
Your computer can describe a mango without noticing that the one beside your keyboard has gone bad. It can write a perfume advertisement without knowing what is coming out of the bottle. That gap is easy to forget because a chatbot’s descriptions can sound so convincing. Ask it about freshly brewed coffee and it’ll give you a paragraph. Put a cup in front of it and the smell contributes nothing unless some other system measures it.
Osmo is working on that missing input. In a November 18 report for MIT Technology Review, Amit Katwala describes the startup’s claim that its technology has reached “human-level olfactory intelligence.” Founder Alex Wiltschko says it can match an average person’s smell detection and discrimination, and a junior perfumer’s fragrance creation. Those are company claims reported by the publication, rather than proof that a machine can handle everything your nose does. Read the reporting.
I think this is a more interesting AI story than another chatbot learning to write emails. It involves chemicals, human judgments and machines that have to do something in the physical world. But “AI can smell” also hides several different jobs inside one sentence. Pull those jobs apart and the achievement gets easier to understand. So do the remaining problems.
Access without medium partner: AI Learned to Talk. Osmo Wants to Smell
What exactly has learned to smell?
Start with a distinction that gets lost in the headline: predicting an odor from a chemical structure is different from detecting an unknown smell in a room. Designing a recipe for a fragrance is another task. Recreating that recipe requires physical ingredients and equipment. A system could do one of these well and struggle with the others.
The published science gives us a starting point. A 2023 paper in Science describes a Principal Odor Map, built with graph neural networks. The researchers trained on 5,000 molecules with odor descriptions. In a prospective study involving 400 novel odorants and a panel, the model’s predictions matched the panel’s average descriptions more closely than the median panelist did. That is evidence for predicting odor qualities from molecular structure under the study’s conditions. The paper is available here.
The comparison needs its full sentence attached. Matching a panel average on a defined task doesn’t mean beating every person, identifying every mixture, or understanding someone’s emotional response to a fragrance. The average is also the reference being predicted. A participant who experiences a molecule differently can disagree with that reference without having a defective nose.
That’s where I initially had to slow down. “Human-level” sounds like a single finish line. In practice, the result depends on which human, which task and which scoring method we choose. An average shopper, a trained evaluator and a perfumer aren’t interchangeable benchmarks. Neither are describing a known chemical and investigating an unfamiliar odor drifting through a warehouse.

The training data needs actual noses
Katwala reports that Osmo says it has collected more than 10 million human labels and 800 million “scent tokens.” The latter represent molecules and concentrations in its system. They shouldn’t be treated as equivalent to text tokens, or as 800 million independently tested smells. His reporting also describes trained people recording odor descriptions alongside chemical measurements.
There is a discrepancy worth keeping visible. Osmo’s public website, checked for this article, displays five million human annotations. That differs from the figure in the newer reporting. The pages don’t establish a shared counting method or reporting date, so I wouldn’t turn them into a growth chart. Osmo’s website is evidence of what the company publicly presents, not an independent audit of either number.
And the human work isn’t an embarrassing detail. If the thing you’re predicting is human perception, you need observations of human perception. A list of chemical structures cannot tell you how a group of people describes those chemicals. Someone has to make that connection and check whether the model’s predictions hold up.
The more useful questions concern the quality of those observations. Were the evaluators trained in a consistent vocabulary? Did they rate samples more than once? How much did their answers vary? Did the test include examples that were meaningfully different from the training material? A large dataset can contain repeated judgments, and repeated judgments can be useful, but they don’t automatically provide broad coverage.
I’d rather read a careful account of a smaller, well-controlled evaluation than a giant number with no explanation of its unit. “Scent token” is a memorable phrase. It still needs a definition beside it.
A recipe has quantities, not just ingredients
Here’s an illustrative example, separate from Osmo’s results. Suppose two formulas contain the same three ingredients, which we’ll call A, B and C. The first is mostly A, with a little B and very little C. The second reverses the amounts of A and C. An ingredient checklist would call them identical. A recipe would not.
The animation shows that simple distinction. Its numbers are invented for explanation; it doesn’t predict how either mixture would smell. That’s a useful boundary because a tidy chart can make a scientific explanation appear more settled than it is. Drawing two bars is easy. Establishing the relationship between a chemical mixture and someone’s perception takes experiments.
For a fragrance tool, this creates a practical feedback loop. Generate a candidate formula, make a sample, have people evaluate it, and revise. A text description can help start the process. It can’t replace the physical sample that a customer eventually judges.
Consider a hypothetical brief: a small brand wants a fragrance that feels like a quiet bookshop after rain. There isn’t one objectively correct answer. Different people may picture different shops, paper, wood or weather. The tool needs a way to translate that vague request into candidates, then absorb feedback without pretending the original words specified a unique chemical solution.
This is also why I don’t find the idea of AI-assisted perfumery absurd. Plenty of useful tools begin with incomplete instructions. The value would come from how quickly and consistently they help someone reach a satisfactory result. The perfume still has to survive the moment somebody opens the bottle.
You can’t send a smell through Wi-Fi alone
Osmo’s earlier “scent teleportation” work makes the physical process clearer. In an October 2024 explanation, the company describes capturing a sample, using gas chromatography–mass spectrometry to obtain chemical data, processing that data, then sending a formula to a formulation robot. For a solid object, its account describes sampling the air around it. Osmo explains the workflow here.
The transmitted part is information. The receiving equipment supplies the materials and mixes them. Your phone doesn’t acquire a smell output because a file arrives, any more than receiving a bread recipe produces a loaf in its charging port. Somewhere, physical work still has to happen.
I like that comparison because it makes the engineering easier to question. What did the sample capture? What could the instrument identify? Which ingredients were available to the receiving machine? How was the recreated odor judged against the original? Each stage can succeed or fail separately. A good model cannot recover information that never reached it. A good formula cannot compensate for a missing ingredient. And a pleasant result isn’t necessarily a faithful reproduction.
For a demonstration, a chosen sample and a controlled environment may be reasonable. For a product, I’d want repeated tests with samples chosen by someone outside the company. I’d also want the failures included. Otherwise, the audience sees a performance while the engineers see a much longer spreadsheet.

Perfume is a sensible place to start
Osmo’s current commercial offering is more grounded than the broadest AI-nose headlines suggest. Its website describes generating fragrance formulas from a brief, shipping physical samples, collecting feedback and working with fragrance advisers. It also puts perfumers inside the process. That is an identifiable service with a customer who can judge the result.
The business case doesn’t require proving that machines experience smell. It requires showing that the service helps customers develop fragrances at an acceptable cost and quality. Faster iteration could matter. So could making a previously difficult process accessible to a smaller brand. Those are business possibilities, though, rather than performance claims established by this article.
My first comparison would be against the same job done without the AI tool. Give both teams the same brief and budget. Hide which samples came from which process. Ask target customers to rate them, then track revisions and total time. Count the unsuccessful candidates as well as the chosen one. If the AI-assisted process produces comparable results with less work, that’s useful even if nobody calls it human-level.
I’d also test whether it follows exclusions and responds to corrections. A system that produces a lovely first sample but repeatedly ignores the customer’s requested change can waste everyone’s time. The second and third attempts may tell us more about its usefulness than the opening demonstration.
There are critics worth reading. Smell scientist Avery Gilbert has questioned how clearly Osmo’s marketing establishes the contribution of AI, including in a March 2026 examination of its patent-related claims. His criticism should be read as criticism, not a verdict that the technology cannot work. His argument is here.
My position is that human involvement doesn’t settle the argument either way. A tool can help experts without replacing them. What would settle more of it is a comparison that shows what changed when the tool was added, with the time, cost and rejected attempts included.
A useful detector needs more than a good headline score
The next step beyond fragrance raises a different set of questions. Suppose a future smell system checks food batches for spoilage. This is a hypothetical scenario, not an Osmo product test. A factory would care about missed spoiled batches, but it would also care about good batches stopped by false alarms. One attractive accuracy number could hide an expensive operating problem.
Let’s put numbers on that example. Assume 10,000 batches, of which one percent are spoiled. Our imaginary detector catches 90 percent of the spoiled batches and falsely flags five percent of the good ones. Those assumptions produce 90 correct alerts and 495 false alerts. Ten spoiled batches pass through undetected. Of the 585 total alerts, only about 15.4 percent indicate spoilage.
Nothing in that chart measures Osmo. It’s arithmetic showing why the frequency of a problem matters when judging a detector. Change the assumed spoilage rate or the false-alarm rate and the outcome changes. The right operating choice would also depend on what happens after an alert. A cheap second check is different from throwing away an entire batch.
This is the sort of test I’d look for before accepting a broad deployment claim. Let the proposed user define an acceptable miss rate and an acceptable false-alarm burden. Evaluate the complete system on samples collected where it will operate. Keep a separate test set. Report what happened when conditions changed, rather than presenting only the cleanest run.
A model score answers a model question. A buyer has to decide whether the whole process works well enough to use on Monday morning.

The demonstrations I’d want to see next
For odor prediction, I’d want a blinded evaluation on previously unseen chemicals, with the scoring method fixed before results are examined. The report should explain the human comparison and how much the evaluators agreed with each other. A model can look different depending on whether it is measured against one person or a group’s average.
For fragrance generation, I’d want repeated briefs from actual prospective customers. Include a difficult revision halfway through. Ask for a less sweet version, or remove an ingredient, then measure whether the next samples satisfy the change. Keep identities hidden during judging. Publish how many attempts were needed, because a successful final bottle can conceal a lot of unsuccessful intermediate ones.
For scent reproduction, I’d separate recognition from preference. “I like this” and “this matches the original” are different answers. Have evaluators compare originals and recreations without knowing which is which. Repeat the process with samples the team didn’t select for a demonstration. Document what the equipment couldn’t reproduce.
And for any proposed sensing application, I’d test the input hardware alongside the software. A deployment claim concerns the assembled system. Performance on stored chemical data doesn’t automatically establish performance when a device has to collect that data itself. The evaluation should make that handoff visible instead of burying it behind the word AI.
None of those requests requires dismissing the research. They’re ways to discover where it is useful. They also leave room for an outcome that technology coverage often handles badly: a system can be commercially valuable on a narrow task while its broadest description remains unproven.
A smell reading can become a record
There is another issue I’d raise before putting chemical sensors into everyday spaces: what gets stored? This is a question about possible deployments, not a claim that Osmo is collecting private household data. A tool used inside a fragrance lab has a different relationship with its surroundings from a sensor installed in a shared workplace.
For any proposed monitoring product, I’d ask whether it keeps raw measurements or only an alert. Who can access the records? Can the person being monitored understand what the system inferred? A measurement can be accurate while the explanation attached to it is wrong. An unfamiliar chemical signature might justify investigation without justifying a confident story about its source.
That distinction matters for product design. A dashboard that says “check this sample” gives an operator a different instruction from one that declares a cause. The second message needs evidence for the cause, beyond evidence that something differed from the reference. It’s a small wording decision with practical consequences.
These aren’t reasons to abandon sensing technology. They’re questions to answer while the proposed use is still specific enough to test. Start with the decision someone needs to make, then collect enough information to support that decision. Keeping everything because storage is available makes the questions harder. A useful instrument should come with a clear account of what its output permits someone to conclude, and where that permission ends.
I’d start with the bottle
Osmo gives us something worth following: published work on predicting odor descriptions, a company building physical scent systems, and a commercial process people can evaluate through actual samples. The newer claim of human-level olfactory intelligence deserves its own evidence. It shouldn’t borrow certainty from a narrower study conducted years earlier.
I haven’t tested Osmo’s equipment or smelled its generated samples. My interest is in the practical question its work makes possible: can software help people make better decisions about chemicals they would otherwise have to evaluate through a slower process? That question has a measurable answer, even if the answer differs across tasks.
Give a customer an unlabelled bottle, record what they think, and count the work it took to get there. That’s where I’d start judging an AI nose.