Rationally Ending Intellectual Discussions

Discussions should end gracefully. There should be some clarity about why they’re ending and why that’s rationally acceptable. If someone wants to continue the discussion, they should have the opportunity to see why this ending is reasonable. (Or, failing that, they should have some evidence available to enable them to argue their case that it was unreasonable. They shouldn’t be left with little or no data about what happened.)

If you discuss poorly, it’s important that you can learn from that and do better next time. If you want it, you should have access to some critical feedback, some explanation of the other person’s perspective, or something to enable progress. If you’re running into problems interacting with people, but no one will tell you what the problem is, that’s bad for rational progress.

The more easily discussions can end, the more easily they can start. We don’t want discussing to be a big burden or commitment (it’s OK if a few discussions require high effort, when people have reason to voluntarily choose that, but that shouldn’t be the default).

Discussions can end by mutual agreement. If no one objects to the ending, that’s fine. That’s one way to end discussions. Pretty much everyone agrees on this. The difficulty is that sometimes people don’t agree. Someone wants to stop discussing but someone else wants to pursue the matter further.

I don’t think agreeing to disagree, by itself, is a good reason to end a discussion by mutual agreement. Disagreements can be resolved! There should be an extra factor like agreeing that it’s reasonable to prioritize other stuff (which we may leave implied without explicitly mentioning). There are many problems to work on, and time/energy/etc are scarce resources, so it’s fine to drop a disagreement (for now, indefinitely) if people think they can get more value working on other stuff.

Ending discussions when a single person wants to, for any reason, with no explanation, is problematic. For example, people will end discussions when they (the ideas they are biased in favor of) start losing the argument.

But we don’t want to trap people in discussions. Sometimes one guy has a lot of energy to discuss something forever but you want to stop. There are lots of legitimate reasons to end a discussion.

You can try to explain your reasons for ending a discussion, but no matter what you say, the other guy might disagree. This is a real concern because you don’t want to discuss extra with the most unreasonable people who give low quality counterarguments to all your reasons for stopping discussion. Meanwhile the most reasonable people tend voluntarily to let you out of discussions, so you discuss with them the least!?

There has to be a way to unilaterally end discussions. You end it without agreement. But it should have some protections against abuse. We don’t want it to be acceptable to arbitrarily end any discussion at any time for no reason or awful reasons. This is for your own sake, too, not just for the benefit of others. If I want to end a discussion and the other guy disagrees, I ought to consider that maybe I’m biased. I might be evading the issue, avoiding being corrected, etc. I shouldn’t just fully trust my own judgment. I should want some discussion policies that make it harder for me to be and stay irrational, biased, unreasonable, etc.

Note: Of course anyone can stop talking at any time for no reason. That’s a matter of freedom. No one should be held hostage. The issue is the consequences for your reputation. What is considered reasonable or rational? Some discussion ending behavior ought to be seen negatively by the community. Some good ways of ending discussions out to be encouraged, widespread and normal. Some bad ways of ending discussions should be disincentivized.

Note: Even if a discussion actively continues, you could participate in it on less than a daily basis. Discussions don’t have to be fast. Some norms are needed for what is stalling a discussion out (e.g. one reply per decade would be a way to pretend you didn’t end the discussion when basically you did). In my experience, people are usually pretty reasonable about discussion pacing, with a few notable exceptions. (The main problem I see is people who discuss a bunch initially and then never come back to it as soon as they go do something else or as soon as they sleep once.)

So we want a way to end a discussion, even if other people disagree with ending the discussion, but which has some protection against abuse (bad faith), error, bias, irrationality, etc. It should ideally provide some transparency: some evidence about why the discussion is ending that can be judged neutrally, positively or negatively by the audience. And it should provide some sort of feedback or learning opportunity for the guy who didn’t want to stop here.

So here’s the first draft of a policy: When you want to end a discussion, you are expected to write one final message which explains why you’re ending the discussion. You’re also expected to read one more message from the other guy, so he has one chance to point out that you’re making a mistake and he could potentially tell you why you should change your mind about ending the discussion.

What’s good about this policy? It helps limit abuse. It’s harder to end a discussion for a bad reason if you have to explain yourself. The other guy gets some info about what happened. The other guy has a chance at a rebuttal so you could potentially be corrected. And it’s not very time consuming. It puts a small, strict limit on how much more discussion happens after you decide you’d like to wrap it up.

This is a pretty minimal policy. I think it could be a good default expectation that LW could use for any discussion where people have each written 3+ messages (or maybe 5+ to reduce the burden? The number could be tuned if this was tried out for a while). That way it won’t add an extra burden to really small discussions. Tiny discussions would be discouraged by any overhead at all. Tiny discussions are also lower investment so ending them is a smaller deal. People haven’t formed a reasonable expectation of reaching a conclusion, getting their questions answered, or anything else. They’re just sharing some thoughts on an ad hoc, no-obligation basis. That’s fine. But for more substantive discussions, I think adding a little bit of an ending cost is reasonable.

The minimal policy has some downsides. If we had a policy that takes more effort, we could fix some problems and get some benefits. So I think for discussions that go longer (e.g. 10+ messages each) or when people mutually agree to make it a substantive discussion, then a longer but better approach could be used for unilaterally ending a discussion.

What are problems with the single parting message? It could be unclear. It could ignore a key issue. It could misrepresent the other guy’s positions or misrepresent what happened in the discussion. It could be poorly thought through and show major signs of bias.

What are the problems with a single rebuttal to the parting message that won’t be replied to? If it asks any clarifying questions, they won’t be answered. Any great points could be ignored without explanation.

So as a next step up, we could have a two-part discussion ending. Instead of one more back and forth (parting message + rebuttal), we could have two more back and forths. Initial parting message, rebuttal and questions, final parting message, and then final rebuttal.

BTW, the rebuttals are semi-optional. You can just decide to agree with the guy’s parting message if you want (maybe it makes sense to you once he explains his position). Or instead of a rebuttal or agreement, your other option is to write your own parting message. But you shouldn’t disagree with their parting message and then silently end the discussion with zero explanation of what’s going on.

Note: Parting messages don’t have to be very long. A few clearly written sentences can cover the key points (e.g. your opinion of the discussion, your final comments on some open issues, your reasons for ending). A bit longer is needed if you write fluff. And generally you should write a bit more if the discussion was longer.

With a two back-and-forth discussion ending, it’s still possible to dodge questions, avoid key issues, etc. It can take quite a few iterations to get some stuff cleared up, and that’s if people are being reasonable. Unreasonable people can sabotage discussions indefinitely.

So what about a three back-and-forth discussion ending? Or four or five? Nothing will be perfect or give a guarantee.

Let’s consider other approaches. What about a 5% ending? However many words you wrote in the discussion, you should write 5% of that number in the discussion ending phase. That seems kinda reasonable. That means for every 20 words of discussion you write, you’re potentially obligating yourself to one word later. This might need to be capped for very long discussions. This means your effort to end the discussion gracefully is proportional to the effort you put into the discussion.

This approach still suffers from being a fairly arbitrary cutoff. You just decide to end the discussion, say a few things that hopefully do a good job of exposing your reasoning to criticism and giving the other guy the chance to learn that he’s wrong, and say a few things to wrap up the open issues (like briefly answering a few key questions you hadn’t gotten to, so the discussion is left in a more complete form and your case is left adequately complete that someone could learn from you). I think that’s way better than nothing but still has significant potential to go wrong.

One useful technique is agreeing to escalate the commitment to the discussion. You can say “I will discuss X but only if you’ll agree to a 3 back and forth ending if you want to end the discussion when I don’t (which I’ll also agree to if I want to end it unilaterally).” It sometimes makes sense to not want to invest in a discussion then have it end abruptly in a way that’s unsatisfying and inconclusive from your perspective.

It makes sense to want a discussion to either be productive or else the other guy makes a clear claim – explained enough that you could learn from it – about what you’re doing wrong, so you have the opportunity to improve (or maybe to criticize his error and judge him, rather than being left with a lack of data). Someone could also explain why the discussion isn’t working in a no-fault way, e.g. you and he have some incompatible traits.

Saying “I’m busy” is broadly a bad excuse to end discussions. You were busy when you started, too, right? What changed? Sometimes people actually get significantly busier in an unforeseeable way in the middle of a discussion, but that shouldn’t be very common. Usually “I’m busy” is code for “I think your messages are low quality and inadequately valuable”. That claim isn’t very satisfying for the other guy without at least one example quote and some critical analysis of what is bad about the quote. Often people speak in general terms about low quality discussion without any quoted examples, which also tends to be unsatisfactory, because the person being criticized is like “Uhh, I don’t think I did that thing you’re accusing me of. I can’t learn from these vague claims. You aren’t showing me any examples. Maybe you misunderstood me or something.”

It can be reasonable to say “I thought I’d be interested in this topic but it turns out I’m not that interested.” You shouldn’t say this often but occasionally is OK. Shit happens. You can end a discussion due to your own mistake. When you do, you shouldn’t hide it. Let the other guy and the audience know that you aren’t blaming him. And maybe by sharing the problem you’ll be able to get some advice about how to do better next time. Or if you share the problem, maybe after a bunch of discussions you’ll be able to review why they ended and find some patterns and then realize you have a recurring problem you should work on.

Impasse Chains

Besides trying to end a discussion gracefully with a parting phase where a few things get explained, I have a different proposal: impasse chains.

An impasse is a statement about why the discussion isn’t working. We’re stuck because of this impasse. It’s explaining some problem in the discussion which is important enough to end the discussion (rather than being minor and ignorable). What if no one problem is ruinous but several are adding up to a major problem? Then the impasse is the conjunction of the smaller problems: group them together and explain why the group is an impasse.

Stating an impasse provides transparency and gives the other guy some opportunity to potentially address or learn from the discussion problem.

Impasses are meant to, hopefully, be solved. You should try to say what’s going wrong that, if it was changed to your satisfaction, you’d actually want to continue.

The other guy can then suggest a solution to the impasse or agree to stop. A solution can be a direct solution (fix the problem) or an indirect solution (a workaround or a better way to think about the issue, e.g. a reason the problem is misconceived and isn’t really a problem). You should also try to think about solutions to impasses yourself.

Sometimes the guy will recognize the impasse exists. Other times it’ll seem strange to him. He wasn’t seeing the discussion that way. So there’s some opportunity for clarification. Lots of times that someone wants to end a discussion, it follows some sort of fixable misunderstanding.

So far an impasse is just a way to think about a parting message, and you can hopefully see why continuing at least one more message past the initial impasse claim makes sense. So you may think this impasse approach just suggests having 2-5 messages (per person) to end discussions. And that’s decent – a lot of discussions do way worse – but I also have a different suggestion.

The suggestion is to chain impasses together.

So step 0, we discuss.

Step 1, I say an impasse and we try to solve it. This is an impasse regarding step 0.

Step 2, I decide the problem solving isn’t working in some way (otherwise I’d be happy to continue). So I state an impasse with the problem solving. This is an impasse regarding step 1.

Step 3, we try to solve the second impasse. Either this problem solving discussion satisfies me or I see something wrong with it. If it’s not working, I say an impasse with this discussion. This is the third impasse.

Each time an impasse is stated, the previous discussion is set aside and the impasse becomes the new topic of discussion. (Though a few closing comments on the previous discussion are fine and may be a good idea.) The impasse is either solved (and then you can return to the prior discussion) or leads to a new impasse. This can repeat indefinitely. You can have an impasse about the discussion of an impasse about the discussion of an impasse in the original discussion.

The impasses are chained together. Each one is linked to the previous one. This is different than multiple separate impasses with the original discussion. Here, we’re dealing with one impasse for the original discussion and then the other impasses in the chain are all at meta levels.

Note: If you see multiple impasses with the original discussion, often that means you tried to ignore one. Instead of bringing up the first one and trying to do problem solving, you let problems accumulate. It’s possible for multiple impasses to come up at the same time but it isn’t very common. In any case, you can deal with the impasses one at a time. Pick one to focus on. If it gets resolved, move on to the next one.

It doesn’t make sense to ask someone to discuss X further when he sees an impasse with discussion X (meaning a reason that discussion isn’t working). You’ll have to address that problem in some way or agree to stop. Discussing the problem itself is a different discussion than discussing X, so it should be possible to try it.

The more impasses chain, usually the clearer the situation gets. Each level tends to be simpler than the previous level. There are fewer issues in play. It becomes more obvious what to do. This helps but isn’t nearly enough to make impasse chains get addressed (either solve every impasse or agree to stop) in a reasonable amount of time.

Impasse chains often get repetitive. Problems reoccur. Suppose I think you keep saying non sequiturs. We can’t discuss the original topic because of that. So then we try to discuss that impasse. What happens? More non sequiturs (at least from my point of view)!

Some discussion problems won’t affect meta levels but some are more generic and will. You can try to say “OK given our disagreements about X, Y and Z, including methodology disagreements, what can we still do to continue problem solving which is neutral – which makes sense regardless of what’s correct about those open issues, and makes sense from both of our points of view?” Often what happens is you can’t think of anything. It’s hard. Oh well. Then you can mutually agree to end the discussion since neither of you knows a good way to continue.

When impasse chains get long, you tend to either have a lot of issues that are being set aside (given A, B, C, D, E, F, G, H are unresolved, what can we do?) or you have a lot of repetition (every meta level is the same problem, e.g. my belief that you’re replies are non sequiturs). Repetition is itself a reason to end the discussion. It’s just repeating and neither of us knows a way to fix it, so we can agree to stop.

This kind of ending is satisfying in a way. It gives transparency about why the discussion ended. It means (worst case scenario) I’ve gotten to make my case about what you’re doing wrong, and you’ve failed to answer it substantively, so now I’m reasonably content (not my favorite outcome, but what more would I hope to gain from continuing?).

So a policy can be e.g. to discuss until an impasse chain with the same problem 3 times in a row. Or to discuss until an impasse chain of at least length 5. Generally 5 is plenty. Reaching the 5th metal level can be quite fast and clarifying.

Impasses can be declared in bad faith. People can disagree about what is an impasse. Then what? Discussions have to proceed in ways that make sense to all parties. If someone thinks there is an impasse, then there is one, even if the impasse consists of his misconception. And yes bad faith is possible. What can be done about that? Transparency. Exposing it to daylight. Having systems like this that make bad faith more visible and easier to criticize. Having procedures that create more evidence of bad faith.

In general, by an impasse chain of length 5, if one person is being rational and the other isn’t, it gets really obvious who is who. This gives the rational guy reason to be satisfied with ending the discussion (he knows what happened and why, and he had some chances to try to solve it) and it gives evidence about both parties. If both people are fairly equal in skill or rationality, or both are pretty bad (even if unequally bad), then you can much more easily have muddy waters after an impasse chain of length 5. Oh well. Creating clearer impasse chains is a skill you can work on. You can learn from your attempts to create some clarity and what didn’t work and why, and try to do better next time. And you can try to learn from the other guy’s attempts to.

The impasse chain system is unnecessary for every discussion. It’s a bit heavyweight and high transaction cost to use all the time. But it’s pretty limited. If you agree to a 5 impasse chain, you’re always 5 messages away from getting out of the discussion. The only reason it’d take more is if the other guy said reasonable stuff. But if he says unreasonable stuff, you’re done in 5 messages, and some of those messages will often be just one paragraph or even one sentence long.

This approach is good when people want to claim they are open to debate and that their views have stood up to debate. That leads to disputes over what it means to be open to debate, etc. I propose being willing to enter into discussions terminated by a length 5 impasse chain (or mutual agreement) as a reasonable criterion for a (self-declared) serious intellectual to say he’s actually open to substantive debate about something and is actually addressing critics.

And the impasse chain approach can be requested when you want to have a discussion if and only if there will be a substantial ending to protect your effort investment. If you want to avoid a case of putting in a bunch of effort now and then the guy just leaves, you can ask for an impasse chain precommitment or 5 parting message precommitment or other way to help protect (and therefore enable) your energy investment into the discussion.

Concluding Thoughts

What’s the typical failure case look like? Joe is trying to have a rational discussion and Bob says “eh, your messages are lame; bye” and won’t answer questions or address arguments. Or, worse, Bob explains even less than that. If Bob would explain that much, at least people could see “OK Bob accused Joe of lame messages and gave zero arguments. Therefore Bob is lame.” Impasse chains or even just parting messages help enable problem solving as well as clarifying what happened in bad outcomes.

Often a discussion looks like this: Joe writes a blog post. Bob says some criticism. Joe sees many flaws in the criticism. Joe explains the flaws. Bob stops talking. This isn’t satisfying for Joe. He never got feedback on his post from post-misconception Bob. And Bob probably didn’t change his mind. And Bob didn’t even say what the outcome was. Or if Bob did change his mind about something, it’s often a partial change followed immediately by like a “you win; bye”. People routinely use conceding as a way to end discussions with no followups: no post mortem (learning about why the error happened and how to fix the underlying cause), no working out the consequences of the right ideas, etc. The correction doesn’t go anywhere.

That’s sad for Joe. He didn’t want to correct Bob just for fun. He wanted to correct Bob so it’d lead to something more directly beneficial to Joe. E.g. Joe’s correction could be criticized and that’d have value for Joe (he learns he was actually wrong in some way). Or Joe corrects Bob and then it leads to further discussion that’s valuable to Joe. If correcting people is pure charity – and you usually get ghosted without them admitting they were corrected – then people will even try to do it way less. There should be rewards like some sort of resolution to the issues and continuation. Discuss productively and keep going (and maybe Joe learns something later or, failing that, at least gets a good student who learns a bunch of things and may be able to suggest good ideas in the future), or say the impasse for why it’s not productive.

Often Bob thinks Joe is doing something wrong in the discussion but won’t explain it enough for Joe to have a reasonable opportunity to learn better. Note that cites are fine. If it’s already explained somewhere, link it. Just take responsibility for the material you refer people to: you’re using it as our proxy, to speak for you, so you ought to have similar willingness to address questions and criticisms as if you’d written it yourself (but if it’s popular stuff with lots of existing discussion, then you can address FAQs by referring the guy to the FAQ, can direct him to the forum for that community to get questions answered and only answer them yourself if the forum won’t answer, and you can link other blog posts, books, papers, etc. to address followup issues if those exist, etc.)

Impasse chains help address these problems and help make it harder to end discussions due to your own error and bias. And they provide opportunities to solve discussion problems instead of just giving up at the first problem, or in the alternative at least more transparency about the problems can be achieved.

See also

My prior article on Impasse Chains.

My articles on Paths Forward (about discussing in such a way that if you’re wrong and anyone knows it and is willing to tell you, you never block that off with no way for your error to be corrected), including the article Using Intellectual Processes to Combat Bias.

My debate policy.

---

View this post at Less Wrong.