AI Olfaction Reaches Human Level? Osmo's Model First Has to Tell Camembert and Used Diapers Apart
Osmo trained its foundational olfactory model on 10 million human-labeled scent data points, claiming its performance is nearly on par with average people, and has already developed custard-tart scented perfume, new mosquito-repellent molecules, and a counterfeit sneaker detection system. But critics note that scent perception is inseparable from context, and AI-powered olfactory medical diagnosis is still far from practical application.
In a lab near Newark Liberty International Airport in New Jersey, around 30 trained scent assessors spend their days smelling different items. They jot down descriptions like "nutty", "peachy", and "oceany", then cross-reference these notes with chemical analysis results. All these records are fed to the model of AI startup Osmo. As company founder Alex Wiltschko puts it, the goal is straightforward: give computers a working sense of smell.

Image credit: OSMO AI
Osmo recently claimed its model has achieved "human-level olfactory intelligence". The company defines this standard as: the ability to detect and distinguish scents is roughly equivalent to that of an average person, and the ability to create new scents is on par with entry-level perfume designers. This label, it should be noted, is self-defined by the company, not a universally accepted standard in the academic community. Last month, the company also announced its model outperformed professional in-house perfumers at a major global consumer goods corporation in a blind competition. Osmo's technology has already been applied in the perfume industry, and the company has previously collaborated with P&G on fragrance development. Earlier use cases include: the Bill & Melinda Gates Foundation first invested $5 million, then provided an additional $3.5 million in grants to fund the evaluation of mosquito-repellent molecules, and Osmo successfully identified 8 substances that repel mosquitoes more effectively than DEET; it also partnered with StockX to spot counterfeit sneakers by identifying differences in the chemical composition of shoe materials.
The biggest bottleneck for olfactory AI lies in data. According to Wiltschko's statistics, before Osmo, existing public databases contained fewer than 200,000 human-generated descriptions of odors. Chemical sensors are expensive, and the perfume industry has long kept olfactory research tightly under wraps. Osmo's solution is to build its own proprietary dataset: over 10 million human-labeled scent annotations, 800 million "scent tokens" that describe specific molecules and their corresponding concentrations. The company uses spectrometers and other chemical sensors to break down the molecular composition of fruits, flowers, pantry ingredients, and even biological tissues around the clock. Around 70 trained annotators based in the U.S. and overseas smell samples every day, and record their observations using a standardized set of professional fragrance terminology.
This massive dataset has yielded some surprising comparative insights. Wiltschko says the scent tokens that define the unique smell of a newborn baby overlap heavily with those of a fresh plate of just-cooked Barilla pasta. Even a tiny shift in molecular proportion can drastically change human scent perception, which explains why odors can be decomposed into independent tokens and then recombined by the AI model. Osmo also owns a scent-printing robot the size of a school bus that mixes raw materials according to extremely precise formulas. It can design and print custom scents based on prompts, similar to how ChatGPT generates text. At the Lisbon Web Summit, I asked Wiltschko to create a Portuguese egg tart scented perfume targeted at men aged 18 to 23. The system processed the request for a few minutes, then returned three design proposals with keywords including cream, coconut, and fruity, listing all the required molecules. If I placed a follow-up order, my "Egg Tart Water" perfume would be ready in roughly a week.
This sounds like a flashy demo, but it reveals the real core selling point of the system: it turns intangible scents into retrievable, combinable, and producible structured data. The perfume industry is just the easiest scenario to monetize this technology.
Criticisms have soon followed. Avery Gilbert, a olfactory scientist and author of What the Nose Knows, criticized Osmo on his Substack page, arguing that some of the company's marketing claims are not backed up by public documentation, and its patent filings still rely heavily on human sensory panels and professional perfumers. He wrote that this is somewhat strange for a company that positions itself around cutting-edge, super-capable AI. Osmo's representatives had not responded to requests for comment before press time.
Michael Schmuker, who previously served as a consultant for Osmo, says the company's technology is a genuine remarkable achievement. He and his team created an open-source reverse-engineered version of Osmo's 2023 model, but it lacks the massive volume of proprietary data the company has collected in subsequent years. He warns that scent perception never exists in a vacuum, detached from context. Camembert cheese and used diapers share some identical odor molecules, but humans will never mix the two up because they appear in completely different scenarios. In a controlled environment like a perfume lab, the lack of contextual cues may not cause major problems. But if we want to integrate olfaction as a "sensory channel" into general world models, for example, to spot spoiled food on production lines, or screen for tumors via scent as accurately as detection dogs, both sensors and models still have a long way to evolve.
Medical diagnosis is the biggest potential prize in this field. Wiltschko also acknowledges it is extremely challenging. He notes that many companies rush straight to pursue the top-tier medical use cases and end up failing, because the endeavor is prohibitively expensive and technically difficult. Osmo's strategy is to lay a solid foundation at the foot of the mountain first, send exploration teams up the slope step by step, and map the scent of the entire world first.
Original reporting: Amit Katwala, MIT Technology Review.
发布时间: 2026-11-18 18:20