Special pleading, along with feigned neutrality, is one of the most infuriating symptoms of faulty rhetoric one can utilize in an argument.
Special pleading comes in multiple forms, but the most common is that of claiming a superior framework which is proven to be superior by its own internal criterion. Vulgar Marxism and Freudian psychoanalysis both resort to this tactic by using lines like, “that you would argue against the Revolution is proof you are bourgeoisie and do not understand” or “your denial is proof of your repressed desires.” The point is that any criticism can be fallaciously transformed into proof of the original claim or be fallaciously disregarded because the critic is inherently limited by his or her own paradigm.
Kaj Sotala, Roko Mijic, and Michael Anissimov all use special pleading when critiquing James Hughes’ piece “Liberal Democracy vs Technocratic Absolutism.” The central rebuttal for all of them can be paraphrased as “your critiques of communism, dictatorships, and other authoritarian governments make sense for humans, but don’t apply to Friendly AI because Friendly AI is different than human systems and is genuinely selfless.” Hughes hears echoes of Marxist-Leninist thought in that point.
Some thinkers, including the allegedly brilliant philosopher Slavoj Zizek, continue to defend Marxism using special pleading. Instead of claiming communism isn’t based in humans, they claim Stalin and the USSR were not pure communism, and therefore were doomed to failure because of the corrupting element of capitalism. Thus, thanks to special pleading, Stalin is not proof that communism and authoritarianism are dangerous and bad, but that capitalism is bad and corrupts the pure motives of communism.
The problem is that, like communism, friendly AI, even if derived through the process described by the CEV, will ultimately fail. The reason democracy works even remotely better than authoritarian systems is because it openly admits and aims to minimize the faults in the system. These faults include both the “programming,” that is, the legislation and philosophy underpinning it, and the agents of the system, humans. Democracy, communism, and, yes, AI-based technocratic authoritarianism, are all human systems. They will be imperfect. Democracy, of the three, is the only one that sees itself as imperfect and prone to mistakes and failure. Therein lies the inherent benefits of democracy – it is a radically reflexive system.
As a final point, I think it is very interesting that those who support friendly super-AI don’t see the AI coming to the conclusion that nearly all forms of government, particularly those of an authoritarian breed, are faulty and instead advocating anarchy or a form of hyper-limited government. That the AI would want to govern at all is a further assumption I don’t understand. Assuming it’s an AI, it should be volitional, which would make forcing it to govern a restriction in its will or it would make it a program, not a genuine AI. There are just too many problems here.
Kyle Munkittrick is a graduate student at New York University where he studies critical theory and transhumanism. His areas of focus are personhood, sexuality and gender, identity construction, biopolitics/cyborg politics, and reproductive technology. He also maintains a blog, Pop Transhumanism which is an effort to apply these studies to current events and pop culture.
Print •
Email •
permalink •
Discuss in Forums •
Send to: ¡
¡
COMMENTS
"Special pleading, along with feigned neutrality, is one of the most infuriating symptoms of faulty rhetoric one can utilize in an argument. "
Your discussion of special pleading was very interesting, but I was actually hoping you'd also discuss feigned neutrality. Arguing with someone who is feigning neutrality is indeed infuriating, but then again, it is supposed to force you to deal with the issues on the table instead of focussing on the personality/leanings of your opponent.
I never understand why people continue to refer to Friendly AI in the present tense, as if it already exists. Sure, I can //define// FAI as "that which would be immune to the faults of Marxism", but then, gosh, I'd actually have to INVENT it!
When technocrats then invent something they claim to be FAI, subject the world to it's rule, and end up creating a disaster, they can then just claim "Oh well - I guess it wasn't REALLY Friendly AI after all."
Friendly AI DOES NOT EXIST in the present tense. It makes no sense to argue about it until we have a candidate we can examine.
In you final point you included a sub-point which I don't understand: your statement "Assuming it’s an AI, it should be volitional, which would make forcing it to govern a restriction in its will or it would make it a program, not a genuine AI" make no sense.
What is "genuine" intelligence, be it artificial or natural? Humans and other animals are intelligent and volitional, yet they can be controlled by others. Intelligence has evolved in niches and can be designed to operate in niches. Of course, as you would probably agree, a system operating in the niche of world governance may be very faulty and/or not what the instigators had hoped for.
> Democracy, of the three, is the only one that sees itself as
> imperfect and prone to mistakes and failure. Therein lies
> the inherent benefits of democracy – it is a radically
> reflexive system.
You don't notice the CEV dynamic as utilizing this reflexivity that (traditional) democracy also utilizes? In what way do you claim traditional democracy to be more reflective i.e. can you present an example of a situation where traditional democracy would catch a potential mistake that the CEV dynamic would not catch?
In what way do you think the CEV dynamic essentially differs from an elaborate polling system?
Also, if you claim that the plan to build a CEV-implementing FAI "sees itself as perfect", you don't seem to have read the CEV page. The very first paragraph there comments on how the writer sees the framework presented as being faulty. (Some of the other preliminary paragraphs explain that he anyway feels he needs to say something about "What should we do if we knew how to build FAI?", since people keep asking that question even though it's not a problem that currently should be focused on.)
> As a final point, I think it is very interesting that those who
> support friendly super-AI don’t see the AI coming to the
> conclusion that nearly all forms of government, particularly those
> of an authoritarian breed, are faulty and instead advocating
> anarchy or a form of hyper-limited government.
If you actually asked us what we thought the output of a CEV dynamic would be, you'd see we'd tend to guess it to be a rather "hyper-limited government". You are making weird assumptions about what we think.
> That the AI would want to govern at all is a further assumption
> I don’t understand.
I consider Nick Bostrom to have proven that due to evolutionary pressures, a total absence of a governing/controlling structure would lead to outcomes that we humans very much dislike:
http://www.nickbostrom.com/fut/evolution.html
That means some form of "hyper-limited goverment" is necessary (for a more precise formulation, see the previous link).
@Giulio:
Oh well - I guess they weren't REALLY engineers after all. ;p
Snark aside, I don't understand. Are you saying that I'm saying that no one should attempt to build FAI at all because... we definitely haven't built one yet?
Maybe science-fiction could be used to try some thought experiments. Iain M. Banks' Culture cycle is a very interesting way to develop philosophical and political reflections on the potential role of “intelligent” machines in an advanced society. On the Culture as a sort of “computer-aided” anarchy, see: http://yannickrumpala.wordpress.com/2010/01/14/anarchy_in_a_world_of_machines/
Let's assume my understanding of the CEV is faulty, because it probably is.
My argument was focused primarily on those saying that a technocratic authority based in the CEV would be better than previous authoritarian governments. If the CEV is as limited as you claim, then proponents of it would have no reason to defend authoritarian rule. If it isn't, there is a dubious more authority = more freedom line of thought being perpetrated.
I am new to the CEV and so may be making an error in response to pure CEV theory, but this post was specifically addressed to the FAI proponents critiquing Hughes' piece on democracy vs technocracy.
Kyle: "Democracy, communism, and, yes, AI-based technocratic authoritarianism, are all human systems.
But your interlocutors are quite specifically denying that AI behavior can be successfully predicted by analogy to humans and human organizations. You seem to dismiss this as special pleading, but you don't address the specific reasons we have for expecting powerful, non-anthropomorphic AI. That is: human behavior is the contingent product of our species's evolutionary history and cultural legacies. The entire human way of existing---where we have individual people, who have emotions and self-identities, and who think on our timescale and can't be copied---doesn't seem to be inherent in the nature of reality or intelligence itself . A generic artificial optimization process isn't going to share human characteristics except insofar as the designers specifically engineer those features in. This is by no means a guarantee that AIs will be "perfect"; it just means that whatever the failures of communism were, any AI failures are probably going to be different. FAI is an engineering problem. To the extent that functioning systems must "[see themselves] imperfect and prone to mistakes and failure," then competent AI designers should endeavor to write code that sees itself as imperfect and prone to mistakes and failure.
Kyle:
> My argument was focused primarily on those saying that a
> technocratic authority based in the CEV would be better than
> previous authoritarian governments. If the CEV is as limited as you
> claim, then proponents of it would have no reason to defend
> authoritarian rule.
Kaj Sotala etc. did not say that they'd support technocratic absolutism (they were quite explicit about this), or that implementing CEV would result in technocratic absolutism, even though they *did* point out that much of James's criticism of AI-authoritarianism was based on the mistaken assumption that all theoretically possible AIs would necessarily be similar to humans.
It is possible to point out mistakes in criticisms of AI-authoritarianism without being in favor of AI-authoritarianism. That is what happened here. (Though other things also happened, like pointing out that CEV isn't AI-authoritarianism even though some people who e.g. haven't read the page describing it think so.)
And even though I'm not in favor of AI-authoritarianism either, I hereby point out that you are mistaken in seeing special pleading where you see it:
Wikipedia describes "special pleading" as "someone attempting to cite something as an exemption to a generally accepted rule, principle, etc. without justifying the exemption".
The key is "without justifying the exemption". We probably all agree that sometimes there are exemptions to general rules that actually *are* justified.
So the question becomes, is it justified to claim that a hypothetical AI-dictator could be an exemption to e.g. "all dictators are self-interested beings" etc.
James Hughes claims that all AIs necessarily are self-interested beings in the way that humans are, while us others have been attempting to point out that the laws of physics do in fact allow other kinds of goal systems, and an AI can in principle have a goal system very different from those of humans. In soldier ants etc. we already see goal systems that aren't really self-centered.
@Nato Snark aside, I don't understand. Are you saying that I'm saying that no one should attempt to build FAI at all because... we definitely haven't built one yet?
Well, this is how I interpret "Friendly AI DOES NOT EXIST in the present tense. It makes no sense to argue about it until we have a candidate we can examine.". It is difficult to design something without arguing about it.
@Giulio: Fair enough. What I meant was that it makes no sense to argue about, say, whether to put an AI //in charge// until we have some more details to examine and test, not whether or not to design or build one. It's a little early in the development stage to say as much as whether a purported FAI is immune from certain cognitive biases or not.
I'm very glad this is just a case of me not being very clear.
@Nato: Agreed. I don't take FAI debates too seriously at this moment, because I don't think we know enough yet. Also, I assume if an AI is much smarter than us, it will quickly shed design constrains and choose its own goals (this is not computer science but logic: it is the definition of much smarter.
But I consider engineering AIs much smarter than us as doable, and desirable. So I am not very much interested in the friendliness issue, but I am very much interested in the ongoing AI research, and I am confident it will produce spectacular result sometime in the first half of this century.
Aleksei, you can say they were "explicit" about not supporting AI governance, but here are some quotes that refute you:
Anissimov: "give them [AI] positions of higher responsibility than most humans"
Mijic: "for agents with bounded intelligence who might be preyed upon by would-be dictators with oh-so-convincing arguments in favor of their totalitarian utopia, the only winning cognitive algorithm might be to just not accept any arguments, no matter how convincing they are. Hughes would rather tolerate a highly suboptimal outcome - life without an FAI - than give in to his innate sense of the wrongness of totalitarianism."
In both cases, the phrasing isn't "I support FAI authoritarianism" it is, "if it works this well, why wouldn't you support it?" That's why you, Aleksei can claim the CEV derived FAI is both authoritarian and somehow minarchist at the same time and never have to take ownership of the actual consequences of these thought experiments.
As for special pleading, here it is, explicitly:
Sotala: "This would be an understandable objection if we were talking about making an AI that ran things the way *it* liked, with little regard to what humans wanted to. But we are expressly talking about an AI that wants nothing else than what humans do."
Sotala and Mijic in particular (Anissimov much less so) use precisely the same argument as Communists - that the "system" gives you what you really want, even if you don't know it, it works that well - but despite the enormous similarities, is immune to Communist critique because FAI is *different* and critics just *don't understand* and in fact, as Sotala puts it, shouldn't even be allowed to critique FAI unless they are experts in the field, and clearly someone isn't an expert if they don't support FAI.
So no, Aleksei, I am not mistaken. Furthermore, I'm not going to play this little game of rhetoric where CEV and FAI supporters openly critique Hughes advocacy of democracy and then turn around and say "oh, well, we don't advocate absolutism, just, you know, pointing out that it's not so bad." That's an extremely disingenuous form of argumentation and doesn't move the debate forward.
I have yet to see a clear proposal as to how the CEV/FAI would be integrated into a modern system of government. Until I see one, I am going to presume that CEV/FAI advocates support some form of technocratic authoritarian rule because that is the form of government they (Mijic and Sotala) have been defending.
> Aleksei, you can say they were "explicit" about not supporting
> AI governance, but here are some quotes that refute you:
But that's not at all what I said. I said they didn't support "AI-authoritarianism" or "technocratic absolutism". Supporting "AI governance" doesn't mean those things any more than our current political system -- which is "human governance" -- can be described as "technocratic absolutism" or "human-authoritarianism".
You probably understand that "human governance" doesn't automatically mean any of these criticized political systems (instead, it can mean any system that has ever existed in history), but for some reason you assume that "AI governance" does.
> In both cases, the phrasing isn't "I support FAI authoritarianism"
> it is, "if it works this well, why wouldn't you support it?"
Do you understand that asking such a question doesn't mean that one supports the political system under discussion? Even bad political systems can be criticized on faulty or insufficient grounds, and one thing that Sotala did was wonder why James Hughes didn't really manage to criticize even the bad systems very well (in Sotala's view).
> That's why you, Aleksei can claim the CEV derived FAI is both
> authoritarian and somehow minarchist at the same time and
> never have to take ownership of the actual consequences of
> these thought experiments.
Where have I claimed that "CEV derived FAI" would be authoritarian?
And actually I have not claimed it to be minarchist either, only that my *guess* is that it would be. I actually *don't know* what the final political system would be, since finding that out is the whole point of doing the extrapolation of human thinking that CEV is supposed to be built to do. If we already knew what the final political system should be, we wouldn't want to build this complex polling mechanism known as CEV.
> as Sotala puts it, shouldn't even be allowed to critique FAI unless
> they are experts in the field, and clearly someone isn't an expert
> if they don't support FAI.
It's really amazing that you're able to fantasize that Sotala would think something like that. You need to try to calm down, clear your head and try to actually read what people are saying.
I'd also guess that you still haven't read the page describing CEV, even though you "criticize" it so much.
Kyle: "Furthermore, I'm not going to play this little game of rhetoric [...] That's an extremely disingenuous form of argumentation and doesn't move the debate forward."
I don't think anyone's being disingenuous; communication is truly difficult, especially when trying to talk about topics such as superintelligence, where our common language doesn't seem to have adequate concepts. And it cannot be stressed enough that FAI proponents are expecting a genuine superintelligence, an entity much, much smarter than humans and human organizations. So when we say that AIs are going to be in control and not us, it's not out of a love for absolutist government; it's based on the theoretical prediction that larger, smarter, faster minds have a greater share in determining the future. CEV is a rough sketch of how one might go about building an AI that does what we want, because the alternative to an AI that does what we want is either an AI that does things we don't want, or no AI at all.
Of course, all this is assuming that superintelligence with arbitrarily programmable values is feasible; if that's not true, then arguments for FAI obviously fail. If you believe that superintelligence is impossible, or that superintelligences necessarily have some particular sort of values, then it's better to move the discussion back to that point. Until can we agree on some general features of what AI would be like, any discussion of the interaction of AI and politics is doomed to fail.
"I have yet to see a clear proposal as to how the CEV/FAI would be integrated into a modern system of government."
If we must speak in analogies, it's probably better to think of the rise of a new intelligent species, rather than an absolutist government. What would it even mean, to integrate human intelligence into a system of chimpanzee tribal governance? Or what would it mean to try to give dogs an equal say in human affairs? We have concepts and problems that dogs and chimpanzees simply aren't equipped to comprehend. So it would be with superintelligences and us. We all want a good outcome, the FAI proponents are just saying that securing good outcomes is a mind design problem, rather than a political problem.
Zack M. Davis:
> We all want a good outcome, the FAI proponents are just saying
> that securing good outcomes is a mind design problem, rather
> than a political problem.
Well, it's *also* a political problem, since the mind design choices are also political choices. It's just that practically any political viewpoint can in principle be well-represented by making appropriate choices during mind design (the exception being if you just hate all AIs, even transitory self-deleting ones). Like representing liberal democrary by building a superintelligence that polls all humans in an elaborate way and implements whatever sufficiently common ground can be found in what humans want (which very well might include deleting the AI after it fixes a few things in the world).
(Yes, you probably actually meant this, that mind design is political too, but people here tend to jump at you if there's an opportunity for an unfavorable interpretation.)
(One of these unfavorable interpretations that people will probably make from this message is that I'd claim that we'd actually be able to build such a superintelligence as described -- I do claim it to be theoretically possible, but that does not mean practical possibility.)
Kyle, when you are rebutting someone in the future, I would suggest using specific quotes, rather than paraphrasing. I think it would help clarify the disagreement. Thanks!
Kyle, if I wasn't using special pleading, i.e., if you don't have quotes to back it up, then please remove my name from the list of names in this post.
Michael, Hughes already quoted you directly, but here is where you engage in special pleading:
"Power corrupts humans for evolutionary reasons—if one is on top of the heap, one had better take advantage of the opportunity to reward one’s allies and punish one’s enemies. This is pure evolutionary logic and need not be consciously calculated. AIs, which can be constructed entirely without selfish motivations, can be immune to these tendencies."
IEET Blog |
email list |
newsletter |
The IEET is a 501(c)3 non-profit, tax-exempt organization registered in the State of Connecticut in the United States.
Contact: Executive Director, Dr. James J. Hughes,
Williams 229B, Trinity College, 300 Summit St., Hartford CT
06106 USA
Email: director @ ieet.org phone:
860-297-2376
Some thinkers, including the allegedly brilliant philosopher Slavoj Zizek, continue to defend Marxism using special pleading. Instead of claiming communism isn’t based in humans, they claim Stalin and the USSR were not pure communism, and therefore were doomed to failure because of the corrupting element of capitalism. Thus, thanks to special pleading, Stalin is not proof that communism and authoritarianism are dangerous and bad, but that capitalism is bad and corrupts the pure motives of communism.
IEET Blog