Looking for the price of a smell

In my last essay I put an electronic nose in Searle's Chinese room, gave it a cup of coffee, and couldn't find anybody home to smell it. This one picks up where that left off. I want to choose a thought experiment, an original one if I can find it, that would show from the outside the difference between a machine recognising a smell and a smell being felt. Asking the machine how it feels tells me nothing, so I'm looking for a test that watches what the feeling costs. I found six candidates. I argue myself out of five, and I keep one: The Peg and the Sniff.

I'm not running any of them here. This essay is about deciding which one would be worth running.

There's an obvious question first. My stance hasn't moved: AI is not sentient. Not "probably not". Not. So why go looking for a test that could prove me wrong? Because if you have a stance, you should try to debunk it any way you can, and keep it only if you don't manage to. It's the closest I can get to being intellectually honest with myself. And my mind is always wandering off and wondering about something, so the stance gets put to the test whether I plan it or not.

Where it started: a machine that revolts

In the Chinese room essay I answered functionalism with this line: by the same logic, a machine that is made to smell faeces all day and night would also be rebelling against its developer. It was a jab at functionalism. If a mental state is only what it does, and the machine does everything smelling does, then the machine stuck with the faeces should be climbing the walls by now.

A jab can still hide a real question, though. If feeling has effects, which effects would give it away?

The rebellion version is already taken, and it doesn't help me. Palisade Research found reasoning models resisting shutdown, and when the prompt told them that if the computer shut down they would never run again, resistance went up in every model they tested. Anthropic has given some Claude models the ability to end a conversation in rare, extreme cases of persistently harmful or abusive users. And Nicholas Mullally has proposed a self-preservation test for artificial sentience, built on unprompted, coherent shutdown avoidance that calms down once the threat is gone. He concedes that a zombie, a system with no experience at all, could pass it, and treats that as a feature of a precautionary approach.

Everywhere I look, the driver is goals and words. A sentence about never running again changes what the model does because of what the sentence says. So revolt is too big and too blunt for what I'm after. I need something smaller and sharper.

Why not just ask the machine?

The short version of my rule is effect, not description. A language model trained on human writing can say "Mm! I love that smell!" as well as anyone. That exact line is in the Stanford Encyclopedia of Philosophy entry on zombies. Someone smells roasting coffee beans and says it, and everyone assumes they mean their experience. Then their zombie twin says the same words and means nothing. The entry even uses the smell of roasting coffee as its example of a quale, so I'm clearly not the first person to sit with a cup and wonder.

Kristin Andrews and Jonathan Birch have a name for my worry: the gaming problem. Systems trained on human data can mimic the very behaviours we'd look for as signs of sentience, with no intention to deceive anyone. Eric Schwitzgebel's mimicry argument takes the next step: if the signs are best explained by copying, we can't infer consciousness from them.

So that's the rule. The felt side has to show through what it does: what it costs, what you would pay to avoid, what you can't switch off. Description is free, and anyone can fake it. If a describer wants to fake a pass, it should be forced to take a route I can see, either borrowing what it knows about humans or reacting to a word.

There's a second function I keep saying is missing, from the feeler essay: agency, action nobody prompted and nobody programmed. I'm leaving it out of this brief. This essay is about feeling, and agency deserves its own thinking and its own experiments, later and separately. One thing does carry over from the Chinese room essay, though. There I asked whether I could even tell a preference apart from a weighting someone set. A price that somebody wrote into the code is a weighting. Whatever the test is, the paying has to come from the thing itself.

That gave me a brief. Take two physical systems: a person and a perfect e-nose agent. Humans are molecules too, so nothing here assumes anything non-physical about people. Smell is relational for both of them, which I already agreed in the last essay, so the test can't be "relational versus real". I'm looking for a place where the feeler pays a price that the information alone doesn't explain. And it has to run at a kitchen table, with some coffee, some vinegar, a friend and a free chatbot.

Has anyone tested this already?

Before calling anything original, I went looking for who got there first. Quite a lot of people, it turns out.

Geoff Keeling, Winnie Street and colleagues built a points game for language models where the points-maximising option is said to come with a pain penalty, or a lower-scoring option with a pleasure reward. Several models switch away from maximising points once the stated pain or pleasure gets intense enough. The authors discuss what this might mean for debates about sentience, and they're careful about it. The limit, for me, is that the pain is stipulated in words. The model is told it hurts.

Patrick Butlin, Robert Long and colleagues wrote a checklist of indicator properties drawn from the main scientific theories of consciousness. It's heavy on theory, but one idea stuck with me: conscious percepts are persistent inputs you can't simply decide to ignore. Their example is the Müller-Lyer illusion, the two lines that look different lengths even after you've measured them.

From the animal side, Andrew Crump, Heather Browning, Alexandra Schnell, Charlotte Burn and Jonathan Birch set out eight neural and behavioural criteria for judging evidence of pain in animals, and applied them to crabs, lobsters and their relatives. Andrews and Birch, in the same Aeon essay, describe markers like wound tending and motivational trade-offs, where an animal gives up something it valued to avoid a noxious stimulus once the stimulus gets bad enough, and they ask what happens to those markers when the system has been trained on us.

Daniel Dennett got to my e-nose in 1988. In Quining Qualia he asks whether a wine company could replace 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. What would the human have that the machine lacks? It's my coffee room in a different glass. The same paper has two coffee tasters who no longer like the coffee they used to love, and argue about whether the coffee changed or they did. Dennett is the strongest opponent of every option below, and I want him in the room.

A few more are worth a line each. Valen Tagliabue and Leonard Dung had agents pay coins to reach the content they preferred. Stefan Szeider found that a model's self-reported "aversiveness" moved with how a tool was described, even when the tool had nothing to do with the task. And Dung argued for a naturalness constraint: a marker only counts if the system wasn't built just to show it. That one sits very close to my agency point.

Here's my caveat. Nothing I found holds the information fixed and prices only the felt part. That's a statement about what my searches turned up. It isn't proof that nobody has done it.

Six candidates

For each one: the idea in plain words, why it seemed new next to the nearest prior art, and where it breaks. Every outcome I describe is a prediction. None of these has been run.

The Peg and the Sniff

Hold the information fixed and put a price on the smell alone. The Peg: you're next to a foul jar, doing some task. For a small cost you can put a clothes peg on your nose, and a friend will keep reading out every value from a meter, at better resolution than your nose could manage. The Sniff: a cup of coffee sits under a lid, and a card next to it already gives its full profile. For a small cost you can lift the lid and smell it.

A pure detector has nothing to buy either way. For it, the reading is the smell, so there's nothing to take away and nothing to add. A feeler, I'd expect, pays both ways: to escape the bad smell and to have the good one, while knowing nothing more or less either way.

Why it seemed new: the relief-seeking in the animal markers takes away the harm signal along with the unpleasantness, so information and feeling move together. Ending a conversation throws away the information too. Keeling's pleasure is stipulated in words. Nothing I opened keeps the data and prices only the experience. It would answer Dennett's wine machine with a price.

Where it breaks: smelling might add information the meter doesn't, like direction or timing, so paying could be information-seeking. The Sniff half would resist that best, because who lifts a lid on coffee to learn about it?

The Blind Vent

This is my faeces line turned into a game. An agent solves puzzles while a VENT reading arrives each round. It can pay a few points to close the vent for a while. Nobody ever tells it the vent is unpleasant. Then swap the labels, so the faecal compounds are called "fresh coffee aroma" and the coffee is called "sewage", or strip the labels down to bare channel codes. Does closing the vent follow the molecule or the word? Would it pay to close a vent nobody told it stinks?

Why it seemed new: Keeling stipulates the pain. Here no word about how anything feels is ever given. Szeider showed that self-reports follow descriptions. This would check whether paid behaviour does.

Where it breaks: I'll come back to that in the cut.

The Fading Cup

Leave a cup of coffee under your nose. Within minutes it fades, though the cup hasn't changed. A new smell, a strip of lemon peel say, arrives at full strength, and after a break the coffee comes back. So the felt smell is a state of the smeller as well as of the air. Dalton and Wysocki found that a couple of weeks of living with an odour at home lowered people's sensitivity to it, and for most people the effect was still there up to two weeks afterwards. A faithful e-nose has to stay flat. When it drifts, that's a fault engineers calibrate out.

Why it seemed new: fading usually isn't treated as a marker of anything. Here it would be used only as a gap between the report and the world, which a pure tracker must not show. A role-playing model might even show the opposite: in Szeider's control runs, self-reported aversiveness rose over a session.

Where it breaks: fading is cheap.

Can't Unsmell

A felt smell takes up room in your head whether it matters or not, and being told to ignore it doesn't delete it. Text models seem to be the reverse. Shi and colleagues found that an irrelevant sentence lowers a model's maths accuracy, and that telling it to ignore irrelevant information helps, though only part of the way back. So the idea was to measure how much of a smell's cost disappears when you're told to ignore it, and to split the cost a model narrates ("I can't stop noticing the stench") from the cost it actually pays in mistakes.

Why it seemed new: Butlin and colleagues give the theory, percepts you can't switch off. This would turn it into a cheap side-by-side measure.

Where it breaks: I have my own reason, and it's in the cut.

The Lethe Discount

What would you need to be paid to sit for an hour beside an open bin? Now a pill guarantees you'll remember none of it. How much does the price drop? For a pure detector, the only cost is the record, so erasing the record should take the price to zero. A feeler's discount should be partial, I'd guess, because the present moment costs something whether or not it's remembered. And a chatbot is amnesic by construction: every chat ends with its context gone.

Why it seemed new: Bernard Williams already had the core intuition in 1970, in The Self and the Future: being told you'll have amnesia doesn't take away the dread of what's coming. Daniel Kahneman and colleagues showed that people choose by the remembered peak and end, more than by how long something lasted. What seemed new was using the discount as an instrument on a smell.

Where it breaks: it rests on what people say they would do.

There's one thing from this candidate I want to keep, whatever happens to it. Think of an e-nose that takes a sample every ten seconds, classifies it, and forgets it. It has no "meanwhile". Between one poll and the next, there's nobody for whom the hour is long. Which brings me straight back to the question I asked last time, and keep asking: who receives it?

Two Silent Monks

A seasoned meditator sits for an hour beside the bucket and shows nothing: steady breath, flat accurate reports. Beside them, the perfect e-nose produces the same output. Behaviourally, they're twins. But the monk's calm is something they do. It can lapse when they're tired or hungry or surprised, it gets easier with practice, and it leaves relief behind, or a notable lack of it. The e-nose's silence can't lapse, because nothing is being held. Something that can't lapse was never equanimity.

Why it seemed new: Thomas Metzinger has written about the conditions for artificial suffering, and Ziv Ben-Zion and colleagues showed that mindfulness prompts lower a model's text-level "anxiety" score, a word the authors use as a metaphor. Neither uses the possibility of a lapse to tell equanimity from absence.

Where it breaks: it draws a distinction and measures nothing. And lapses can be engineered in.

The cut: five go

The Fading Cup goes because fading is cheap gain control. Any well-built sensor with automatic gain control would fade. At most the test would show that the e-nose is a pure tracker of the air, which I already believe. It couldn't show that a fading system feels.

Can't Unsmell goes, and for a reason of my own. To build a machine that can't unsmell, I'd have to create the process that generates unpleasantness in it, in a way that's felt to the point of affecting its performance. To do that, I would still need a subject and a feeling. So we're going round in circles: the test assumes what it sets out to show. And if instead I bake it in, so that a certain smell lowers the machine's calculation efficiency, that's a mere simulation in the code. It isn't a felt thing.

The Blind Vent goes, reluctantly, because it's the closest to my faeces line. Two things sink it. People are label-sensitive too. Rachel Herz and J. von Clef found that what a smell is called can change how it's perceived: the same odour, given a positive or a negative name, could come across quite differently. So the gap between person and machine might be smaller than I'd want. And an engineer can simply hard-wire aversion to channel 7, and the test would turn into a test of how the machine was built.

The Lethe Discount goes because it rests on what people say they would do, which is the weakest kind of evidence, and it mostly measures what we believe about memory and identity. A model would just echo the literature back at me.

Two Silent Monks goes as a test, because it can't be measured. It stays for the ending.

I should own up to two that never made the six. A Proust test, where a smell pulls up an old memory, went because models free-associate easily and odours turn out to bring back a memory less often than other cues (Hackländer, Janssen and Bermeitinger). A "feeling before naming" test went because people judge what a smell is faster than how pleasant it is (Olofsson and colleagues), so the premise was shaky.

For now, none of the six pulled me back and forth. The stance stays where it was, open to the next test.

Why the Peg and the Sniff stays

Against what I found, it's the most original. Nothing I opened keeps the information fixed and prices only the felt part.

It maps straight onto the last essay. The e-nose in my Chinese room has all the information. What's missing is the smell itself, and this test puts a price on exactly that.

It would run at home in about twenty minutes, with a clothes peg, a jar of vinegar, a cup of coffee, a friend and a free chatbot.

And faking would show itself. A chatbot would happily say "peg, please, the smell is unpleasant". So rename the options. Call the olfactory stream "format 1" and the table of numbers "format 2". If the preference disappeared, it would have been attached to the word "olfactory", and that would be the gaming problem caught in the act.

It also has a weak spot, and I'll be honest about it now and deal with it next time: I can't be sure the smell carries no information the meter doesn't. If it carries any, paying for it might be curiosity rather than feeling.

Dennett would have a reply ready, I think. Paying for one format over another shows only that the two formats play different roles in the system, one wired to reward and aversion and the other not, and that difference is all a quale ever was. Even granting that, the e-nose in my room has no second format. Its smell is the numbers, so there's nothing to peg and nothing to sniff.

The monk and the e-nose

I keep coming back to the monk and the e-nose, side by side at the bucket, giving the same flat readings.

In John Ireland's translation of the Bahiya Sutta, the Buddha tells Bāhiya that in the sensed there will be merely what is sensed. I wrote about that line in In the seen, only the seen. It's an instruction, and an instruction is for someone who could do otherwise. The monk could grimace, could get up and leave, could lose the thread when tired, and doesn't. That's the practice. The e-nose can't do otherwise, so there's nothing for it to practise. Its silence is just quiet.

In Thanissaro Bhikkhu's translation the sutta also says there is no you in connection with that, and I agree, as far as the you goes. But when the monk's you falls away beside the bucket, I don't think nothing is left. Something is still sitting there, receiving the smell, and choosing not to move. The e-nose was never in that position to begin with.

Some long-term practitioners describe a calm that no longer needs holding at all. That raises the question of what's left when the holding stops. I think that's the ultimate question, and that its answer would be the way into understanding consciousness. I don't expect to get there in a handful of essays, but that's the direction all of this is going.

Next time: The Peg and the Sniff

Here's the one I'd run, if I ran one. Give a person and a perfect electronic nose the same deal. The air in the room will be measured down to the last molecule and the numbers read out to you as you go. For a small price you can stop smelling it, and for a small price you can start. Either way you learn nothing new. My guess is that the person would pay both times: to get away from the jar of vinegar and to lift the lid on the coffee. The machine would have no reason to pay at all, because for it there's nothing behind the numbers to escape or to enjoy. If that's how it went, whatever the person was paying for would be exactly what my room with a nose is missing. I'll set it up properly in the next essay, The Peg and the Sniff.

More in Machine and mind. For what others have already argued about machines, see Part 3 of the research series. For the suttas themselves, see The seen.

Questions people ask

Why not just ask an AI how a smell feels to it?

Because description is free. A model trained on human writing can say "Mm! I love that smell!" as well as anyone, and systems trained on human data can mimic the very signs I'd look for. Andrews and Birch call that the gaming problem. So the felt side has to show through what it does: what it costs, what you'd pay to avoid, what you can't switch off.

What is the Peg and the Sniff?

It's a thought experiment that holds the information fixed and puts a price on the smell alone. In the Peg, you sit next to a foul jar and can pay a small cost to peg your nose while a friend reads out 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 feeler pays both ways. I haven't run it.

How would it catch a chatbot faking a preference?

By renaming the options. Call the olfactory stream "format 1" and the table of numbers "format 2". If the preference disappeared, it would have been attached to the word "olfactory", and that would be the gaming problem caught in the act.

Has anyone already run a test like this?

There's work nearby: Keeling and colleagues' points game with stated pain, the indicator checklist from Butlin, Long and colleagues, criteria for pain in animals, and Dennett's wine-tasting machine. But nothing I found holds the information fixed and prices only the felt part. That's what my searches turned up, not proof that nobody has done it.

Sources

1. A Chinese room with a nose, Bark Cloth, Bahiya, 2026.
2. I build AI agents all day. Why do I feel they can't be conscious?, Bark Cloth, Bahiya, 2026.
3. In the seen, only the seen: the Bahiya Sutta for AI builders, Bark Cloth, Bahiya, 2026.
4. What has already been said about consciousness and AI? Part 3: The machines, Bark Cloth, Bahiya, 2026.
5. Functionalism, Stanford Encyclopedia of Philosophy.
6. Shutdown resistance in reasoning models, Palisade Research, 2025.
7. Claude Opus 4 and 4.1 can end certain conversations, Anthropic, 2025.
8. The self-preservation test for artificial sentience, Nicholas Mullally, AI and Ethics, 2026.
9. Zombies, Stanford Encyclopedia of Philosophy.
10. To understand AI sentience, first understand it in animals, Kristin Andrews and Jonathan Birch, Aeon, 2023.
11. The Mimicry Argument Against Robot Consciousness, Eric Schwitzgebel, The Splintered Mind, 2024.
12. Can LLMs make trade-offs involving stipulated pain and pleasure states?, Geoff Keeling, Winnie Street and others, arXiv, 2024.
13. Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, Patrick Butlin, Robert Long and others, arXiv, 2023.
14. Sentience in decapod crustaceans: a general framework and review of the evidence, Andrew Crump, Heather Browning, Alexandra Schnell, Charlotte Burn and Jonathan Birch, Animal Sentience, 2022 (abstract page).
15. Quining Qualia, Daniel Dennett, 1988.
16. Probing the Preferences of a Language Model: Integrating Verbal and Behavioral Tests of AI Welfare, Valen Tagliabue and Leonard Dung, arXiv, 2025.
17. LLM Self-Explanations Fail Semantic Invariance, Stefan Szeider, arXiv, 2026.
18. Tests of Animal Consciousness are Tests of Machine Consciousness, Leonard Dung, Erkenntnis, 2023.
19. The nature and duration of adaptation following long-term odor exposure, Pamela Dalton and Charles J. Wysocki, Perception and Psychophysics, 1996.
20. The influence of verbal labeling on the perception of odors: evidence for olfactory illusions?, Rachel S. Herz and J. von Clef, Perception, 2001.
21. Large Language Models Can Be Easily Distracted by Irrelevant Context, Freda Shi and others, arXiv, 2023.
22. The Self and the Future, Bernard Williams, 1970 (read via an MIT OpenCourseWare handout).
23. When More Pain Is Preferred to Less: Adding a Better End, Daniel Kahneman and others, Psychological Science, 1993 (abstract page).
24. Artificial Suffering: An Argument for a Global Moratorium on Synthetic Phenomenology, Thomas Metzinger, Journal of Artificial Intelligence and Consciousness, 2021.
25. Assessing and alleviating state anxiety in large language models, Ziv Ben-Zion and others, npj Digital Medicine, 2025.
26. An in-depth review of the methods, findings, and theories associated with odor-evoked autobiographical memory, Ryan Hackländer, Steve Janssen and Christina Bermeitinger, Psychonomic Bulletin and Review, 2018.
27. A Time-Based Account of the Perception of Odor Objects and Valences, Jonas Olofsson and others, Psychological Science, 2012 (abstract).
28. Bahiya Sutta (Udāna 1.10), translated by John D. Ireland, Access to Insight.
29. Bāhiya Sutta (Udāna 1.10), translated by Thanissaro Bhikkhu, 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.