How SIAI could publish in mainstream cognitive science journals

by lukeprog 9 min read9th Mar 201178 comments


Newest edit: I just realized that by "philosophy journals" in the original post I really meant "cognitive science" journals. (I made the mistake because for me, philosophy basically just is cognitive science.) So please the read the below in terms of cognitive science journals, not just philosophy journals. 

First edit: Some people apparently read this as an "ultimatum" for SIAI, which was not the intent at all. It's merely an argument for why I think SIAI could benefit from publishing in mainstream journals, and then some advice on how to do it. I'm making recommendations, not "demands" - how silly would that be? Also, it's not like I'm saying SIAI should do a bunch of stuff, and I'm then walking away. For example, I'm actively writing a journal-grade paper on Friendly AI, putting it in the context of existing literature on the subject. And I'd love to do more.

Also, I suspect that many at SIAI already want to be publishing in mainstream philosophy journals. The problem is that it requires a fair amount of resources and know-how to do so (as the below post shows), and that takes time. It doesn't appear that SIAI has anyone whose primary training is in philosophy, because they've (wisely) invested their resources in, you know, math geniuses and people who can bring in funds and so on. Anyhoo...


After reading about 80% of the literature in the field of machine ethics, I've realized that the field hasn't quite caught up to where Yudkowsky's thinking was (on the most important issues) circa 2001.*

One cause of this may be the fact that unlike almost every other 10-year research institute or university research program on the planet, SIAI has no publications in established peer-reviewed journals. This fact has at least two effects: (1) SIAI's researchers are able to work more quickly on these problems when they are not spending their time reading hundreds of mostly useless papers from the mainstream literature, and then composing arduously crafted papers that conform to the style and expectations of the mainstream community, citing all the right literature. And: (2) the mainstream community has not caught up with SIAI's advances because SIAI has not shared them with anyone - at least not in their language, in their journals, to their expectations of clarity and style.

However, I suspect that SIAI may now want to devote some resources doing what must be done to get published in mainstream journals, because (1) many donors do know the difference between conference papers and papers published in mainstream journals, and will see SIAI as more valuable and credible if they are publishing in mainstream journals, (2) SIAI's views will look less cult-like and more academically credible in general if they publish in mainstream journals, and (3) SIAI and LW people will need to spend less time answering dumb questions like "Why not just program the AI to maximize human happiness?" if SIAI publishes short, well-cited, well-argued responses to such questions in the language that everybody else knows how to understand, rather than responding to those questions in a way that requires someone to read a set of dozens of blog posts and articles with a complex web of dependencies and an unfamiliar writing/citation style and vocabulary. Also: (4) Talking in everyone else's language and their journals will probably help some really smart people make genuine progress on the Friendly AI problem! Gert-Jan Lokhorst is a really smart guy interested in these issues, but it's not clear that he has read Yudkowsky. Perhaps he's never heard of Yudkowsky, or if he has, he doesn't have time to risk spending it on something that hasn't even bothered to pass a journal's peer review process. Finally, bringing the arguments to the world in the common language and journals will (5) invite criticism, some of which will be valid and helpful in reformulating SIAI's views and giving us all a better chance of surviving the singularity.

Thus, I share some advice on how to get published in philosophy journals. Much of SIAI's work is technically part of the domain of 'philosophy', even when it looks like math or computer science. Just don't think of Kant or Plato when you think of 'philosophy.' Much of SIAI's work is more appropriate for math and computer science journals, but I'm not familiar with how to get published in those fields, though I suspect the strategy is much the same.

Who am I to share advice? I've never published in a philosophy journal. But a large cause of that fact is that I haven't tried. (Though, I'm beginning on early drafts of some journal-bound papers now.) Besides, what I share with you below is just repeating what published authors do say to me and online, so you're getting their advice, not particularly mine.

Okay, how to get published in philosophy journals...

The easiest way to get published is to be a respected academic with a long publication history, working at a major university. Barring that, find a co-author or two who fit that description.

Still, that won't be enough, and sometimes the other conditions below will be sufficient if you don't match that profile. After all, people do manage to build up a long publication history starting with a publication history of 0. Here's how they do it:

1. Write in the proper style. Anglophone philosophy has, over the years, developed a particular style marked by clarity and other norms. These norms have been expressed in writing guides for undergraduate philosophy students here, here, and elsewhere. However, such guides are insufficient. Really, the only way to learn the style of Anglophone philosophy is to read hundreds and hundreds of journal articles. You will then have an intuitive sense of what sounds right or wrong, and which structures are right or wrong, and your writing will be much easier because you won't need to look it up in a style guide every time. As an example, Yudkowsky's TDT paper is much closer to the standard style than his CEV paper, but it's still not quite there yet.

2. Use the right vocabulary and categories. Of course, you might write a paper aiming to recommend a new term or new categories, but even then you need to place your arguments in the context of the existing terms and categories first. As an example, consider Eliezer's Coherent Extrapolated Volition paper from 2004. The paper was not written for journals, so I'm not criticizing the paper. I'm explaining how it would need to be written differently if it was intended for journal publication. Let's pretend it is now 2004, and I am co-writing the Coherent Extrapolated Volition paper with Eliezer, and we want to publish it in a mainstream journal.

First, what is Eliezer's topic? It is the topic of how to design the goal system of an AI so that it behaves ethically, or in ways that we want. For a journal paper, our first goal would be to place the project of our paper in the context of the existing literature on that subject. Now, in 2004, it wasn't clear that this field would come to be called by the term "machine ethics" rather than by other terms that were floating around at the time like "artificial morality" (Danielson, 1992) or "computational ethics" (Allen et al., 2000) or "friendly ai" (Yudkowsky, 2001). So, we would probably cite the existing literature on this issue of how to design the goal system of an AI so that it behaves ethically (only about a two dozen works existed in 2004) and pick the terms that worked best for our purpose, after making clear what we meant by them.

Next, we would undertake the same considerations for the other concepts we use. For example, Eliezer introduces the term volition:

Suppose you're faced with a choice between two boxes, A and B. One and only one of the boxes contains a diamond. You guess that the box which contains the diamond is box A. It turns out that the diamond is in box B. Your decision will be to take box A. I now apply the term volition to describe the sense in which you may be said towant box B, even though your guess leads you to pick box A.

But here, it's unnecessary to invent a new term, because philosophers talk a lot about this concept, and they already have a well-developed vocabulary for talking about it. Eliezer is making use of the distinction between "means" and "ends," and he's talking about "informed desires" or "informed wants" or "what an agent would want if fully informed." There is a massive and precise literature on this concept, and mainstream journals would expect us to pick one variant of this vocabulary for use in our paper and cite the people who use it, rather than just introducing a brand new term for no good reason.

Next, when Eliezer writes about "extrapolating" human volition, he actually blends two concepts that philosophers keep distinct for good reasons. He blends the concept of distinguishing means from ends with the notion of ends that change in response to the environment or inner processes. To describe the boxes example above, a mainstream philosopher would say that you desired to choose box A as a means, but you desired the diamond in box B as an end. (You were simply mistaken about which box contained the diamond.) Eliezer calls this a type of "extrapolation," but he also refers to something else as "extrapolation":

In poetic terms, our coherent extrapolated volition is our wish if we knew more, thought faster, were more the people we wished we were, had grown up farther together; where the extrapolation converges rather than diverges, where our wishes cohere rather than interfere; extrapolated as we wish that extrapolated, interpreted as we wish that interpreted.

This is a very different thing to the mainstream philosopher. This is an actual changing of (or extrapolating of) what one desires as an end, perhaps through a process by which reward signals reinforce certain neural pathways, thus in certain circumstances transforming a desire-as-means into a desire-as-end (Schroeder, 2004). Or, in Yudkowsky's sense, it's an "extrapolation" of what we would desire as an end if our desires-as-ends were transformed through a process that involved not just more information but also changes to our neural structure due to environment (such as growing up together).

This kind of unjustified blending and mixing of concepts - without first putting your work in the context of the current language and then justifying your use of a brand new language - is definitely something that would keep our paper out of mainstream journals. In this case, I think the mainstream language is just fine, so I would simply adopt it, briefly cite some of the people who explain and defend that language, and move forward.

There are other examples, right from the start. Eliezer talks about the "spread" of "extrapolated volition" where a mainstream philosopher would talk about its uncertainty. He talks about "muddle" where a mainstream philosopher would call it inconsistency or incoherence. And so on. If we were writing the CEV paper in 2004 with the intention of publishing in a mainstream journal, we would simply adopt the mainstream language if we found it adequate, or we would first explain ourselves in terms of the mainstream language and then argue in favor of using a different language, before giving other arguments in that brand new language.

Same goes for every other subject. If you're writing on the complexity of wishes, you should probably be citing from, say, OUP's recent edited volume on the very latest affective neuroscience of pleasure and desire, and you should probably know that what you're talking about is called "affective neuroscience," and you should probably know that one of the leading researchers in that field is Kent Berridge, and that he recently co-edited a volume for OUP on exactly the subject you are talking about. (Hint: neuroscience overwhelmingly confirms Eliezer's claims about the complexity of wishes, but the standards of mainstream philosophy expect you to cite actual science on the topic, not just appeal to your readers' intuitions. Or at least, good mainstream philosophy requires you to cite actual science.)

I should also mention there's a huge literature on this "fully informed" business, too. One of the major papers is from David Sobel.

3. Put your paper in context and cite the right literature. Place your work in the context of things already written on the subject. Start with a brief overview of the field or sub-field, citing a few key works. Distinguish a few of the relevant questions from each other, and explain exactly which questions you'll be tackling in this paper, and which ones you will not. Explain how other people have answered those questions, and explain why your paper is needed. Then, go on to give your arguments, along the way explaining why you think your position on the question, or your arguments, are superior to the others that have been given, or valuable in some other way. Cite the literature all along the way.

4. Get feedback from mainstream philosophers. After you've written a pretty good third draft, send it to the philosophers whose work you interact with most thoroughly. If the draft is well-written according to the above rules, they will probably read it. Philosophers get way less attention than scientists, and are usually interested to read anything that engages their work directly. They will probably send you a few comments within a month or two, and may name a few other papers you may want to read. Revise.

5. Submit to the right journals. If you have no mainstream academic publishing history, you may want to start conservatively and submit to some established but less-prestigious journals first. As your mainstream academic publishing record grows, you can feel more confident in submitting to major journals in your field - in the case of CEV, this would be journals like Minds and Machines and IEEE Intelligent Systems and International Journal of Applied Philosophy. After a couple successes there, you might be able to publish in a major general-subject philosophy journal like Journal of Philosophy or Nous. But don't get your hopes up.

Note that journals vary widely in what percentage of submissions they accept, how good the feedback is, and so on. For that, you'll want to keep track of what the community is saying about various journals. This kind of thing is often reported on blogs like Leiter Reports.

6. Remember your strengths and weaknesses. If this process sounds like a lot of work - poring through hundreds of journal articles and books to figure out what the existing language is for each concept you want to employ, and thinking about whether you want to adopt that language or argue for a new one, figuring out which journals to submit to, and so on - you're right! Writing for mainstream journals is a lot of work. It's made much easier these days with online search engines and digital copies of articles, but it's still work, and you have to know how to look for it. You have to know the names of related terms that might bring you to the right articles. You have to know which journals and publishers and people are the "big names." You have to know what the fields and sub-fields of philosophy (and any relevant sciences) are, and how they interact. This is one advantage that someone who is familiar with philosophy has over someone who is not - it may not be that the former is any smarter or creative than the latter, it's just that the former knows what to look for, and probably already knows what language to use for a greatly many subjects so he doesn't have to look it up. Also, if you're going to do this it is critical that you have some mastery over procrastination.

Poring through the literature, along with other steps in the process of writing a mainstream philosophy paper, is often a godawful slog. And of course it helps if you quite simply enjoy research. That's probably the most important quality you can have. If you're not great with procrastination and you don't enjoy research but you have brilliant and important ideas to publish, team up with somebody who does enjoy research and has a handle on procrastination as your writing partner. You can do the brilliant insight stuff, the other person can do the literature slog and using-the-right-terms-and-categories part.

There is tons more I could say about the subject, but that's at least a start. I hope it's valuable to some people, especially if you think you might want to publish something on a really important subject like existential risks and Friendly AI. Good luck!




* This is not to say the field of machine ethics is without valuable contributions. Far from it!