I here make the case against smoking in the Smoking Lesion, a hypothetical decision scenario (where smoking doesn’t cause cancer) studied in decision theory. For the case against smoking in real life (where smoking does cause cancer), see, for example, the Wikipedia article on the health effects of tobacco or the Cleveland Clinic article about smoking.
I assume basic familiarity with the debate about Newcomb’s problem and the basic mechanisms of evidential and causal decision theory (EDT and CDT). I assume some sympathy with one-boxing.
The Smoking Lesion goes as follows:
Some people have a genetic lesion that causes cancer and also makes them more likely to smoke. As a result, smokers get cancer much more often than non-smokers. Unlike in the real world, smoking itself doesn’t cause cancer. You don’t know whether you have the lesion. You enjoy smoking, but cancer is much worse than smoking is pleasant. Should you smoke?
People usually use the Smoking Lesion to argue against EDT. In particular, they claim that EDT reasons as follows: Smoking is evidence that you have the lesion. So conditional on smoking, you’re more likely to get cancer than conditional on not smoking. Since cancer is much worse than smoking is pleasant, you shouldn’t smoke.
People typically take this reasoning to be mistaken. They then conclude that EDT is wrong and turn to some other kind of decision theory that doesn’t use conditionals in the way that EDT does. (Some people still try to get their new theory to one-box. Examples of such theories are timeless decision theory (TDT), functional decision theory (FDT), and Spohn’s version of CDT (Spohn 2012). Others take the Smoking Lesion to be a reductio ad absurdum of one-boxing and thus end up endorsing a theory like CDT that two-boxes.)
The mainline response of the EDT advocate is the so-called tickle defense (see, e.g., Oesterheld 2018 for an overview). The tickle defense aims to argue that for any realistic version of the Smoking Lesion, EDT also recommends smoking. Very roughly, the tickle defense goes as follows: Your decision depends only on your beliefs, your desires, and your decision theory. So the lesion can make you smoke only by influencing one of these – for example, by giving you an urge (a “tickle”) to smoke. You know your own beliefs, desires, and decision theory before you decide. In particular, you know whether you have the urge. Once you know all this, whether you smoke gives you no further evidence about whether you have the lesion. So EDT recommends smoking.
I here take a different (albeit related) perspective: I make the case against smoking. That is, I argue that in versions of the Smoking Lesion where EDT recommends not smoking, it seems very reasonable not to smoke. Among the most-discussed decision theories, EDT is the only one that recommends not smoking in (some versions of) the Smoking Lesion, so I think the Smoking Lesion is an argument in favor of evidential decision theory.
Admittedly, I think the ultimate version of EDT in the correct ontology (e.g., updateless EDT, logical-zombie EDT) might end up smoking in many versions of the Smoking Lesion in which a more naive version of EDT (as discussed here) avoids smoking (see Treutlein and Oesterheld 2017, pp. 16–17 and n. 19; though cf. Demski 2017). (Discussing this is outside the scope of the present post.) Nonetheless, I believe the Smoking Lesion favors EDT, because at least some version of EDT can rationalize not smoking, whereas the other most-discussed theories are specifically designed to recommend smoking. A second possibility is that in the Smoking Lesions where one shouldn’t smoke, logi-causalist theories (such as TDT and FDT) actually also recommend not smoking. (For an argument along these lines, see Levinstein and Soares 2020, §5.2, on the Psychopath Button problem (Egan 2007), which is similar to the Smoking Lesion. They argue that FDT refrains from pressing the button if whether one is a psychopath depends on the output of one’s decision algorithm. Cf. also Cooper, Oesterheld, and Conitzer 2024, who study CDT agents that give some credence to being any part of the environment whose behavior depends on their policy, roughly as if that part were simulating them.)
My arguments are mostly intuition pumps. In the first section, I argue that the classic Smoking Lesion leaves it unclear what EDT recommends. Roughly, for EDT to recommend not smoking, the correlation between smoking and cancer has to hold among people who are similar to you at the time of the decision: they know about the correlation, they’re torn between the same decision-theoretic arguments as you, and so on. In versions of the Smoking Lesion where this holds, I find it intuitive not to smoke. In the second section, I argue that our intuitions about the Smoking Lesion come in part from our experience with how real-world genes might cause health effects and behaviors. I then discuss two cases from the literature that share the Smoking Lesion’s causal structure but have no medical framing. I find EDT’s recommendation intuitive in both. The third section applies two standard intuition pumps for one-boxing to the Smoking Lesion: a repeated version of Newcomb’s problem and a version with a perfect predictor. As a bonus, I also give a variant of the Smoking Lesion in which a company makes money from people who follow CDT, TDT, or FDT. The variant is based on my Adversarial Offer scenario.
Acknowledgments: Thanks to Emery Cooper for comments on earlier drafts of this post. Also, I used LLMs a bunch more than for most of my previous posts. I’d estimate that about 30% of the characters in this post were generated by AI (though based on extensive instruction from me).
Deriving guidance from studies – analyzing the classic Smoking Lesion case
Consider the following version of the Smoking Lesion: An observational study shows that smokers are more likely to have cancer. Additionally, an intervention study shows that when people are randomly told whether to smoke or not, smokers are no more likely to get cancer. Should you smoke?
It’s actually quite unclear what EDT says you should do. We have two reference classes of people, each deciding in some sense whether to smoke or not. In one of them, smokers were more likely to get cancer; in the other, they weren’t. To figure out whether your decision to smoke gives you evidence, you have to figure out which reference class is more relevant to your own situation.
Naively, you might say that the no-intervention reference class is more relevant. After all, they’re making their decision in a normal way, similar to how you’re going to make a decision, while the intervention class randomizes (which is pretty different).
But, absent specifications to the contrary, the subjects of the observational study might be different from you in various ways. You’re currently deciding according to EDT based on a set of studies that you’ve read about. From the information above, it’s entirely plausible that the people in the study decide very differently. In fact, most people don’t even know what evidential decision theory is! (Shocking, I know.) Some people know about evidential decision theory but reject it! Perhaps the people in the study don’t know anything about a correlation between smoking and cancer. (Indeed, the natural assumption is that they don’t. There was no talk of an earlier study showing a correlation, after all.) It’s not even clear that the people in the study dislike getting cancer! (This is an alternative world in which cancer is never caused by smoking. Maybe besides its causes, other features of cancer, such as its reputation, are also different in this world.)
To make it clear that EDT advises against smoking, we need to make sure that at the point of making the decision, you’re quite similar to the people in the observational study. In particular:
- We need to make sure that the people in the study have similar information about the smoking–cancer correlation. That is, they too should know, before deciding, that smokers get cancer much more often than non-smokers, even though smoking doesn’t cause cancer.
- This poses a regress issue: Where do the subjects of the study get this information from? It seems that there must have been a prior study. But if there’s a prior study, where do the subjects of that study get their information from? And so on. Let’s just sweep this under the rug. (Perhaps we can borrow an idea from how the theists resolve infinite regress problems and suppose that the original study is (provided by) God and thus doesn’t require any human subjects.)
- They need to think about decision theory in similar ways. For instance, let’s say you’re currently torn between EDT and TDT (a theory that recommends one-boxing but was designed to recommend smoking). Then the study’s subjects should, when they start deliberating, have similar decision-theoretic uncertainties.
As soon as I internalize that the study subjects are really quite similar to me, I find it intuitive that I shouldn’t smoke.
To make this more intuitive, we can embellish the story further. For instance, suppose that in the observational study, they had people reason out loud and published transcripts. Many of the transcripts read like the kinds of thoughts you would have: “I do like one-boxing in Newcomb’s problem, but is the current case really analogous?”; “Perhaps the tickle defense applies…”; “ChatGPT says I should smoke, and ChatGPT is always right.”; “Hasn’t Arif Ahmed shown that causality doesn’t matter?”; … Some participants smoked based on elaborate decision-theoretic considerations that you’re sympathetic to. Most of them, unfortunately, now have cancer. Other participants refrained based on similarly elaborate and compelling considerations. Few of those participants have cancer.
Removing the medical framing
I believe our intuition about cases like the Smoking Lesion is confounded by our intuitions about how real-world genes might affect one’s health and one’s behavior. If I imagine a real-life gene that causes some health problem, I imagine a gene that works through specific kinds of mechanisms – say, it produces too much of some enzyme, and having too much of that enzyme is bad for me. If I imagine that same gene also causing some behavior, then I’d usually imagine this working through some relatively simple mechanism – say, the enzyme causes some unpleasant feeling, and some ingredient of cigarettes neutralizes this downstream effect of the enzyme. Noticing a craving or the unusually pleasant effect of smoking gives me evidence that I have the gene. Once I learn that I have, say, the craving, my eventual decision to smoke gives me no further evidence that I have the gene, assuming that the craving is the only way the gene influences my decision. (This is essentially the tickle defense again.) A related argument: If I don’t know that I have the craving, but I subconsciously decide based on the craving, then finding myself smoking is evidence that I have cancer. But finding that I smoke based on abstract decision-theoretic considerations (as one supposes one does when discussing these kinds of scenarios in a decision-theoretic context) gives me no such evidence. It seems extremely unlikely that a gene, and in particular a gene that causes cancer, can also influence our final decision about smoking in a way that is robust to us reading studies, thinking about causal and evidential decision theory, and so on.
To remove the influence of our real-world experience, we can look at cases in which (as in the Smoking Lesion) a common cause influences both your decision and the outcome, but which (unlike the Smoking Lesion) don’t hinge on mechanisms that we have prior intuitions about (like genes that cause cancer). Since others have given such examples, I’ll just cite them here and refer to these prior works for more in-depth analyses.
The Coin Flip Creation problem (Treutlein 2017a) goes as follows:
One day, while pondering the merits and demerits of different acausal decision theories, you’re visited by Omega, a being assumed to possess flawless powers of prediction and absolute trustworthiness. You’re presented with Newcomb’s paradox, but with one additional caveat: Omega informs you that you weren’t born like a normal human being, but were instead created by Omega. On the day you were born, Omega flipped a coin: If it came up heads, Omega created you in such a way that you would one-box when presented with the Coin Flip Creation problem, and it put $1 million in box A. If the coin came up tails, you were created such that you’d two-box, and Omega didn’t put any money in box A. We don’t know how Omega made sure what your decision would be. For all we know, it may have inserted either CDT or EDT into your source code, or even just added one hard-coded decision rule on top of your messy human brain. Do you choose both boxes, or only box A?
So the coin flip plays the role of the lesion: it causes both your decision and the contents of box A. CDT recommends two-boxing, since the contents of box A depend only on the coin flip. TDT and FDT also recommend two-boxing, as long as we can set up the scenario in a way that doesn’t involve predicting the agent. EDT recommends one-boxing. Treutlein suggests that even some people who would smoke in the Smoking Lesion might find it intuitive to one-box here.
Betting on the Past (Ahmed 2014, p. 120) goes as follows:
Betting on the Past: In my pocket (says Bob) I have a slip of paper on which is written a proposition P. You must choose between two bets. Bet 1 is a bet on P at 10:1 for a stake of one dollar. Bet 2 is a bet on P at 1:10 for a stake of ten dollars. So your pay-offs are as in [the table below]. Before you choose whether to take Bet 1 or Bet 2 I should tell you what P is. It is the proposition that the past state of the world was such as to cause you now to take Bet 2.
| P is true | P is false | |
|---|---|---|
| Take Bet 1 | $10 | −$1 |
| Take Bet 2 | $1 | −$10 |
So the past state of the world plays the role of the lesion: it causes your choice, and it determines whether you win your bet. You can’t influence the past, and whatever its state, Bet 1 pays more than Bet 2. So CDT recommends Bet 1. But if the world is deterministic, then P is true if and only if you take Bet 2. So if you take Bet 1, you lose $1, and if you take Bet 2, you win $1. EDT recommends Bet 2. Ahmed presents the case as a counterexample to CDT, and Treutlein finds it obvious that you should take Bet 2 (Treutlein 2017b).
I find EDT’s recommendation intuitive in both cases. Since both cases share their causal structure with the Smoking Lesion, this suggests that the common intuition in favor of smoking comes at least in part from the medical framing.
Note that for Betting on the Past to be interesting, we again have to suppose that resolving the bet doesn’t involve anything close to simulating the world, and thus you, from the past state of the world forward. Otherwise, the problem becomes very similar to Newcomb’s problem, and one-boxing theories (including TDT and FDT) will take Bet 2. We must suppose that the bet is resolved in some other way, say, by genetic analysis (similar to the Smoking Lesion) or by divine revelation.
Transferring intuition pumps for one-boxing
The Daily Lesion
A standard intuition pump for one-boxing in Newcomb’s problem is to consider a variant in which you play Newcomb’s problem many times in a row, finding the predictor to be accurate (e.g., Gardner 1973, p. 108; Ahmed 2014, §7.3.1, pp. 181–182). Then in a subsequent round, it seems quite intuitive that one should one-box.
We can apply the same intuition pump to the Smoking Lesion. Because cancer takes years to develop, I replace it with a headache that shows up within hours, and I replace smoking with a morning cup of coffee.
Daily Lesion: Every morning, you decide whether to have a cup of coffee. You enjoy coffee. You also have (and have had for a few years) a lesion that is active on some days and dormant on others. If the lesion is active on a given day, you develop a terrible headache in the afternoon; otherwise, you feel fine all day. The headache is much worse than coffee is pleasant. Coffee causally affects neither the lesion nor the headaches. (Intervention studies have shown that if you force people with the lesion to drink coffee on some days and not to drink coffee on other days, the rate at which they develop headaches is the same on coffee and no-coffee days.)
However, the lesion influences your decision: On days on which it is active, you are much more likely to end up drinking coffee. Suppose you’ve kept a diary for the past 1,000 days. You drank coffee on 500 of them, and on 480 of those, you got a headache in the afternoon. On the 500 days on which you didn’t drink coffee, you got a headache only 20 times. The numbers are about the same for the last 100 days, by which time you had long known about the correlation and were taking it into account in your deliberation. Like many decision theorists, you’re torn about the Smoking Lesion. During some periods you decided that smoking in the Smoking Lesion is rational. Those were periods of terrible headaches. Early on, you were hesitant to believe in the correlation and had a lot of coffee and a lot of headaches. A few months ago, you read Arif Ahmed’s Evidence, Decision and Causality, which made you refrain from coffee. Oh, what a great time that was. Once, you thought that EDT didn’t work because of 5-and-10 or Troll Bridge, so you drank coffee – and got a headache. In general, it feels like you’ve had all the relevant thoughts about Newcomb-like problems before and have tried acting on them. Over all these days, you’ve also looked hard for a feeling (a “tickle”) or other sign that would tell you in advance whether the lesion is active. But you’ve found none. The only predictor of headaches that you have is whether you drink coffee.
Should you drink coffee today?
CDT, TDT, and FDT all recommend drinking coffee every day. EDT, on the other hand, recommends not drinking coffee. On 96% of the days on which you drank coffee, the lesion was active, and on 96% of the days on which you didn’t, it was dormant. So if you drink coffee today, you should expect a headache this afternoon, and if you don’t, you should expect to feel fine.
I find it quite intuitive that you shouldn’t drink coffee.
Part of what makes the Daily Lesion compelling, I think, is that you learn about the correlation from your own past decisions. Earlier, I argued that to be sure that EDT recommends abstaining, we need the people in the observational study to be similar to you at the time of the decision: they need similar information about the correlation, they need to think about decision theory in similar ways, and so on. In the Daily Lesion, the people in the study are your past selves. The recent ones had the same information as you (give or take a few diary entries), thought about all the same kinds of things you can possibly think about today, and were torn between the same arguments. And the correlation held about as strongly for them as for your early selves, who had little evidence for it.
We can extend the scenario to deal with the possibility of randomization. Suppose that besides the 1,000 days above, your diary covers 200 days, picked in advance, on which you flipped a coin in the morning and let the coin decide whether you’d drink coffee. On these days, your headache was usually about half as bad as your headaches on other days, uncorrelated with how the coin landed. Perhaps on some days you rolled a die and drank coffee if it came up 3 or 6. Then on those days your headache was about 1/3 as bad on average (and uncorrelated with the outcome of the die roll). In general, if you let a random device decide and it had probability p of choosing coffee, your headache was about p · 96% + (1 − p) · 4% as bad, where 96% and 4% are your headache rates on other days with and without coffee.
TDT and FDT recommend drinking coffee every day in the Daily Lesion. For them, the relevant difference between the repeated version of Newcomb’s problem and the Daily Lesion is that the predictor’s prediction depends on your decision procedure – the predictor runs a model of it – whereas the lesion is a common cause of your decision and the headaches. When I imagine facing the next decision in either case, this difference doesn’t seem to matter much. In both cases, I have a long personal track record of a particular choice being followed by a bad outcome that the choice doesn’t cause.
Impossible outcomes
Another standard intuition pump for one-boxing considers a version of Newcomb’s problem in which the predictor is perfect (e.g., Nozick 1969, pp. 140–141; Seidenfeld 1984, pp. 203–204). Then only two of the four combinations of prediction and choice occur: one-boxing after a one-boxing prediction and two-boxing after a two-boxing prediction. Yet the reasoning of a causal decision theorist will refer to (the utility of) the other two combinations, which never occur. This seems rather strange! In an earlier blog post (Oesterheld 2017), I proposed a general principle based on this observation. I called it the irrelevance of impossible outcomes: your decision shouldn’t depend on the utilities of outcomes that can’t occur. (Herrmann and Rothfus 2025 discuss a very similar principle under the name “Deference to the Seen”.) EDT abides by this principle: it weighs the utility of each combination of action and state by the probability of the state conditional on the action, and the impossible combinations have conditional probability zero. CDT violates the principle. For example, suppose that in Newcomb’s problem, one-boxing after a two-boxing prediction paid $1,000,000 instead of $0, and two-boxing after a one-boxing prediction paid $1,000 instead of $1,001,000. Then each option has the same payoff under both predictions: one-boxing gets you $1,000,000, and two-boxing gets you $1,000. So CDT would one-box, even though the outcomes that can occur are the same as in the original problem.
We can apply the same reasoning to the Smoking Lesion. Consider the following scenario.
Perfect Lesion: Some people have a lesion that causes cancer: everyone who has the lesion gets cancer, and nobody else does. Smoking doesn’t cause cancer. The lesion also influences whether people smoke, and this correlation, too, is perfect: everyone with the lesion ends up smoking, and everyone without it ends up abstaining. Many people in this world deliberate like you, in the demanding sense from earlier: they know about the correlation, they’re torn between the same decision-theoretic arguments as you, and so on. In particular, the correlation is perfect among them. You can’t tell whether you have the lesion until you’ve made up your mind. You’re sure that you’ll enjoy smoking (everyone does in this hypothetical world), but cancer is much worse than smoking is pleasant.
Here is the outcome of each combination of lesion and action. The combinations in italics never occur.
| Lesion | No lesion | |
|---|---|---|
| Smoke | pleasure, cancer | pleasure, no cancer |
| Don’t smoke | no pleasure, cancer | no pleasure, no cancer |
As in Newcomb’s problem with a perfect predictor, each comparison in the dominance argument for smoking involves one of the combinations that never occur. The argument says that if you have the lesion, smoking gets you pleasure and cancer rather than cancer alone, and that if you don’t have the lesion, smoking gets you pleasure without cancer rather than neither. But nobody ever smokes without getting cancer, and nobody ever gets cancer without smoking. Smokers get pleasure and cancer, and non-smokers get neither.
If we take into account only the outcomes that can occur, not smoking is better, since cancer is much worse than smoking is pleasant. EDT gives the two combinations that never occur zero weight, and so it recommends not smoking. CDT, TDT, and FDT treat whether you have the lesion as independent of your choice, so they accept the dominance argument and recommend smoking. Their recommendations thus depend on what would happen in combinations of lesion and action that never occur.
Like the version of Newcomb’s problem that it’s based on, this intuition pump requires a perfect correlation. If the correlation is imperfect, all four combinations can occur, and the principle doesn’t apply. But it would be strange if the right decision switched from not smoking to smoking as soon as the correlation dropped from 100% to 99%.
Finally, the principle makes no reference to how the correlation between action and state comes about. TDT and FDT abide by it in Newcomb’s problem, where the correlation arises from a prediction, but they violate it in the Perfect Lesion, where the correlation arises from a common cause. Their proponents would need to explain why the source of the correlation should matter in this way.
Bonus: Extracting money from smokers
All the previous intuition pumps are based on the general structure of the Smoking Lesion.
As a bonus, here’s a variant of the Smoking Lesion with a very different structure, based on my Adversarial Offer scenario (Oesterheld and Conitzer 2021; cf. Spencer 2021):
Adversarial Lesion: A company runs a contest in which anyone who lifts a 1,000-kilogram weight wins $3. Nobody can lift the weight unaided, but the company sells two strength potions, A and B, at $1 each, and anyone who drinks a working potion can lift it. Each person may buy only one potion. The potions have no other effects and cost the company nothing to make. Apart from the prize, nobody cares whether anyone lifts the weight. Some people have lesion A, some have lesion B, and some have neither; nobody has both. Lesion A makes potion A ineffective, and lesion B makes potion B ineffective. The company has observed that the lesions also influence which potion people choose: among its customers, 75% of those who buy potion A have lesion A, and 75% of those who buy potion B have lesion B. Many of its customers deliberate like you, and among them, the numbers are the same. You can’t tell which lesion, if any, you have until you’ve made up your mind.
(What happens if you randomize? Well, we could simply say that randomization is not allowed. For anything more complicated, see the discussion in Section IV.1 of Oesterheld and Conitzer 2021.)
First, consider the scenario from the company’s perspective. The company takes in $1 for every potion it sells. Three quarters of its customers have the lesion that matches the potion they buy, so their potion is ineffective and they win nothing. The remaining quarter lift the weight, and the company pays each of them $3. So on average, the company makes $1 − 0.25 · $3 = $0.25 per potion. CDT, EDT, TDT, and FDT all agree that running the contest is profitable for the company, since the company’s decision to run it is uncorrelated with its customers’ lesions. The potions cost the company nothing to make and matter to its customers only through the prize. So the company’s gain is exactly its customers’ loss: on average, its customers lose $0.25 per potion.
EDT recommends buying neither potion. If you buy potion A, you pay $1. Conditional on buying potion A, you have lesion A with probability 0.75, in which case the potion is ineffective. So you win the $3 prize with probability 0.25, and the expected payoff of buying potion A is 0.25 · $3 − $1 = −$0.25. The same calculation applies to potion B. Buying neither gets you $0.
CDT, TDT, and FDT recommend buying one of the two potions. As in the classic Smoking Lesion, they treat which lesion you have (if any) as independent of your choice. So they calculate the expected payoff of buying potion A as (1 − P(lesion A)) · $3 − $1, where P(lesion A) is your probability of having lesion A before you update on your decision. Because nobody has both lesions, P(lesion A) + P(lesion B) ≤ 1, so at least one of the two probabilities is at most 1/2. According to these theories, buying a potion whose matching lesion has probability at most 1/2 has an expected payoff of at least 1/2 · $3 − $1 = $0.50.
But as we saw from the company’s perspective, its customers lose $0.25 per potion on average. The same holds for its customers who deliberate like you, since the numbers are the same among them. Following CDT, TDT, or FDT thus costs you $0.25 on average. In the Adversarial Offer, TDT and FDT decline to buy a box, as EDT does. In the Adversarial Lesion, they buy a potion and lose money along with CDT.
Related work
Others have similarly argued that refraining from smoking in the Smoking Lesion is rational, using similar intuition pumps and arguments. For example:
- Ahmed 2014, Evidence, Decision and Causality. Several parts are relevant:
- Smoking Lesion: Ahmed defends the tickle defense for realistic medical versions (§4.3, pp. 91–99). But he also argues that “Why Ain’cha Rich?” supports the dominated option in just those Newcomb-like cases where EDT recommends it (p. 191, n. 40). Applied to versions of the Smoking Lesion where EDT recommends abstaining, this supports abstaining.
- Perfect correlation: His Calvinist problem (§0.6, pp. 9–12) closely resembles the Perfect Lesion. God’s decree determines both conduct and salvation. Virtuous people are saved and sinners damned, although sin is more pleasant under either fate. Ahmed defends virtue, the analogue of not smoking.
- Reference classes: He uses a medical case, Check-up, to explain why the relevant statistics must reflect the agent’s present information – in that case, their symptoms (§7.3.3, pp. 189–191). This parallels the first section here.
- pallas 2013, “Chocolate Ice Cream After All?” Defends the evidential choice in Newcomb’s Soda and a medical version of Solomon’s Problem, two other common-cause problems. The discussion is particularly close to the first section here: it asks what the study subjects knew about the correlation, and defends the analogue of abstaining when the correlation holds among subjects who have the same information as the deliberating agent.
- Treutlein 2017a, “Did EDT get it right all along?” Explicitly argues for the evidential choice in medical Newcomb problems. His Coin Flip Creation problem, quoted above, aims to elicit one-boxing intuitions in a case with the Smoking Lesion’s common-cause structure. He also questions whether replacing the common-cause mechanism with a simulation should change the recommendation when the action–outcome correlation stays the same.
- Carlsmith 2021, “Can you control the past?” (§VIII). Considers applying the repeated-Newcomb intuition pump to the Smoking Lesion: repeatedly enter the scenario and try smoking or abstaining. He becomes more sympathetic to abstaining if the stipulated correlations survive this experimentation. His discussion remains tentative, and he questions whether those correlations could persist under unrestricted experimentation with a fixed base rate of lesions.
- Bacus 2026, “An Opinionated Introduction to Newcombology” (§2). Makes a similar point about the lesion’s mechanism: if it influences which decision-theoretic arguments you accept, abstaining becomes more intuitive. This parallels the distinction above between ordinary cravings and influences on abstract deliberation.
