ModelsArchitectures & capability
Osmo says its AI smell model matches human noses and outperformed perfumers at a large consumer-goods company
Google Brain spinout Osmo says its smell model now detects and tells odors apart like an average person and designs scents like a junior perfumer. The model builds on its 2023 odor-mapping work, and critics question the company's patent and performance claims.
From the MIT Technology Review · AI article of November 18, 2026 and cited sources: Osmo uses its own definition of human-level smell. [1]
SOURCES[1] This smell startup says its AI nose is as good as a… technologyreview.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Osmo uses its own definition of human-level smell. [1]
- Original source
- MIT Technology Review · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2510.19660 (Oct 2025) [8] date from arXiv ID
- When
- November 18, 2026 article published
- Covered
- Nov 20 edition
- Evidence
- 8 sources across 8 sites
- 1This smell startup says its AI nose is as good as a…technologyreview.com
- 2AI Predicts What Chemicals Will Smell like to a Human —…scientificamerican.com
- 3Principal odor map news release — EurekAlerteurekalert.org
- 4The Great and Powerful Os(mo) — Avery Gilbert, Substackaverygilbert.substack.com
- 5A principal odor map unifies diverse tasks in olfactory…academicnewsletter.sufe.edu.cn
- 6OpenPOM — GitHubgithub.com
Reach: not measured yet

Osmo uses its own definition of human-level smell. Under that definition, the model detects and distinguishes odors about as well as an average person and designs scents about as well as a junior perfumer. The company also says the model beat in-house perfumers at a major consumer-goods firm last month. These results are reported by Osmo alone, and it has not named the firm or described how the test was run. [1]
The peer-reviewed basis is the 2023 Principal Odor Map. It is a graph neural network trained on about 5,000 perfumery-labeled molecules and tested in advance on 400 new odorants. It matched some trained panelists at describing odors, but only for single molecules, not mixtures. OpenPOM is an open reimplementation of this model. [2] [3] [5] [6]
The current commercial system turns a text brief into scent profiles and the molecules needed to make them, and a mixing robot produces samples in about a week. Osmo says the model is trained on more than 10 million human labels and 800 million scent tokens. Its architecture and benchmarks are not public, and Procter & Gamble is named as a customer. [1]
Critics dispute the company's framing. Avery Gilbert argues that Osmo presented patent applications as if they were granted patents and that its 10x success-rate claim has no supporting documentation. Michael Schmuker notes that the same molecule can be perceived differently depending on context, and that sensors must improve a lot before uses such as medical diagnosis. [1] [4]
The only peer-reviewed evidence covers single-molecule odor prediction from 2023. If the mixture-design claims hold up, fragrance R&D could move faster from brief to sample.
Read the full assessment
Buyers should first ask for evaluation protocols and run their own blinded sensory panels.
The sources are gathered; writing the dossier now.
Executive brief
Osmo, a startup spun out of Google Brain, says its smell model has reached "human-level olfactory intelligence." By the company's own definition, that means the model detects and tells smells apart about as well as an average person, and designs scents about as well as a junior perfumer. It also says the model beat in-house perfumers at a large consumer-goods company last month (MIT Technology Review). No independent evaluation of either claim has been published. The most recent peer-reviewed benchmark comes from Osmo's 2023 model.
What changed and event timeline
Preprint says AI predicts odor
The model's odor descriptions came closer to the panel average than a typical panelist did for 53% of molecules. It was tested on single molecules only, not mixtures ().
Principal Odor Map published in Science
Lee et al. describe a graph neural network that maps molecular structure to perceived odor. It matched some trained human sniffers at describing odor quality ().
Company says AI beat perfumers
Osmo says its model outperformed in-house perfumers at one of the largest consumer-goods firms. Neither the company nor the test design has been disclosed ().
Critic questions Osmo's claims
Smell scientist Avery Gilbert argues that Osmo described five published patent applications in a way that suggested granted patents, and that its "10x success rate" claim has no supporting documentation ().
"Human-level" claim reported
MIT Technology Review publishes Wiltschko's claim together with criticism from Gilbert and Michael Schmuker ().
Capabilities and access
- Model: Osmo calls it a "foundation model for smell." No name, version number or model card has been published.
- In the reporter's demo, the system returned three scent profiles plus the molecules needed to make each one.
- Procter & Gamble is named as one. A sneaker anti-counterfeiting pilot with StockX is also reported.
Read the full section
- Model: Osmo calls it a "foundation model for smell." No name, version number or model card has been published.
- Interface: Users type a text brief. In the reporter's demo, the system returned three scent profiles plus the molecules needed to make each one. A large mixing robot can then produce samples in about a week (MIT Technology Review).
- Access: Commercial only, for fragrance customers. Procter & Gamble is named as one. A sneaker anti-counterfeiting pilot with StockX is also reported.
Technical analysis for researchers and developers
- 2023 architecture: A message-passing graph neural network trained on about 5,000 molecules labeled with perfumers' odor descriptors.
- Current system: Osmo reports more than 10M human labels and 800M "scent tokens" (exact molecules and concentrations), drawn from spectrometers and about 70 trained labelers (MIT Technology Review).
- Reproducibility: OpenPOM, led by Schmuker, reimplements the 2023 model in DeepChem. It does not have access to Osmo's later data.
Read the full section
- 2023 architecture: A message-passing graph neural network trained on about 5,000 molecules labeled with perfumers' odor descriptors. It places molecules in a 256-dimensional space (Scientific American). The model was tested prospectively on 400 new odorants, each rated by a trained panel (Science summary).
- Current system: Osmo reports more than 10M human labels and 800M "scent tokens" (exact molecules and concentrations), drawn from spectrometers and about 70 trained labelers (MIT Technology Review). Its architecture, evaluation protocol and benchmarks are not public.
- Reproducibility: OpenPOM, led by Schmuker, reimplements the 2023 model in DeepChem. It does not have access to Osmo's later data.
Claims and evidence
- "Human-level" detection and discrimination: This is Osmo's own definition and has no published method behind it (MIT Technology Review).
- Beat in-house perfumers: Reported by the company only. No independent corroboration has been found.
- Odor-quality prediction on par with panelists: Peer-reviewed, but only for single molecules in 2023 (EurekAlert).
Read the full section
- "Human-level" detection and discrimination: This is Osmo's own definition and has no published method behind it (MIT Technology Review).
- Beat in-house perfumers: Reported by the company only. No independent corroboration has been found.
- Odor-quality prediction on par with panelists: Peer-reviewed, but only for single molecules in 2023 (EurekAlert).
- Eight molecules more repellent than DEET: Reported in Gates-funded work. Secondary coverage cites about 400 molecules screened (MIT Technology Review; citybiz).
- Patent leadership and "10x" success rates: Disputed by Gilbert (Substack).
Context and prior work
Before Osmo, odor databases held fewer than 200,000 human descriptions, according to Wiltschko (MIT Technology Review). The 2023 Principal Odor Map gave the field a shared coordinate system for odor, much as wavelength does for color (EurekAlert). Hardware for electronic noses and embedded AI for smell is a separate, active line of work (Mershin et al., arXiv).
Read the full section
Before Osmo, odor databases held fewer than 200,000 human descriptions, according to Wiltschko (MIT Technology Review). The 2023 Principal Odor Map gave the field a shared coordinate system for odor, much as wavelength does for color (EurekAlert). Hardware for electronic noses and embedded AI for smell is a separate, active line of work (Mershin et al., arXiv). Osmo's current pitch moves from predicting how one molecule smells to designing whole mixtures and using smell as an input channel for AI "world models."
Limitations, safety and contested findings
- Narrow early tests: The 2023 model was trained on perfumery data that leans toward pleasant smells. It was not tested on mixtures (Scientific American).
- Labels depend on language: Oxford's Asifa Majid notes that smell vocabularies differ across languages (same source).
- Missing context: Schmuker says the same molecule can mean different things depending on context; Camembert and used nappies share a compound.
Read the full section
- Narrow early tests: The 2023 model was trained on perfumery data that leans toward pleasant smells. It was not tested on mixtures (Scientific American).
- Labels depend on language: Oxford's Asifa Majid notes that smell vocabularies differ across languages (same source).
- Missing context: Schmuker says the same molecule can mean different things depending on context; Camembert and used nappies share a compound. He adds that sensors need to get much better before uses like medical diagnosis (MIT Technology Review).
- Human panels behind the patents: Four of Osmo's five patent filings rely mainly on human panels and Master Perfumer ratings (Gilbert). Wiltschko replies that human perception is the training signal, not a workaround.
Business and practitioner implications
- Fragrance R&D: Text-to-formula design could shorten early brief-to-sample cycles. Brands should still run their own blinded sensory panels before relying on it.
- Buyers: Ask for the evaluation protocol behind any "human-level" or "beat perfumers" claim. None has been published.
- Developers: OpenPOM is a usable open baseline for predicting odor from single molecules. Osmo's data advantage cannot be reproduced.
Read the full section
- Fragrance R&D: Text-to-formula design could shorten early brief-to-sample cycles. Brands should still run their own blinded sensory panels before relying on it.
- Buyers: Ask for the evaluation protocol behind any "human-level" or "beat perfumers" claim. None has been published.
- Developers: OpenPOM is a usable open baseline for predicting odor from single molecules. Osmo's data advantage cannot be reproduced.
- Security and health uses: Airport screening and disease diagnosis remain aspirational and depend on better sensors. Wiltschko himself calls diagnostics expensive and difficult.