A Chinese room with a nose: could AI ever smell coffee?

When I wake up, I drink my coffee and set up AI agents. Somewhere in that routine the same thought keeps coming back, and it always takes me to the Chinese room thought experiment. Could one of these agents smell this coffee?

So I want to run my own version of it here. My answer is no. That's my own view, and I'll say so up front, as I do across Machine and mind. What others have argued about machines is in Part 3 of the research series. This page is me working out why it feels that way, and where the experiment leaves me with more questions than I started with.

Could an AI agent smell the coffee?

Let's set up the room. There's an agent with an electronic nose wired into it, a cup of coffee on the desk, and one job: smell the coffee.

I'd build it the way I build anything else. If I wanted an agent or a piece of software that smells the coffee, the smelling part is doable. Capture the molecular idiosyncrasies of the coffee aroma and send them to a classifier. The classifier says "coffee".

Run it with me for a minute. Picture the agent getting it right every single time. Now look around the room. What's in there?

What can an electronic nose already do?

This part already exists. A 2024 review traces electronic noses back to Persaud and Dodd (1982): an array of gas sensors plus pattern recognition.

Coffee is a popular test. Mutz and colleagues built an e-nose from 12 metal oxide sensors and used it to classify coffee by species, by roast level and by expiry, with accuracy above 90% on every task. Lee and colleagues trained a graph neural network on molecular structure to build a principal odour map which, in their words, "paves the way toward digitizing odors".

So imagine my agent can, in principle, detect any smell. It can tell coffee from tea and fresh from stale. It can compare each reading with every other smell it has on file.

Here's where I get stuck. Telling smells apart against a molecular reference doesn't mean the e-nose enjoys any of them, or that it has a preference. It sorts the stale coffee from the fresh. Does it mind which one it gets?

Where is the smell of coffee in the software?

When I look at what the agent actually has, I can't find the smell of coffee. There's a reading, a label, and a distance from other readings. In the software the smell is only real in relation to the molecular detection, or by comparison with other smells. It's self-referential. In and of itself, it doesn't exist. In me, it does.

There's a fair objection here, and I agree with it. "It's just molecules under the hood" is true of me as well. My receptors are molecules. My memory of coffee is chemistry. So whatever separates us, it isn't the mechanism. Only one of us feels it.

My smell of coffee is partly relational too. It's coffee partly because it isn't tea, and because of all the mornings behind it. So I can't claim the machine's smell is relational and mine is real. Both are relational, and only one is felt.

Say I bolt a perfect language model onto the perfect e-nose, and it says "this smells like my grandmother's kitchen". Does that change anything for me? No. I know the only reference that model would have under the hood is the molecular composition, or a reference to another molecular composition. Not the experience. My own reference is chemistry too, as far as anyone can measure. But mine is felt, and I can't find anything in the model that feels it.

Philosophers got here long before me, and one of them used a smell. The Stanford Encyclopedia of Philosophy cites C. D. Broad, writing in 1925, on an archangel who knew the full molecular structure of ammonia. He still couldn't predict how it smells "unless someone told him so or he had smelled it for himself" (SEP, Qualia: The Knowledge Argument). Frank Jackson's Mary, from 1982, knows every physical fact about colour vision from inside a black and white room, yet seems to learn something when she first sees red (SEP, Qualia).

My agent is Broad's archangel, with one difference. It never gets to smell the ammonia for itself. What would that even mean for a classifier?

The missing subject: who receives it?

My first note on this idea said I'd be missing one function and two objects. I can name them now. The function is the perceiving. The object is the smell of coffee. And the subject (another object, in software terms) is the perceiver.

The perceiver is the end object that would receive the smell, and it's missing too. It's the same gap I keep coming back to in I build AI agents all day. Why do I feel they can't be conscious?: the final receiver, the feeler. Even if the sensing and the sorting both work perfectly, who receives the smell?

Thomas Nagel put the question in 1974. An organism has conscious states, he wrote, "if and only if there is something that it is like to be that organism" (Nagel, What Is It Like to Be a Bat?). Is there something it is like to be my agent, smelling coffee? I can't find anyone for it to be like.

This is also where the Bahiya Sutta comes back to me. In John Ireland's translation the Buddha tells Bāhiya: "In the seen will be merely what is seen; in the heard will be merely what is heard; in the sensed will be merely what is sensed..." (Ud 1.10, Access to Insight; see the translation note in Sources). I wrote about that line in In the seen, only the seen.

At first glance the e-nose looks like it's already there. Just the reading, with no story about anyone's kitchen added on top. So would it be the perfect meditator? It would, if there were a subject perceiving the sensed as "just the sensed". That's the catch. The instruction is for someone. Take away the someone and there's no practice left, only a sensor.

How is this different from Searle's Chinese Room?

John Searle's Chinese Room, from 1980, goes like this. Someone who speaks no Chinese sits in a room with a rulebook. Chinese questions come in, the person follows the rules, and fluent Chinese answers go out. From outside, the room understands Chinese. Inside, nobody understands a word (SEP, The Chinese Room Argument).

My version swaps the person for software and the Chinese symbols for smells. So we're trying to build a piece of software or an AI agent that does the understanding. It can detect any smell, but purely by digital comparison, not by experience.

I should admit that it's a hybrid. Searle's room is about meaning: does the system understand? Mine is about feeling: is there a smell, and is there a smeller? That puts it closer to Broad's archangel and Jackson's Mary than to Searle. I've borrowed his room, put a nose in it, and asked a different question once inside.

The replies I keep hearing

The Chinese Room has collected replies for over forty years, and most of them could be pointed at my coffee room too. Here's where I land on four of them.

The systems reply. The person in the room doesn't understand Chinese, but the whole room does. So maybe the whole agent smells the coffee, even if no single part does. My answer: still nobody home.

The robot reply. Give the system a body and senses, so its symbols connect to the world. But my agent already has a sense. The e-nose is the whole point. The SEP notes Searle's own answer, that sensors only add more syntactic input. I'm not bothered by this reply. It still doesn't explain the experience, and the experience doesn't happen in the nose.

Functionalism. A mental state is what it does. If the agent does everything that smelling does, it smells. But by the same logic, a machine that is made to smell faeces all day and night would also be rebelling against its developer. I've said before that without a feeler there's no compulsion. Here the compulsion would be to get away from the smell. Would the machine feel it?

Illusionism. The felt smell of coffee is an introspective illusion, so there's nothing extra to explain. My question back: who is being fooled? We're back at square one. Again, it's a self-referential statement. For it to be an illusion, there must be a subject that is the victim of the illusion in the first place.

Questions I can't answer yet

If my own smell of coffee is molecules and relations too, what is the feeling adding? And why is it there at all?
Would I know if an e-nose ever started to prefer one coffee? Could I tell a preference apart from a weighting someone set?
Broad's archangel could learn the smell by smelling it. Is there any version of "smelling it for itself" open to software, or is that just asking for a subject under another name?
The sutta says there is no you in the seeing. I keep saying the e-nose needs a subject to be a meditator at all. Are those the same subject, or am I asking for two different things?
Is it fair to borrow Searle's room for a question about feeling, or does the hybrid blur both?

FAQ

Can AI smell?

Electronic noses detect and classify odours using gas sensor arrays and pattern recognition. One coffee e-nose reached over 90% accuracy on species, roast and expiry. Whether anything is actually smelled, rather than measured, is the open question I'm asking here.

What is the Chinese Room argument?

John Searle's 1980 thought experiment. A person who speaks no Chinese follows a rulebook and produces fluent Chinese answers without understanding a word. Searle argues that following rules for symbols is not enough for meaning.

What is Broad's mathematical archangel?

In 1925 C. D. Broad imagined a "mathematical archangel" with unlimited mathematical skill who could perceive the structure of atoms. It would know ammonia's structure but could not predict its smell "unless someone told him so or he had smelled it for himself".

What is Mary's room?

Frank Jackson's 1982 thought experiment. Mary knows every physical fact about colour vision but has lived in a black and white room. When she first sees red, does she learn something new?

How is a Chinese room with a nose different from Searle's?

Searle's room asks whether a system understands. Mine asks whether there is a smell and a smeller. That puts it closer to Broad's archangel and Jackson's Mary.

Sources

The Chinese Room Argument, Stanford Encyclopedia of Philosophy.
Minds, brains, and programs, John R. Searle, Behavioral and Brain Sciences, 1980.
What has already been said about consciousness and AI? Part 3: The machines, Bark Cloth, Bahiya.
Qualia, Stanford Encyclopedia of Philosophy.
Qualia: The Knowledge Argument, Stanford Encyclopedia of Philosophy.
The Mind and Its Place in Nature, C. D. Broad, 1925 (Internet Archive scan).
Functionalism, Stanford Encyclopedia of Philosophy.
Illusionism as a Theory of Consciousness, Keith Frankish, Journal of Consciousness Studies, 2016 (book edition 2017).
What Is It Like to Be a Bat?, Thomas Nagel, 1974 (read via philosophy.uconn.edu).
Effectiveness of an E-Nose Based on Metal Oxide Semiconductor Sensors for Coffee Quality Assessment, Yhan S. Mutz and others, Chemosensors, 2025.
Electronic Noses: From Gas-Sensitive Components and Practical Applications to Data Processing, Zhenyu Zhai and others, Sensors, 2024.
Analysis of discrimination mechanisms in the mammalian olfactory system using a model nose, Krishna Persaud and George Dodd, Nature, 1982.
A principal odor map unifies diverse tasks in olfactory perception, Brian K. Lee and others, Science, 2023 (read via pmc.ncbi.nlm.nih.gov).
Bahiya Sutta (Udāna 1.10), translated by John D. Ireland, Access to Insight.
In the seen, only the seen: the Bahiya Sutta for AI builders, Bark Cloth, Bahiya.
I build AI agents all day. Why do I feel they can't be conscious?, Bark Cloth, Bahiya.

Translation note: this essay quotes Ireland's translation because it keeps the word "sensed". My other Bahiya Sutta essay, In the seen, only the seen, uses Bhikkhu Sujato's translation, With Bāhiya (Udāna 1.10), SuttaCentral.