking to reporter Amit Katwala, Wiltschko said Osmo’s model had reached what the company calls “human-level olfactory intelligence” in recent months. Osmo defines that as roughly matching an average person at detecting and distinguishing smells, while matching a junior perfume designer at creating them. Those are the company’s capability claims, not an independently verified finding.
Building smell data by hand—and nose
Wiltschko spun Osmo out of his work at Google Brain in 2023. A central obstacle, he says, was a shortage of training material: existing databases contained fewer than 200,000 human descriptions of smells, by his count. Chemical sensors are expensive, and fragrance companies closely guard their olfactory research, making the raw material for a smell model difficult to assemble.
Osmo says it has since built a dataset containing more than 10 million human labels and 800 million “scent tokens.” The tokens describe the molecules and concentrations that make up odors. Spectrometers and other chemical sensors run around the clock, analyzing materials ranging from fruit and flowers to pantry ingredients and organs.
Human judgment remains part of that process. Around 70 trained labelers in the United States and overseas smell samples and describe them using a standardized fragrance vocabulary. Their descriptions connect a chemical measurement to the way people perceive it. The training operation therefore combines instrument readings with a substantial amount of human sensory work.
A text prompt becomes a scent formula
Katwala saw the design interface during a demonstration in Lisbon. He requested a fragrance resembling a Portuguese custard tart, aimed at men aged 18 to 23. After a few minutes, the system returned three possible scent profiles and the molecules needed to make them. He did not complete the order; the finished fragrance would have been ready about a week later.
Turning a formula into a sample requires physical machinery. Osmo uses a school-bus-sized mixing robot to combine ingredients precisely for research and paying customers. Its technology has already helped create retail fragrances, including work for Procter & Gamble. That named customer should not be confused with the unnamed company in Osmo’s October performance claim.
Smell is super complex. Context is very important for smell.
Michael Schmuker, a neural-computation researcher who previously consulted for Osmo
The limits of the human comparison
Smell scientist Avery Gilbert has challenged Osmo’s marketing on his Substack, arguing that some claims are undocumented. In March, he also pointed to patent filings that relied on human sensory panels and perfumers, questioning that reliance for a company built around AI. An Osmo representative did not respond to MIT Technology Review in time for publication.
Michael Schmuker, a former University of Hertfordshire professor who previously consulted for Osmo, calls its technology an achievement. But he warns that chemical similarity is not the whole experience of smell. Camembert and used nappies partly share an odor molecule, he notes; people distinguish them through context. A controlled fragrance lab may make that limitation less important than other settings would.
Wiltschko’s ambitions extend to airport threat detection and medical diagnosis, but those remain future possibilities. Schmuker says reaching such uses will require substantially better sensors as well as better models.
Wiltschko himself describes medical diagnostics as expensive and difficult, acknowledging that earlier companies have failed at it. Osmo’s current commercial work is in fragrance; diagnosing disease is a different, unfinished climb.
Reader comments
Newest comments first. Replies stay oldest first.