Skip to content
Theme
OnDevice Tools

AIAI Models & Research

Osmo says its AI smell model has reached human-level scent detection

The startup Osmo claims its foundation model for smell can detect and distinguish odours as well as an average person. Independent scientists question parts of that claim.

Illustration of an AI system analysing scent molecules beside aperfume bottle and chemical sensors

Key takeaways

  • Osmo claims its AI smell model matches an average person at detecting and distinguishing odours and works like a junior perfumer when creating scents.
  • The company built its own dataset of over 10 million human labels and 800 million scent tokens using sensors and teams of trained sniffers.
  • Its technology is used in retail fragrances, including with Procter & Gamble, and it identified eight insect repellents more effective than DEET for the Gates Foundation.
  • Independent scientists question parts of the claim, noting that patents still rely on human panels and that context is central to how people experience smell.
In this article
  1. What the startup claims
  2. Building a training set for smell
  3. Printing scent to order
  4. Commercial uses so far
  5. Why scientists are not fully convinced
  6. What comes next

What the startup claims

Osmo, a startup working on artificial scent, says its model has reached what it calls human-level olfactory intelligence. According to MIT Technology Review, founder Alex Wiltschko defines this as the system detecting and telling apart smells about as well as an average person, and creating scents roughly as well as a junior perfume designer.

The company says its software is already used by fragrance firms to help design and make new scents. It also claims that last month the model outperformed the in-house perfumers at one of the world's largest consumer goods companies. That comparison is the company's own and has not been independently verified.

Building a training set for smell

Illustration: AI, business, building, plane and the labels Fragrance, Gates Foundation, Osmo

The biggest obstacle, Wiltschko told MIT Technology Review, was a shortage of usable data. By his count, fewer than 200,000 human descriptions of smells existed in databases before Osmo began. Chemical sensors are costly, and the fragrance industry keeps much of its olfactory research secret.

Osmo's answer was to build its own dataset. Spectrometers and other sensors run around the clock, breaking down the chemistry of fruit, flowers, pantry staples and other materials.

A team of about 70 trained labelers in the United States and overseas spends its days smelling samples and describing them with a standard fragrance vocabulary. In a lab near Newark Airport in New Jersey, around 30 experienced sniffers log notes such as nutty, peachy and marine, cross-checked against chemical analysis.

Printing scent to order

Osmo also makes its own smells. A robot the size of a school bus mixes ingredients to precise formulas, producing samples for research and fragrances for paying customers. The company says scents can be designed from a prompt in the style of a chatbot.

In a demonstration reported by MIT Technology Review at Web Summit in Lisbon, the system produced three scent profiles, along with the molecules needed to make them, matching a brief for a custard-tart fragrance aimed at young men. Wiltschko said a finished fragrance would take about a week.

Commercial uses so far

The technology has been used to create retail fragrances, including work with Procter & Gamble. The Gates Foundation invested $5 million when Osmo launched and later awarded a $3.5 million grant to study molecules that repel insects carrying disease. Osmo says it identified eight substances more repellent than DEET.

The company has also worked with the online marketplace StockX on spotting counterfeit sneakers by detecting chemical differences in the materials. Wiltschko believes scent-based AI could eventually help with airport threat detection or with diagnosing disease.

Why scientists are not fully convinced

Not everyone accepts the company's claims. Smell scientist Avery Gilbert, author of What the Nose Knows, has criticised Osmo on his Substack, arguing that some marketing statements are undocumented and that its patent filings still depend on human sensory panels and perfumers rather than AI. MIT Technology Review reports that a representative for Osmo did not respond before publication.

He and colleagues built an open-source, reverse-engineered version of Osmo's 2023 model, though it lacks the company's later data.

He notes, however, that context matters greatly in human smell. The same molecule appears in Camembert cheese and in used nappies, yet people tell the two apart because of the setting. That could limit machines in uncontrolled environments, even if a fragrance lab is forgiving.

What comes next

Wiltschko has suggested the model could become an olfactory channel for AI world models, sensing spoiled food on production lines or tumours the way dogs have in pilot studies. Schmuker says that for such uses, sensor hardware and the models themselves must improve considerably, and medical diagnostics remains the hardest target.

Wiltschko describes the strategy as building foundations first. He says many companies have attacked diagnostics directly and failed because it is expensive and difficult, so Osmo is instead trying to catalogue the entire sensory world of smell before climbing towards those applications.

Sources

  1. This smell startup says its AI nose is as good as a human’sMIT Technology Review AI

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.