This smell startup says its AI nose is as good as a human’s

MIT Technology Review
Read full story· 6 min
Osmo, an AI startup, has developed a foundation model for smell that matches human olfactory ability, aiding fragrance design and potentially useful for threat detection and medical diagnosis. The model was trained on a proprietary dataset of over 10 million human scent labels and 800 million molecular scent tokens.

What happened

Osmo, an AI startup based near Newark Airport in New Jersey, has developed a foundation model for smell that it says matches human ability to detect and create scents. The company built its model using a proprietary dataset collected by a team of trained sniffers and chemical sensors, analyzing millions of odors. Osmo's technology is already being used in the fragrance industry, and the company claims its AI has outperformed in-house perfumers at a major consumer goods company. The model can also design custom scents and has been used in projects ranging from insect repellent research to counterfeit sneaker detection.

Key details

  • Osmo's dataset includes over 10 million human scent labels and 800 million molecular scent tokens.
  • The company employs around 30 human sniffers and 70 data labelers for scent analysis.
  • Osmo was spun out of Google Brain in 2023 by founder Alex Wiltschko.
  • The Gates Foundation invested $5 million and later provided a $3.5 million grant for insect repellent research.
  • Osmo's technology has been used by Procter & Gamble and StockX.

Why it matters

Fragrance companies and consumer goods manufacturers can use Osmo's AI to design new scents more efficiently and potentially reduce reliance on human perfumers. Organizations involved in public health and safety, such as those working on disease prevention or counterfeit detection, gain a new tool for analyzing chemical signatures. The availability of this technology introduces new capabilities for industries that depend on scent analysis, with concrete applications in product development and quality control.


More in Machine Learning