Smell, the Last Sense AI Still Lacks

Opening

Hello, dear subscriber, this is Oz’s Knowledge Talking. Do you happen to like perfume?

Let me start with something a bit unusual today. We’re living in an era where AI passes medical and bar exams, AI-designed cancer drugs enter clinical trials, and AI agents sort our email for us. But an article I recently read in Noema magazine raised an interesting question.

“Why is research on giving AI a sense of smell moving so slowly?”

Let me give you one number first. Between 2015 and 2025 — a full 10 years — the number of papers on machine olfaction¹ came to less than 1% of the papers on computer vision and natural language processing combined. Major AI conferences like NeurIPS, ICLR, and ICML barely touch the topic. Even most of the humanoid robot startups flooding the market today are designing their machines without any olfactory sensors at all.

This might look like a simple matter of priorities — “vision and language first, smell later.” But the reason I’m bringing this up today is that I don’t think this is a question of sequence at all. I think it’s a question of structure. Let’s unpack what that structure actually is.

Overhyped Headlines and the Real Gap

The phrase “AI that can smell” isn’t new. Headlines like “Computers Are Learning to Smell” (The Atlantic) and “AI is digitizing our sense of smell” (World Economic Forum) have been recycled for years. But the Noema article points out that most of these headlines are overstated. Take a close look at the BBC Future piece “An AI started ‘tasting’ colours and shapes”: what actually happened is that the LLM simply repeated the human associations baked into its training data — “sweet things are pink and round.” The AI didn’t smell anything; it copy-pasted text that humans had already written about smell. In other words, it didn’t “smell” anything — it merely described smell.

So why does this matter now? As discussions of “LLMs hitting a ceiling” have gained traction, AI researchers have started looking to world models² — systems that internally reconstruct the structure of the world using multiple sensory channels like vision, hearing, and touch — as the next breakthrough. And by volume of daily sensory input, smell ranks third for humans, right after sight and hearing. Yet most world-model research leaves this third channel out entirely.

🧠 Why Smell Is Part of Intelligence

“The sense of smell is of extremely slight service.”

That’s Charles Darwin, writing in 1874. The philosopher Kant went even further in 1798:

“Smell is the most dispensable sense, and the least worthy of cultivation.”

For most of Western intellectual history, smell was treated as an inferior sense. But neuroscience over the past twenty years has been overturning that verdict. It’s hard for us to imagine now, but let me lay out a few reasons why people back then thought this way.

First, smell connects directly to the brain’s oldest circuitry. Every other sense passes through the thalamus before reaching the cerebral cortex, but smell alone bypasses the thalamus entirely, feeding straight into the hippocampus and amygdala — the regions responsible for memory and emotion. That’s why the smell of food can suddenly trigger a childhood memory.

Second, olfactory training genuinely boosts cognitive ability. A 2023 study from a UC Irvine neuroscience team is especially striking. Researchers split 43 adults aged 60 to 85 into two groups. One group was exposed to alternating scents (7 kinds, including rose, orange, and eucalyptus) for two hours every night while sleeping; the other group received only a trace level of scent. After 6 months, the scent-exposed group scored 226% higher than the control group on verbal memory tests. MRI scans even showed structural improvement in the left uncinate fasciculus³ — the pathway that integrates memory and emotion.

Third, loss of smell is an early warning sign for roughly 70 neurological and psychiatric conditions — including Alzheimer’s, Parkinson’s, schizophrenia, and alcoholism. Research is also piling up on the cognitive decline seen in people who lost their sense of smell after COVID-19.

In other words, smell isn’t a “nice-to-have” feature — it’s closer to the foundational infrastructure of intelligence, memory, and emotion. And the AI we’re building is skipping this infrastructure entirely.

📚 Why Did AI Put Smell Last? 5 Structural Gaps

The most systematic answer to this question came in a position paper presented at EurIPS 2025. The author is Kordel K. France, a robotics researcher and engineer at Toyota North America, along with his co-authors. Here are the five “structural gaps” they identify.

• Uncertainty in olfactory theory itself. We still don’t fully understand how smell receptors work. The “shape theory” — that odor molecules fit into receptors like keys into locks — is competing with “vibration theory” — that the vibrational frequency of molecules is what gets detected. Vision developed on top of an agreed-upon theory, like the RGB 3-color model. Smell is still stuck at the stage of basic scientific consensus.
• The absence of a data standard. This is the real crux of it. Images have JPG and PNG. Sound has WAV and MP3. Those formats made it possible to build massive datasets like ImageNet, and deep learning exploded on top of them. Smell still doesn’t have its own JPG. Detection methods vary wildly — metal-oxide sensors, electrochemical sensors, optical sensors, sensors that transplant insects’ biological receptors — and there’s no unified spec for representing the resulting data.

3. The limits of subjective labeling. Labels like “lemon scent” or “floral scent” vary enormously across cultures and individuals. The same molecule might be called “refreshing” by one person and “sour” by another. A 2008 experiment even showed that the same smell is perceived differently depending on what it’s called — label something “rose” and people perceive it as sweeter.

4. The absence of benchmarks. Language models have MMLU. Vision has CIFAR and ImageNet, with live leaderboards on HuggingFace. Smell has no such competitive yardstick yet. Without a shared metric for comparison, it’s hard to say objectively who’s doing better and who isn’t.

5. The absence of a community. This is the result of the previous four gaps compounding on each other. With no standards, no data, and no benchmarks, there’s little incentive for researchers to jump in. Papers are hard to publish, reproducibility is hard to verify, and industrial applications remain fragmented.

Reading this paper from a data background, the second point is the one that made me nod hardest. Regardless of sensory modality, a standard format has to exist before datasets can grow, and datasets have to grow before models can learn from them. The order is fixed.

🔬 Those Trying to Break the Deadlock — the Sensor Front and the Perfume Front

Right now, two major fronts are trying to fill this gap.

Kordel France, mentioned earlier, is the leading figure in this camp. His recent project, Sigma, is a portable smell-recording device that connects to a smartphone — designed to “record” smell the way a recorder or camera records sound or image. The data it collects feeds into an open multimodal dataset called ScentNet. You can probably guess the reference: it’s a deliberate echo of ImageNet, which Stanford’s Fei-Fei Li built in 2009. It’s close to a declaration of intent to build the ImageNet of the olfactory era.

France also works with Toyota, where he argues that even today’s basic olfactory sensors, paired with cameras on a robot, enable inferences like “ethanol, methane, and heptane detected → the vehicle engine is running, or there’s a fire.” Chemical signals fill in context that vision alone can’t provide.

Meanwhile, a $10 billion market is already running on the other front. The most closely watched player is Osmo, founded in 2022 by Alex Wiltschko. He’s the researcher who built Alphabet’s olfactory research group at Google Brain, and in March 2025 he officially launched Generation by Osmo, an AI-driven fragrance design house. Osmo runs a proprietary AI model called Olfactory Intelligence (OI)⁴, which takes text or image prompts and generates scent molecule combinations. In 2024, the company announced it had successfully digitized the scent of a “freshly cut summer plum” at the molecular level and “teleported” — recreated — it elsewhere.

Givaudan’s Carto and Philyra, which IBM built for Symrise, occupy similar territory — meaning Switzerland’s major fragrance houses are already using AI-driven scent-composition tools.

But there’s a structural distinction here that I think matters. In the perfume world, the final judge is still a human perfumer’s subjective evaluation. The AI proposes molecules, and the perfumer smells them and decides. Strictly speaking, this isn’t “a machine smelling” — it’s “a machine proposing a fragrance formula.” A robot that actually detects smell in its environment in real time and acts on it — true embedded olfaction — is still a ways off.

⚙️ Why Better Sensors Alone Won’t Solve This

The natural next question is whether better sensor technology alone would fix this. The reality is trickier than that.

Gas chromatography-mass spectrometry (GC-MS), the most precise molecular detection equipment available, is still refrigerator-sized, costs about $500,000, and takes 6 hours to analyze a single sample. You can’t strap that onto a robot. The alternative — compact “electronic noses” (e-noses) — comes with three limitations of its own:

• Drift: sensitivity degrades over time. Mammalian olfactory neurons keep regenerating; artificial sensors don’t.
• Limited detection range: they work well in lab conditions but degrade sharply in the real world.
• The gap between detection and recognition: catching individual molecules is one thing, but identifying a complex blend as “coffee beans” is a completely different problem.

And there’s one more fundamental question left: is smell even measurable at all? The estimated number of possible odor molecule combinations is on the order of 10^60 — 1,000 times the number of atoms in our solar system. And most smells are “chords.” Even strawberry scent is a combination of hundreds of molecules, each reinforcing or suppressing the others. When we say “the smell of the sea,” we’re essentially naming a ghost.

This is decisively different from vision. Vision can approximate nearly every color using just three RGB channels. Smell has no such low-dimensional axis.

Oz’s Lens

Layering in my own experience working in GTM strategy, I read this story not as a philosophical question about the senses, but as an industrial map problem about standards and data infrastructure.

Looking back, computer vision didn’t explode because GPUs got better. It exploded because standard formats — JPG, PNG — already existed, a shared dataset called ImageNet emerged, and competitive arenas like Kaggle and CIFAR were built on top of it. The bottleneck was never the algorithm — it was always standards, data, and benchmarks. Natural language followed the exact same path.

Smell is still a step before that starting line. If a Korean company is eyeing this market, I’d argue there’s far more leverage in getting in early on standard-setting and dataset-building than in racing to build models. The machine olfaction standard IEEE P2520 is already under discussion, and open datasets like ScentNet have only just begun. Applications will clearly open up — early disease diagnosis in healthcare, quality inspection in smart factories, gas leak detection in semiconductor and display manufacturing. Whether that happens within 3 years or 10 depends entirely on how quickly these five gaps get filled. In this game, positioning near the standard beats racing to be first.

Closing

• AI research today is concentrated on language and vision, but a growing body of argument holds that truly embodied intelligence can’t do without low-bandwidth, high-context senses like smell.
• Olfactory AI is slow not for lack of technology, but because of gaps in data standards, benchmarks, and community. It hasn’t yet walked the path computer vision already walked.
• The fragrance industry has already started moving, but embedded olfaction remains stuck at the standard-setting stage. Whoever gets involved at this stage will shape the next 10 years.

If you want to go deeper, I recommend skimming France’s position paper. The section on “why smell is well-suited to neuromorphic computing⁵” is worth reading even just as a story about next-generation semiconductor architecture. Once I wrap up this thread of study, I’ll put together a piece on the opposing camp — Yann LeCun’s “abstract reasoning first” argument.

References & Further Reading

Primary sources

• Maughan, Philip. “Why AI Needs A Sense Of Smell.” Noema Magazine, April 16, 2026. — The essay that started today’s newsletter. It ranges from the causes of stagnation in AI olfaction research to field interviews with Osmo and Sigma, and even scent-based dating experiments. It’s long, but it’s the deepest explanation I’ve found of why smell is part of intelligence.
• France, Kordel K., et al. “Position: Olfaction Standardization is Essential for the Advancement of Embodied Artificial Intelligence.” arXiv, 2025. — The position paper that systematized the “five structural gaps” discussed above. If you’re short on time, Section 3 (data standards) and Section 5 (affinity with neuromorphic computing) are enough.
• 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. — The original study behind the 226% memory improvement figure. I’d recommend checking the sample size (N=43) and methodology (2 hours nightly, 7 scents, 6 months) yourself. Keep in mind this is a small-scale RCT.

Background

• 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. Good for understanding how OI technology is being commercialized — as a fragrance design house plus B2B licensing.
• France, Kordel K. “Machine Olfaction in Artificial Intelligence and Robotics.” 2025. — A survey paper covering the entire field of machine olfaction. Recommended if you’re new to this area. It also covers ML techniques that pair well with olfaction, like active learning and multi-agent reinforcement learning.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Machine Olfaction: A field of technology in which machines or robots use chemical sensors to detect and classify smells. Also called the “electronic nose” (e-nose). Think of it as an attempt to mimic smell with sensors, the way a camera mimics sight. ↩

2. World Model: A method by which AI internally simulates the structure of the world it’s situated in, so it can predict what happens next and plan its actions accordingly. Think of a child predicting, before throwing a ball, “if I throw with this much force, it’ll land over there.” This is a different approach from LLMs trained only on text. ↩

3. Uncinate Fasciculus: A bundle of nerve fibers connecting the temporal lobe (responsible for memory) and the frontal lobe (responsible for judgment and decision-making). It weakens with age and is one of the first regions damaged in early Alzheimer’s. Think of it as a “highway connecting memory and judgment.” ↩

4. Olfactory Intelligence (OI): A term coined by Osmo, referring to an AI model trained on the relationship between the chemical structure of scent molecules and the scent labels humans perceive. Think of it as the olfactory counterpart to NLP for text and computer vision for images. ↩

5. Neuromorphic Computing: An architecture that, unlike conventional CPUs and GPUs that operate on a fixed clock cycle, activates only when events occur — the way the brain does. Its advantages are low power consumption and strong parallel processing. It pairs well with data that arrives “sparsely, in bursts” — much like smell. ↩