Osmo says its AI can identify and design smells, drawing on millions of human labels and hundreds of millions of chemical “scent tokens”. The work points to a striking new use for machine learning, but claims of human-level ability and possible medical applications still need careful scrutiny.
Watch Desk analysis
What happened
The startup, spun out of Google Brain in 2023, is building a model that maps chemical compositions to human descriptions of odours. MIT Technology Review reports that Osmo has created a dataset with more than 10 million human labels and 800 million scent tokens, using chemical sensors alongside trained people who describe what they smell.
The system is already being used in fragrance work, according to founder Alex Wiltschko. The report describes a demonstration in which a prompt produced three proposed fragrance profiles and the molecules needed to make them. Osmo also says its technology has been used to create retail fragrances and to identify substances that repel disease-carrying insects. These are company claims reported by the magazine, not independent assessments of performance.
Why it matters
Giving computers a way to connect chemical measurements with human descriptions could make smell more useful in industries where it is currently difficult to measure consistently. Fragrance design is the clearest near-term example in the reporting. Osmo has also explored checking trainers for counterfeit materials by detecting differences in chemical composition.
The bigger possibilities, including detecting threats or disease, remain possibilities rather than demonstrated applications. Smell is unusually dependent on context: the same molecule can suggest very different things depending on what surrounds it. Michael Schmuker, a former professor who has consulted for Osmo, says that limitation could matter outside controlled settings such as a fragrance lab. Smelling a molecule is not quite the same as understanding what it means.
Our read
This is an inventive attempt to turn a famously slippery human sense into data a model can work with. The useful story is not that computers have acquired a nose, but that Osmo is building a large labelled dataset and putting it to work in fragrance design. That is already a more grounded proposition than the medical-diagnosis mountain on the horizon.
The reported results deserve interest, not a standing ovation just yet. Smell scientist Avery Gilbert has criticised some of the company’s marketing claims as undocumented, while the report notes that Osmo did not respond to a request for comment before publication. For now, readers should separate reported commercial uses from broader claims about what the technology may eventually do.
What to watch
- Whether Osmo publishes independent evaluations of its model’s performance.
- How well its approach handles smell outside controlled fragrance settings.
- Whether the reported uses in product design and counterfeit detection produce measurable results.
Discussion spark: Should an AI smell model be judged mainly by how well it performs in controlled fragrance work, or should claims about uses such as disease detection require a higher bar before they shape public expectations?
Sources and evidence
- This smell startup says its AI nose is as good as a human’s (18 November 2026, 10:20 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.