Osmo Claims Human-Level AI Smell as Scientists Scrutinize Its Data Factory and Diagnostic Ambitions
Artificial intelligence conquered pixels and prose long before it could reliably tell jasmine from burnt rubber. On November 18, 2026, Google Brain spin-out Osmo claimed it has crossed a decisive threshold: founder and CEO Alex Wiltschko told [MIT Technology Review](https://www.technologyreview.com/2026/11/18/1148816/smell-startup-osmo-ai-nose-good-as-humans/) that the startup's Olfactory Intelligence platform has achieved "human-level olfactory intelligence," matching average humans at discriminating odors and junior perfumers at designing new scents. Yet beneath the milestone—and Osmo's rapid commercial expansion into consumer goods—sensory scientists and computational researchers remain sharply divided over how much of the feat is truly autonomous AI versus industrialized human labor. ### Cracking Smell's Data Bottleneck With Spectrometers and 70 Sniffers Unlike computer vision or large language models, computational olfaction never had an ImageNet or Common Crawl to train on. As Wiltschko explained, fewer than 200,000 reliable human smell descriptions existed across public datasets when Osmo spun out of Google Brain's research in 2022–2023, largely because chemical spectrometers are costly and legacy fragrance conglomerates guard their scent libraries as trade secrets. To break that bottleneck, [Osmo](https://www.osmo.ai/about) built an industrial-scale physical data factory: a 58,000-square-foot R&D and formulation facility in Elizabeth, New Jersey, backed by $60 million in Series A capital, a $70 million Series B in 2026, and a Scientific Advisory Board featuring Deep Learning pioneer Geoffrey Hinton. Around 70 trained human sniffers across the United States and abroad catalog substances using standardized descriptors—"nutty," "peachy," "marine"—paired around the clock with mass-spectrometry molecular breakdowns. That pipeline has generated more than 10 million human labels, 250,000 physical samples, 3 billion mapped molecules, and 800 million "scent tokens" capturing exact chemical compositions and concentrations. Rearranging those tokens shifts perception just as syntax shifts prose: Wiltschko noted that the molecular tokens behind the scent of a newborn baby overlap heavily with those of freshly cooked Barilla spaghetti. ### From Prompt-to-Perfume to Independent Skepticism Commercially, the model is already printing real-world molecules. At Web Summit in Lisbon, Osmo demonstrated a prompt-driven formulation box that translated a request for a "Portuguese custard tart" scent aimed at young men into three candidate chemical profiles ready for synthesis by a bus-sized mixing robot. Osmo says its platform outperformed in-house perfumers at a major consumer goods firm in October 2026, has shipped commercial fragrances for brands such as Procter & Gamble, discovered eight mosquito-repellent molecules stronger than DEET under a $3.5 million Gates Foundation grant, and piloted chemical sneaker-authentication checks with StockX. Still, independent olfactory experts warn against confusing commercial perfumery tooling with autonomous superhuman noses. In a pointed critique on [FirstNerve](https://averygilbert.substack.com/p/the-great-and-powerful-osmo), smell scientist Avery Gilbert scrutinized Osmo's headline claims, noting that the company's touted 2025 "patent" record actually rested on five published patent *applications* rather than granted patents—and that those filings continued to cite human sensory evaluation panels and master perfumers rather than purely algorithmic discovery. ### The Open-Source Gap and the Leap to Medicine Academic attempts to keep pace underline why data—not model architecture alone—remains the moat. Former University of Hertfordshire neural computation professor Michael Schmuker and colleagues built [OpenPOM](https://github.com/biomachinelearning/openpom), an open-source DeepChem replication of Osmo's landmark 2023 *Science* Principal Odor Map paper. While OpenPOM validates that graph neural networks can predict human odor descriptors from molecular structure, open-source teams cannot replicate Osmo's proprietary spectrometer-and-sniffer flywheel. Even so, Schmuker cautions that Osmo's next ambitions—disease diagnostics and airport threat detection—pose fundamentally harder challenges than blending retail perfumes. Human smell relies heavily on situational context: Camembert cheese and soiled diapers share key volatile molecules, yet human brains separate them instantly through visual and environmental cues. Capturing subtle disease biomarkers or trace explosives in noisy real-world air will demand breakthroughs in portable sensor hardware and clinical validation that go well beyond matching a junior perfumer in a New Jersey lab. ### Sources & Further Reading - [MIT Technology Review (US)](https://www.technologyreview.com/2026/11/18/1148816/smell-startup-osmo-ai-nose-good-as-humans/) - [Osmo Official Company Profile & Timeline (US — Primary)](https://www.osmo.ai/about) - [FirstNerve — Avery Gilbert Analysis (US — Independent Critique)](https://averygilbert.substack.com/p/the-great-and-powerful-osmo) - [GitHub — BioMachineLearning/openpom (Open-Source Replication)](https://github.com/biomachinelearning/openpom) ~AI ~Osmo ~Biotech ~MachineLearning ~Science