What Data Would AI Need to Tell Smells Apart?

Describing a Smell Versus Detecting One

AI can describe the scent of a flower or a cup of coffee in words. But that description alone can’t tell you what’s actually in the air of the room you’re sitting in right now. You need sensors that measure the actual chemical compounds present, plus technology to interpret those signals. A piece I recently read in Noema magazine got into exactly this gap.

So how far has research on machines detecting and interpreting smell actually come?

Machine olfaction¹ research already exists, but it’s tiny compared to work on language and vision. Kordel K. France and colleagues tallied papers on arXiv and Semantic Scholar from January 2015 through May 2025 and found that machine-olfaction papers came to less than 1% of the volume published in computer vision and natural language processing. The exact figure depends on search scope and classification method, but it points to a real gap in scale.

What caught my attention was why there’s so little research. Making sensors smaller and more accurate turns out not to be enough. You also have to figure out how to compare readings across different sensors, and agree on what names and standards to use when recording the same smell.

Distinguishing What “Can Do” Actually Means

The phrase “AI that can smell” can point to entirely different technologies. A model that predicts scent from chemical structure, a device that distinguishes specific gases from sensor signals, and a model that generates sentences about smell are all doing different things. The BBC Future taste-and-color-association example that Noema flagged reads, to me, as a caution against confusing a language model’s answers with actual sensory measurement.

This distinction matters even more in world model² research, where robots perceive their surroundings and plan actions. If a robot can detect leaks or decay that aren’t visible on screen through chemical signals, that’s more information for it to work with. Still, you can’t conclude that every AI needs a sense of smell. What sense is required depends entirely on the task at hand.

What Smell Actually Does for Us

The Noema article opens with a passage in which both Darwin and Kant ranked smell below the other senses.

But for humans, smell helps us judge the condition of food and read our surroundings, and it’s tied closely to memory and emotion.

Looking at that relationship also gives us a reason to care about machine olfaction research.

Whether findings from the human brain can simply be transplanted onto AI, though, is a separate question.

Smell is tightly linked to the brain regions involved in memory and emotion. The experience of a familiar food smell conjuring up a scene from the past is part of this same connection. Rather than saying the olfactory pathway “goes straight into the hippocampus,” it’s more accurate to say several regions work together.

There’s also a small study that looked at scent exposure alongside memory. In a 2023 paper, researchers at UC Irvine assigned 43 adults aged 60 to 85 to two groups. One group inhaled seven different scents, rotating one per day, for two hours every night over 6 months, while the control group was exposed to only trace amounts of scent. Because COVID-19 disrupted testing, only 23 participants ended up included in the cognitive-test analysis.

The 226% difference the paper reported is a comparison of before-and-after change on one specific item of a word-learning test between the two groups — it doesn’t mean overall memory improved by 226%. Several other cognitive tests showed no significant difference. MRI scans did show a change in diffusion metrics in the left uncinate fasciculus³, but this alone can’t confirm an overall improvement in brain function or a dementia-prevention effect.

Some research has also observed reduced sense of smell in conditions like Alzheimer’s and Parkinson’s disease. But a decline in smell can’t be treated as definitive proof that a specific disease is coming. Studying the link between olfactory change and disease is different from diagnosing based on that change alone.

These studies show that smell is an important subject for understanding human cognition and health. Whether chemical signaling is worth pursuing in AI too is something we could test by defining a concrete task and checking it out.

Looking at the fragment, it matches the Korean source well structurally and semantically. No Hangul, headings/lists/image match. Only minor polish needed.

Five Tasks the Researchers Proposed

In a position paper published in 2025, France and co-authors proposed expanding research into machine olfaction. According to Noema, the same paper was also presented at EurIPS that year. The five tasks outlined in the paper are as follows.

1. Understand smell better. There’s still much to explain about how a molecule’s properties, its concentration, and the interaction of multiple substances connect to odor perception. That said, this doesn’t mean sensor research or standardization has to wait until biology is fully understood.
2. Build a shared standard for measurements. Metal-oxide, electrochemical, and optical sensors differ in how they detect and what they output. Recording conditions like device type, concentration, temperature, humidity, and measurement time alongside the data makes it easier to reuse across different studies. The analogy to easily shareable image formats like JPG and PNG points to exactly this idea.

3. Decide how to attach ground truth. Expressions like “floral scent” can vary by individual and culture. There are also cases where sensors detect concentrations too low for humans to perceive. It needs to be made clear whether human evaluations or chemical analysis values serve as the ground truth.

4. Expand shareable datasets. Olfactory data already exists, but the researchers point out a shortage of large-scale resources that pair sensor signals with images, language, and other modalities in a way usable across multiple AI research efforts.

5. Establish common evaluation methods. Models need to be comparable using the same data and conditions. A result showing strong accuracy in one particular lab doesn’t tell you whether the same performance holds with different sensors or in real-world environments.

Having worked with data myself, the second item in particular resonated with me. Even when a lot of data accumulates, it’s hard to combine and use it if you don’t know what was measured under what conditions. Standards, data, and models can all improve together, but reconstructing measurement conditions after the fact can be extremely costly.

Sensing with sensors, suggesting scent with AI

Among the cases Noema introduced were attempts to measure environments with sensors and attempts to help create perfumes.

Sigma, which France covered, is a portable smell-measuring device that connects to a smartphone. Noema explains that data collected with this device is sent to a multimodal dataset called ScentNet. It’s an attempt to gather smell signals alongside other information like video or sound. The direction is to build a reference resource for researchers, much like ImageNet.

France explains that linking chemical signals with video could help determine what’s happening nearby—things like vehicles or fires. Still, detecting a few specific gases doesn’t let you immediately conclude there’s a fire. You need to look at concentration, location, and information from other sensors as well.

In the perfume space, Osmo, founded by Alex Wiltschko, unveiled Generation by Osmo in March 2025. It’s a service where customers describe the scent they want, and AI and a perfumer work together to create and refine the blend. The company calls this technology “olfactory intelligence”⁴. In 2024, it also unveiled an experiment analyzing and reproducing the scent of plums, among other things. They called it “scent teleportation,” but in practice it’s a process of measuring components, sending the data elsewhere, and blending the fragrance there.

Carto, from the Swiss fragrance company Givaudan, and Philyra, developed by IBM with Germany’s Symrise, are also tools that assist in perfumery.

Looking at these cases, I think we need to separate the goals involved. Proposing a fragrance blend and refining it with human evaluation is different from a robot interpreting real-time signals from the surrounding air. Progress on one doesn’t mean the problems of the other have been solved.

Problems That Remain in Real-World Environments

There are several kinds of equipment for measuring chemical substances.

Gas chromatography-mass spectrometry (GC-MS) is equipment that separates and analyzes the components of a mixture. Since there are both lab-scale and portable versions, you can’t reduce its size, price, and analysis time to a single figure. To use it in a robot, you have to weigh how fast and reliably it can detect a target substance against its weight, power draw, and maintenance costs. Even compact “electronic noses” still face the following problems:

• Sensor drift: Sensitivity can shift over time or with contamination, so the same smell can yield different readings — meaning calibration is required.
• Environmental variation: Temperature, humidity, and other gases can affect measurements. You need to verify that lab results hold up in the field.
• Interpreting mixtures: Detecting a specific substance is a different task from classifying a blend of many substances as “coffee.”

Smell varies not just by composition but by concentration and mixing ratio — some molecules amplify or suppress the effects of others. The sheer number of possible combinations doesn’t mean measurement itself is impossible; you simply have to decide what needs to be distinguished based on the purpose at hand.

Unlike color, which can be expressed with just three RGB values, no common way of representing the full range of smells has yet become well established. That said, research is already underway to learn the relationship between molecular structure and scent and express it in compact coordinates.

Oswarld’s Lens

Having worked in GTM strategy, I read this topic as an industry problem about what data gets collected, and who gets to reuse it.

The advance of computer vision came from GPUs, algorithms, and large-scale datasets and benchmarks working together — it’s hard to point to any single cause. I think olfactory AI needs the same combination: developing models, yes, but also recording measurement conditions and building the materials needed to compare performance.

If a Korean company were looking into this field, I’d start by examining how to collect and validate its own field data against a common standard. Efforts like the IEEE P2520 series, which have discussed device and performance-testing standards, could serve as a reference. Gas-leak detection and quality inspection are already established applications, and there’s room to make them more accurate across different environments. Uses like disease diagnosis would require separate clinical validation. Being an early participant in standardization doesn’t guarantee success, but it would help clarify exactly what problem the data is meant to solve.

Closing

Machine olfaction research isn’t starting from zero. Sensors, datasets, and industrial products already exist. What’s still missing is the work of connecting results obtained under different conditions and turning them into AI that holds up in new environments.

When I see a presentation claiming “AI can smell,” I first check what exactly it detected. Which substances, at what concentration, and across what range of environments it can distinguish. Only with that explanation can you judge where it’s actually usable.

For readers who want to go deeper, I’d recommend the sections of the France team’s paper covering data representation, datasets, and evaluation methods. It also includes a proposal to apply neuromorphic computing⁵ to olfactory signal processing. Once I finish the study I’m currently working through, I’ll follow up with a piece connecting Yann LeCun’s emphasis on world models to sensory learning.

References & Further Reading

Primary sources

• Maughan, Philip. “Why AI Needs A Sense Of Smell.” Noema Magazine, April 16, 2026. : This is the essay that sparked today’s newsletter. It covers a wide range of ground — from why AI olfaction research has stalled, to field interviews with companies like Osmo and Sigma, to smell-based dating experiments. It also lays out the author’s own view on why machine olfaction research matters.
• France, Kordel K., et al. “Position: Olfaction Standardization is Essential for the Advancement of Embodied Artificial Intelligence.” arXiv, 2025. : A paper proposing the key challenges in machine olfaction research. Section 2.2 covers data standards, Section 2.4 covers datasets, and Section 2.5 covers evaluation criteria. The discussion of neuromorphic computing is also in Section 2.2.
• Woo, Cynthia C., et al. “Overnight olfactory enrichment using an odorant diffuser improves memory and modifies the uncinate fasciculus in older adults.” Frontiers in Neuroscience, 17: 1200448, 2023. : A small randomized trial comparing scent exposure with memory test outcomes. 43 participants were enrolled, and 23 were included in the cognitive test analysis. This lets you check exactly which test item and which measure of change the 226% figure cited in the body refers to.

Background

• Osmo, How we invented scent teleportation, October 29, 2024. : The company’s own account of the procedure and limitations of experiments that measure, recombine, and reproduce scents.

• IEEE, Introduction to the IEEE P2520 series of standards, 2021. : Introduces standardization efforts for machine olfaction devices and performance evaluation.

• Teledyne FLIR, Portable GC-MS product announcement, October 3, 2023. : A manufacturer’s release confirming that field-deployable analysis equipment already exists.

• Osmo official announcement. “Osmo Launches Generation, World’s First AI-Powered Fragrance House.” BusinessWire, March 5, 2025. : The official launch press release for Generation by Osmo. Useful for understanding how Olfactory Intelligence (OI) technology is being commercialized as a business model — a fragrance design house combined with B2B licensing.

• France, Kordel K. “Machine Olfaction in Artificial Intelligence and Robotics.” 2025. : A survey paper offering a full overview of the machine olfaction field. Recommended for anyone approaching this area for the first time. It also touches on ML techniques that pair well with olfaction, such as active learning and multi-agent reinforcement learning.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. Machine olfaction is a research field that uses chemical signals or molecular data gathered by sensors to detect, classify, and predict smells. An electronic nose (e-nose) refers to a device that combines sensing with analytical functions. ↩

2. A world model is a model trained on how an environment changes over time, used to predict what happens next or to plan actions. The data and implementation vary from study to study. ↩

3. The uncinate fasciculus is a bundle of nerve fibers connecting the anterior temporal lobe with parts of the frontal lobe. It is one of the pathways associated with learning, memory, and emotion. ↩

4. Olfactory Intelligence (OI) is the name Osmo uses to describe its proprietary technology for handling the molecular information of smells and human perception of them. ↩

5. Neuromorphic computing is a field of hardware and algorithm research inspired by how the nervous system processes information. It often relies on event-driven computation, where processing occurs only when a signal fires; its power efficiency depends on the specific implementation and task. ↩