Osmo Claims "Human-Level" AI Olfactory Intelligence—Why Fragrance Works While Medical Diagnostics Lag
While tech giants pour billions into text, video, and robotic manipulation, Google Brain spinout **Osmo** is betting on a sensory modality AI has largely ignored: smell. In a November 18, 2026 report from *MIT Technology Review*, founder Alex Wiltschko claimed the company’s olfactory foundation model has now reached what Osmo defines as "human-level olfactory intelligence"—matching average human odor discrimination and performing on par with a junior perfumer—while outperforming in-house fragrance teams at a major consumer goods company in a recent test. Yet beneath the commercial headlines, independent smell scientists caution that turning digital olfaction from a perfume-design assistant into a reliable detector for disease or security threats faces severe chemical, hardware, and contextual hurdles. ### How Osmo Built an 800-Million-Token "Olfactory Dataset" When Wiltschko spun Osmo out of Google Brain in 2023, the fundamental bottleneck was not model architecture—it was data scarcity. According to Wiltschko, public scientific databases held fewer than 200,000 human smell descriptions, while chemical spectrometers remained costly and major fragrance houses guarded their proprietary formulas closely. To break that bottleneck, Osmo built its own data factory: a 60,000-square-foot facility in Elizabeth, New Jersey, anchored by roughly 30 experienced human sniffers near Newark Airport (part of a roughly 70-person global labeling team). Around-the-clock spectrometers break down raw samples—from fruits and flowers to pantry staples—into molecular profiles, while human labelers log standardized descriptors like "nutty," "peachy," or "marine." That pipeline has produced more than **10 million human labels** and **800 million "scent tokens"** mapping exact molecular concentrations to perceived odors. Much as token order transforms meaning in language models, tiny shifts in molecular proportions radically alter human smell perception—creating counterintuitive overlaps in latent space (Wiltschko noted that the scent tokens for a newborn baby overlap heavily with fresh cooked spaghetti). Using a school-bus-sized scent-printing robot, Osmo can now take a text prompt—demonstrated to *MIT Technology Review* at Web Summit in Lisbon with a request for a Portuguese custard tart fragrance aimed at young men—and formulate physical sample options in roughly a week. ### Where It Works Today—and Why Critics Push Back Commercially, Osmo’s clearest traction sits inside controlled industrial labs rather than messy public spaces. The company has worked with **Procter & Gamble** on retail fragrances, partnered with **StockX** to detect counterfeit sneakers via volatile chemical signatures, auctioned proprietary AI-developed fragrance ingredients, and used a $3.5 million Gates Foundation grant (following an earlier $5 million launch investment) to identify eight insect-repellent molecules stronger than DEET. Still, independent researchers urge skepticism toward broad "human-level" framing. Smell scientist **Avery Gilbert**, author of *What the Nose Knows*, has pointed out on his Substack that several of Osmo’s boldest public claims lack peer-reviewed documentation and that its own patent filings still lean heavily on human sensory panels and perfumers rather than autonomous AI judgment. Neural computation researcher **Michael Schmuker**—formerly of the University of Hertfordshire, who previously consulted for Osmo and co-built an open-source reverse-engineered version of its 2023 model—highlights an even deeper scientific obstacle: **context**. Identical volatile molecules can register completely differently depending on environment and human expectation; French Camembert cheese and soiled diapers, for instance, share overlapping chemical compounds that a nose separates only through surrounding cues. ### The Long Road to Medical Diagnostics That context gap, paired with physical chemical-sensor limits, explains why Wiltschko himself compares medical diagnostics and airport threat detection to distant mountain peaks while Osmo builds revenue in fragrance "foothills." Unlike digital cameras or microphones, today’s portable chemical sensors remain slow, expensive, and easily overwhelmed in open air. Bridging AI scent models from controlled lab formulation into real-world clinical screening will require major hardware breakthroughs alongside vastly richer biological datasets—making Osmo’s current wins promising in industrial chemistry, even as autonomous AI "noses" in hospitals remain years away. ### Sources & Further Reading - [MIT Technology Review (US): This Smell Startup Says Its AI Nose Is as Good as a Human’s](https://www.technologyreview.com/2026/11/18/1148816/smell-startup-osmo-ai-nose-good-as-humans/) - [Osmo Official Site (Primary): About Osmo & Digital Olfaction](https://www.osmo.ai/about) - [Cosmetics Business (UK): Osmo to Auction AI-Developed Fragrance Ingredients](https://cosmeticsbusiness.com/osmo-to-auction-ten-ai-developed-fragrance-ingredients) ~AI ~Osmo ~Biotech ~MachineLearning ~Science