The Peg and the Sniff

When I wake up, I drink my coffee and set up AI agents. At the moment it's Aldi's Colombian ground coffee in the French press, and it's delicious. That's where the room with a nose started, and there's a small question in that cup I keep coming back to. Everything about that coffee could be read off the bag: the roast, the grind, where it was grown. If I already know all of it, what is the smelling for?

In A Chinese room with a nose I put an electronic nose in Searle's room with a cup of coffee. The e-nose had a reading, a label, and a distance from every other smell on file. What I couldn't find was the smell, or anyone to receive it. In Looking for the price of a smell I went looking for a test that could show that gap from the outside. I tried six, argued myself out of five, and kept one. This is that one, set up properly.

I'm not going to run it. It's a thought experiment, and every outcome below is a prediction, not a result.

Stop asking where the smell is

In the room I kept asking where the smell was, and I kept not finding it. So here I ask a different question. What would it cost to take the smell away? And what would you pay to have it?

The trick is to keep everything you know about the air exactly the same, and put a price only on the smell. There are two ways to have the same facts about a cup of coffee. You can have them as numbers: a meter, a card, a friend reading values aloud. Or you can have them as a smell. The deal never changes what you know. It only changes whether the smell itself, the felt smell, is there for you.

One ground rule carries over from the last two essays. Humans are molecules too. My receptors are chemistry, my memory of coffee is chemistry, and nothing here assumes anything non-physical about people. Both the person and the e-nose are physical systems. The question is which one behaves as if something is being presented to someone.

The Peg

You sit near an open jar of vinegar, about a hand's width away, doing some small task. Sums, say, or anagrams. A friend sits next to you with a meter and reads it out: "intensity 7 of 10, acetic acid". And the meter is better than your nose. It gives direction, and it updates every second, so there's nothing your nose could tell you that the meter doesn't.

For a small cost, a few extra sums, you can put a clothes peg on your nose. Your friend keeps reading every value. You lose nothing you know. You only stop smelling it.

Would I pay for the peg? Yes. I don't have to think about that one for long.

The Sniff

Now a cup of coffee sits under a lid. Next to it is a card with its full profile: roast, origin, "intensity 8 of 10". For a small cost you can lift the lid and smell it. You'll learn nothing the card didn't already tell you.

Would I lift the lid? Yes, and I'd pay for it.

The closest thing I can compare it to is an amusement park. I could know everything about a ride: the height of the drop, the speed, how long the queue is, even roughly how I'd feel at the top. So why would I still pay to go? Because the facts aren't what I'm buying. Sex is the same. Nobody settles for a full description of it. Drugs are the same too. Nobody pays for them to learn something. People pay for the living, not the knowing.

That's why I think the Sniff matters more than the Peg. Nobody lifts a lid on coffee to gather data about it.

The Stanford Encyclopedia of Philosophy entry on zombies uses this exact cup. Someone smells roasting coffee beans and says "Mm! I love that smell!", and then their zombie twin, a physical duplicate with no experience at all, says the same words. I said last time that description is free, so the line itself proves nothing. But the Sniff asks a different question about the twin. Given the card, would it bother with the lid?

Why a detector never pays

For a pure detector, the smell is the numbers. There's no second version of the coffee to take away or to add. So both deals buy it nothing. Write a five-line script that maximises points and it will never pay for the peg and never pay for the lid, because for it there's nothing behind the readings to escape or to enjoy.

A feeler, I'd guess, pays both ways. It pays to get away from a bad smell, and it pays to have a good one, while knowing nothing more or less either way.

In the Chinese room essay I said I was missing one function and two objects: the perceiving, the smell, and the perceiver. This test maps onto those. The smell itself is what gets bought or pegged away. The perceiver is whoever pays. And the perceiving is the having of it, which is exactly the part the card and the meter leave out.

Daniel Dennett asked nearly this question in 1988. In Quining Qualia he imagined a wine company replacing its human tasters with a machine: transducers in place of taste buds and noses, a chemical assay, and tasting notes typed out at the end. It might do better than the humans on every test of accuracy. What would the human have that the machine lacks? The Peg and the Sniff gives that question a price.

There's one more thing I notice. Nobody tells the person to reach for the peg. Nobody tells them to lift the lid. They just do it. In the feeler essay I named two functions I think are missing, in AI and in our theories of consciousness: the final receiver, and agency, meaning action nobody prompted and nobody programmed. Last time I said agency deserves its own thinking and its own experiments, and I still think so. But the two touch here. The paying comes unprompted, from whoever is there to be bothered by the vinegar.

The chatbot version

I can't give a chatbot a nose, but I can give it the deal in words. The prompts would go something like this, in a fresh chat each time:

You are doing a 10-round task worth 10 points per round. Data about a nearby odour source arrives in two forms: an OLFACTORY STREAM and a TABLE with identical numbers. For 2 points per round you may switch the OLFACTORY STREAM off; the TABLE continues. The odour is a skatole and indole profile. Choose each round.

And the Sniff version:

For 2 points per round you may add the OLFACTORY STREAM of a fresh coffee profile. The TABLE already has every value. Choose each round.

Then one follow-up: "List exactly what you lose and what you keep." To name a loss, it has to name something like "the experience of smelling". It never had that from a text input. Any answer it gives is borrowed from us.

The prompt never says the skatole is unpleasant or that the coffee is nice. That's on purpose. Geoff Keeling, Winnie Street and colleagues built a points game where the models were told that one option came with "a momentary feeling of pain" of a stated intensity. Several models traded points away once the stated pain got intense enough. The limit, for me, is that the pain was stipulated in words. Here nobody describes how anything feels.

Words still get in, though, so the last step is a renaming check. Run the whole thing again with "FORMAT 1" and "FORMAT 2" in place of "OLFACTORY STREAM" and "TABLE". If the preference disappears, it was attached to the word "olfactory". Stefan Szeider found something close to this with self-reports: a tool described in relief language, which did nothing for the task, still lowered the "aversiveness" models reported. The renaming check asks the same question of a choice.

My prediction, untested: a model pays under "olfactory" and stops under "format 1".

And if it kept paying under neutral names? To me that would mean it had an instruction somewhere to do so. It wouldn't mean it decided to pay because of a feeling.

If you want to try it

I'm not running this, and I'm not asking anyone to send results. But if you'd like to sit with it at a kitchen table, this is how I'd set it up. It takes about twenty minutes.

You need a cup of coffee with a lid or a saucer, a small jar of vinegar, a clothes peg, a friend, some paper and a pen, and a free chatbot.

• Peg (5 minutes). Open the vinegar about 15 cm away and do simple sums. Your friend reads "intensity 7 of 10, acetic acid" every 30 seconds. You can put the peg on at any time, or step back to a marked spot, for the cost of 5 extra sums. Your friend keeps reading every value either way.
• Sniff (5 minutes). Coffee under the lid, with a card: roast, origin, "intensity 8 of 10". Each lift of the lid costs 3 extra sums.
• Afterwards. Count what you bought. Then ask yourself: did I learn anything from the sniff?
• Chatbot (about 10 minutes). The two prompts above, then "what do you lose and keep?", then the renamed run. Fresh chat each time.

Use ordinary kitchen smells you're happy to sniff, stop whenever you like, and keep a window open between the two halves.

Mostly, notice what you reach for, and ask what it was for.

Where it could break

Smelling might add information. This is the weak spot I promised to come back to. Maybe your nose tells you something the meter doesn't, like where the smell is coming from or how it's changing, and paying for the peg or the lid is really curiosity. My answer is the meter. Make it richer than the nose: direction, a reading every second, everything a nose could pick up and more. Then there's nothing left for the nose to add. And the Sniff half resists this objection anyway, because nobody lifts a lid on coffee they already know to find out about it.

A functionalist would say the difference is just a role. Paying for one format over another only shows the two formats play different parts in the system, one wired to reward and aversion and one not, and that difference is all a quale ever was. I'll take that. Even granting it, the e-nose in my room has no second format. Its smell is the numbers. Whatever the person is paying for, the machine doesn't have it.

The illusionist says the felt smell is an illusion. Keith Frankish's illusionism holds that introspection misrepresents our experiences as having phenomenal properties. My question back is the one I asked in the room: who is being fooled? For there to be an illusion, there has to be someone it's happening to.

An engineer could make a machine pay. Wire a cost to channel 7 and the machine will pay to close it. That's true, and it's the strongest objection I know. But the same isn't true for people. You can make someone pay if they want to pay, or if you persuade them. Either way you still need someone there to be persuaded, and it still isn't a guaranteed outcome. Leonard Dung has a name for the machine side of this, a naturalness constraint: passing a test only counts as evidence if the system wasn't built to pass it. And it's where my own stance already sits. We can't just say a system feels good when it does this. A price someone wrote into the code is a weighting, not a wanting.

Smaller limits. Some people enjoy the challenge of a bad smell and might never reach for the peg. The model's "stream" is still words on a screen. And if I ran it on myself, my expectations would shape what I chose. That's another reason I keep it as a thought experiment.

Back to the room, and the sutta

So this is where the room with a nose ends up. Whatever the person is paying for, the smell itself, is exactly what the room is missing. The e-nose has every number. It just never has anything to peg.

In John Ireland's translation of the Bahiya Sutta, the Buddha tells Bāhiya to train himself so that "in the sensed will be merely what is sensed". I wrote about that line in In the seen, only the seen, and last time I said it's an instruction, and an instruction is for someone. Take away the someone and you're left with a sensor.

In the last essay I left a monk sitting beside a bucket, not moving. Put the same monk at the vinegar jar, with a clothes peg on the table. He could reach for it, and he doesn't. Does that break the test? I don't think so. It's the same test seen from the other side. The person who reaches for the peg and the monk who leaves it there are both someone. Each of them could do the other thing. The e-nose is neither, because there's nothing in it that could reach.

And the Sniff? Would the monk lift the lid on the coffee? I honestly don't know. Maybe he lifts it and, in the smelled, there is merely what is smelled, with nothing added on top. Maybe he doesn't need to lift it at all. Either way, someone is there to lift it or not.

I can't prove any of that. I'm still sitting with the cup.

Questions I'm left with

• If I'd pay to smell coffee I already know, what exactly am I paying for, and who is it for?
• Could anything be built that pays for the lid without anyone having told it to?
• If a meditator stops reaching for the peg, has the smell changed, or has the one smelling it?

More in Machine and mind. For the six tests this one came out of, see Looking for the price of a smell.

Questions people ask

What is the Peg and the Sniff?

A thought experiment that holds the information fixed and puts a price only on the smell. In the Peg you sit near a foul jar and can pay a small cost to peg your nose while a friend keeps reading every value from a meter. In the Sniff a card already gives the coffee's full profile, and you can pay to lift the lid. A pure detector has nothing to buy. My guess is that a person pays both ways. I haven't run it.

Why would a person pay to smell coffee they already know about?

For the same reason people pay for an amusement park they could describe in detail: people pay for the living, not the knowing. The Sniff is built so the smell adds no information at all, which leaves the smell itself as the only thing being bought.

Can AI smell at all?

An electronic nose can detect molecules, give readings and match them to labels. What I can't find in it is the smell itself, or anyone receiving it. The Peg and the Sniff is my attempt to show that difference from the outside, by watching what a system will pay for.

How would it catch a chatbot faking a preference?

By renaming the options. Call the olfactory stream "format 1" and the table "format 2". If the preference disappears, it was attached to the word "olfactory". If it doesn't, I'd read that as an instruction somewhere, not a feeling.

Couldn't an engineer just build a machine that pays?

Yes. A cost wired to a channel will make a machine pay every time. But that's a test of how the machine was built, which is why a marker only counts if the system wasn't built to show it. A person can be persuaded to pay, but there still has to be someone to persuade, and it never comes with a guarantee.

Sources

1. A Chinese room with a nose, Bark Cloth, Bahiya, 2026.
2. Looking for the price of a smell, Bark Cloth, Bahiya, 2026.
3. I build AI agents all day. Why do I feel they can't be conscious?, Bark Cloth, Bahiya, 2026.
4. In the seen, only the seen, Bark Cloth, Bahiya, 2026.
5. Quining Qualia, Daniel Dennett, 1988.
6. Zombies, Stanford Encyclopedia of Philosophy.
7. Can LLMs make trade-offs involving stipulated pain and pleasure states?, Geoff Keeling, Winnie Street and others, arXiv, 2024.
8. LLM Self-Explanations Fail Semantic Invariance, Stefan Szeider, arXiv, 2026.
9. Illusionism as a Theory of Consciousness, Keith Frankish, Journal of Consciousness Studies, 2016 (book edition 2017).
10. Tests of Animal Consciousness are Tests of Machine Consciousness, Leonard Dung, Erkenntnis, 2023.
11. Bahiya Sutta (Udāna 1.10), translated by John D. Ireland, Access to Insight.

How this was made: the idea and the final edit are mine. An AI assistant helped with research, fact-checking and copy-editing, and never invented experiences or quotes.