In this article I present a critique of Nick Bostrom’s book Superintelligence. For purposes of brevity I shall not devote much space to summarising Bostrom’s arguments or defining all the terms that he uses. Though I briefly review each key idea before discussing it, I shall also assume that readers have some general idea of Bostrom’s argument, and some of the key terms involved. Also note that to keep this piece focused, I only discuss arguments raised in this book, and not what Bostrom has written elsewhere or others who have addressed similar issues. The structure of this article is as follows. I first offer a summary of what I regard to be the core argument of Bostrom’s book, outlining a series of premises that he defends in various chapters. Following this summary, I commence a general discussion and critique of Bostrom’s concept of ‘intelligence’, arguing that his failure to adopt a single, consistent usage of this concept in his book fatally undermines his core argument. The remaining sections of this article then draw upon this discussion of the concept of intelligence in responding to each of the key premises of Bostrom’s argument. I conclude with a summary of the strengths and weaknesses of Bostrom’s argument.
Summary of Bostrom’s Argument
Throughout much of his book, Bostrom remains quite vague as to exactly what argument he is making, or indeed whether he is making a specific argument at all. In many chapters he presents what are essentially lists of various concepts, categories, or considerations, and then articulates some thoughts about them. Exactly what conclusion we are supposed to draw from his discussion is often not made explicit. Nevertheless, by my reading the book does at least implicitly present a very clear argument, which bears a strong similarity to the sorts of arguments commonly found in the Effective Altruism (EA) movement, in favour of focusing on AI research as a cause area. In order to provide structure for my review, I have therefore constructed an explicit formulation of what I take to be Bostrom’s main argument in his book. I summarise it as follows:
Premise 1: A superintelligence, defined as a system that ‘exceeds the cognitive performance of humans in virtually all domains of interest’, is likely to be developed in the foreseeable future (decades to centuries).
Premise 2: If superintelligence is developed, some superintelligent agent is likely to acquire a decisive strategic advantage, meaning that no terrestrial power or powers would be able to prevent it doing as it pleased.
Premise 3: A superintelligence with a decisive strategic advantage would be likely to capture all or most of the cosmic endowment (the total space and resources within the accessible universe), and put it to use for its own purposes.
Premise 4: A superintelligence which captures the cosmic endowment would likely put this endowment to uses incongruent with our (human) values and desires.
Preliminary conclusion: In the foreseeable future it is likely that a superintelligent agent will be created which will capture the cosmic endowment and put it to uses incongruent with our values. (I call this the AI Doom Scenario).
Premise 5: Pursuit of work on AI safety has a non-trivial chance of noticeably reducing the probability of the AI Doom Scenario occurring.
Premise 6: If pursuit of work on AI safety has at least a non-trivial chance of noticeably reducing the probability of an AI Doom Scenario, then (given the preliminary conclusion above) the expected value of such work is exceptionally high.
Premise 7: It is morally best for the EA community to preferentially direct a large fraction of its marginal resources (including money and talent) to the cause area with highest expected value.
Main conclusion: It is morally best for the EA community to direct a large fraction of its marginal resources to work on AI safety. (I call this the AI Safety Thesis.)
Bostrom discusses the first premise in chapters 1-2, the second premise in chapters 3-6, the third premise in chapters 6-7, the fourth premise in chapters 8-9, and some aspects of the fifth premise in chapters 13-14. The sixth and seventh premises are not really discussed in the book (though some aspects of them are hinted at in chapter 15), but are widely discussed in the EA community and serve as the link between the abstract argumentation and real-world action, and as such I decided also to discuss them here for completeness. Many of these premises could be articulated slightly differently, and perhaps Bostrom would prefer to rephrase them in various ways. Nevertheless I hope that they at least adequately capture the general thrust and key contours of Bostrom’s argument, as well as how it is typically appealed to and articulated within the EA community.
The nature of intelligence
In my view, the biggest problem with Bostrom’s argument in Superintelligence is his failure to devote any substantial space to discussing the nature or definition of intelligence. Indeed, throughout the book I believe Bostrom uses three quite different conceptions of intelligence:
- Intelligence(1): Intelligence as being able to perform most or all of the cognitive tasks that humans can perform. (See page 22)
- Intelligence(2): Intelligence as a measurable quantity along a single dimension, which represents some sort of general cognitive efficaciousness. (See pages 70,76)
- Intelligence(3): Intelligence as skill at prediction, planning, and means-ends reasoning in general. (See page 107)
While certainly not entirely unrelated, these three conceptions are all quite different from each other. Intelligence(1) is mostly naturally viewed as a multidimensional construct, since humans exhibit a wide range of cognitive abilities and it is by no means clear that they are all reducible to a single underlying phenomenon that can be meaningfully quantified with one number. It seems much more plausible to say that the range of human cognitive abilities require many different skills which are sometimes mutually-supportive, sometimes mostly unrelated, and sometimes mutually-inhibitory in varying ways and to varying degrees. This first conception of intelligence is also explicitly anthropocentric, unlike the other two conceptions which make no reference to human abilities. Intelligence(2) is unidimensional and quantitative, and also extremely abstract, in that it does not refer directly to any particular skills or abilities. It most closely parallels the notion of IQ or other similar operational measures of human intelligence (which Bostrom even mentions in his discussion), in that it is explicitly quantitative and attempts to reduce abstract reasoning abilities to a number along a single dimension. Intelligence(3) is much more specific and grounded than either of the other two, relating only to particular types of abilities. That said, it is not obviously subject to simple quantification along a single dimension as is the case for Intelligence(2), nor is it clear that skill at prediction and planning is what is measured by the quantitative concept of Intelligence(2). Certainly Intelligence(3) and Intelligence(2) cannot be equivalent if Intelligence(2) is even somewhat analogous to IQ, since IQ mostly measures skills at mathematical, spatial, and verbal memory and reasoning, which are quite different from skills at prediction and planning (consider for example the phenomenon of autistic savants). Intelligence(3) is also far more narrow in scope than Intelligence(1), corresponding to only one of the many human cognitive abilities.
Repeatedly throughout the book, Bostrom flips between using one or another of these conceptions of intelligence. This is a major weakness for Bostrom’s overall argument, since in order for the argument to be sound it is necessary for a single conception of intelligence to be adopted and apply in all of his premises. In the following paragraphs I outline several of the clearest examples of how Bostrom’s equivocation in the meaning of ‘intelligence’ undermines his argument.
Bostrom argues that once a machine becomes more intelligent than a human, it would far exceed human-level intelligence very rapidly, because one human cognitive ability is that of building and improving AIs, and so any superintelligence would also be better at this task than humans. This means that the superintelligence would be able to improve its own intelligence, thereby further improving its own ability to improve its own intelligence, and so on, the end result being a process of exponentially increasing recursive self-improvement. Although compelling on the surface, this argument relies on switching between the concepts of Intelligence(1) and Intelligence(2). When Bostrom argues that a superintelligence would necessarily be better at improving AIs than humans because AI-building is a cognitive ability, he is appealing to Intelligence(1). However, when he argues that this would result in recursive self-improvement leading to exponential growth in intelligence, he is appealing to Intelligence(2). To see how these two arguments rest on different conceptions of intelligence, note that considering Intelligence(1), it is not at all clear that there is any general, single way to increase this form of intelligence, as Intelligence(1) incorporates a wide range of disparate skills and abilities that may be quite independent of each other. As such, even a superintelligence that was better than humans at improving AIs would not necessarily be able to engage in rapidly recursive self-improvement of Intelligence(1), because there may well be no such thing as a single variable or quantity called ‘intelligence’ that is directly associated with AI-improving ability. Rather, there may be a host of associated but distinct abilities and capabilities that each needs to be enhanced and adapted in the right way (and in the right relative balance) in order to get better at designing AIs. Only by assuming a unidimensional quantitative conception of Intelligence(2) does it make sense to talk about the rate of improvement of a superintelligence being proportional to its current level of intelligence, which then leads to exponential growth. Bostrom therefore faces a dilemma. If intelligence is a mix of a wide range of distinct abilities as in Intelligence(1), there is no reason to think it can be ‘increased’ in the rapidly self-reinforcing way Bostrom speaks about (in mathematical terms, there is no single variable which we can differentiate and plug into the differential equation, as Bostrom does in his example on pages 75-76). On the other hand, if intelligence is a unidimensional quantitative measure of general cognitive efficaciousness, it may be meaningful to speak of self-reinforcing exponential growth, but it is not necessarily obvious that any arbitrary intelligent system or agent would be particularly good at designing AIs. Intelligence(2) may well help with this ability, but it’s not at all clear it is sufficient – after all, we readily conceive of building a highly “intelligent” machine that can reason abstractly and pass IQ tests etc, but is useless at building better AIs.
Bostrom argues that once a machine intelligence became more intelligent than humans, it would soon be able to develop a series of ‘cognitive superpowers’ (intelligence amplification, strategising, social manipulation, hacking, technology research, and economic productivity), which would then enable it to escape whatever constraints were placed upon it and likely achieve a decisive strategic advantage. The problem is that it is unclear whether a machine endowed only with Intelligence(3) (skill at prediction and means-ends reasoning) would necessarily be able to develop skills as diverse as general scientific research ability, the capability to competently use natural language, and perform social manipulation of human beings. Again, means-ends reasoning may help with these skills, but clearly they require much more beyond this. Only if we are assuming the conception of Intelligence(1), whereby the AI has already exceeded essentially all human cognitive abilities, does it become reasonable to assume that all of these ‘superpowers’ would be attainable.
According to the orthogonality thesis, there is no reason why the machine intelligence could not have extremely reductionist goals such as maximising the number of paperclips in the universe, since an AI’s level of intelligence is totally separate to and distinct from its final goals. Bostrom’s argument for this thesis, however, clearly depends adopting Intelligence(3), whereby intelligence is regarded as general skill with prediction and means-ends reasoning. It is indeed plausible that an agent endowed only with this form of intelligence would not necessarily have the ability or inclination to question or modify its goals, even if they are extremely reductionist or what any human would regard as patently absurd. If, however, we adopt the much more expansive conception of Intelligence(1), the argument becomes much less defensible. This should become clear if one considers that ‘essentially all human cognitive abilities’ includes such activities as pondering moral dilemmas, reflecting on the meaning of life, analysing and producing sophisticated literature, formulating arguments about what constitutes a ‘good life’, interpreting and writing poetry, forming social connections with others, and critically introspecting upon one’s own goals and desires. To me it seems extraordinarily unlikely that any agent capable of performing all these tasks with a high degree of proficiency would simultaneously stand firm in its conviction that the only goal it had reasons to pursue was tilling the universe with paperclips. As such, Bostrom is driven by his cognitive superpowers argument to adopt the broad notion of intelligence seen in Intelligence(1), but then is driven back to a much narrower Intelligence(3) when he wishes to defend the orthogonality thesis. The key point to be made here is that the goals or preferences of a rational agent are subject to rational reflection and reconsideration, and the exercise of reason in turn is shaped by the agent’s preferences and goals. Short of radically redefining what we mean by ‘intelligence’ and ‘motivation’, this complex interaction will always hamper simplistic attempts to neatly separate them, thereby undermining Bostrom’s case for the orthogonality thesis – unless a very narrow conception of intelligence is adopted.
In the table below I summarise several of the key outcomes or developments that are critical to Bostrom’s argument, and how plausible they would be under each of the three conceptions of intelligence. Obviously such judgements are necessarily vague and subjective, but the key point I wish to make is simply that only by appealing to different conceptions of intelligence in different cases is Bostrom able to argue that all of the outcomes are reasonably likely to occur. Fatally for his argument, there is no single conception of intelligence that makes all of these outcomes simultaneously likely or plausible.
||Intelligence(1): all human cognitive abilities
||Intelligence(2): unidimensional measure of cognition
||Intelligence(3): prediction and means-ends reasoning
|Develops all cognitive superpowers
|Absurd ‘paperclip maximising’ goals
|Resists changes to goals
|Can escape confinement
Premise 1: Superintelligence is coming soon
I have very little to say about this premise, since I am in broad agreement with Bostrom that even if it takes decades or a century, super-human artificial intelligence is quite likely to be developed. I find Bostrom’s appeals to surveys of AI researchers regarding how long it is likely to be until human level AI is developed fairly unpersuasive, given both the poor track record of such predictions and also the fact that experts on AI research are not necessarily experts on extrapolating the rate of technological and scientific progress (even in their own field). Bostrom, however, does note some of these limitations, and I do not think his argument is particularly dependent upon these sorts of appeals. I therefore will pass over premise 1 and move on to what I consider to be the more important issues.
Premise 2: Arguments against a fast takeoff
Bostrom’s major argument in favour of the contention that a superintelligence would be able to gain a decisive strategic advantage is that the ‘takeoff’ for such an intelligence would likely be very rapid. By a ‘fast takeoff’, Bostrom means that the time between when the superintelligence first approaches human-level cognition and when it achieves dramatically superhuman intelligence would be small, on the order of days or even hours. This is critical because if takeoff is as rapid as this, there will be effectively no time for any existing technologies or institutions to impede the growth of the superintelligence or check it in any meaningful way. Its rate of development would be so rapid that it would readily be able to out-think and out-maneuver all possible obstacles, and rapidly obtain a decisive strategic advantage. Once in this position, the superintelligence would possess an overwhelming advantage in technology and resources, and would therefore be effectively impossible to displace.
The main problem with all of Bostrom’s arguments for the plausibility of a fast takeoff is that they are fundamentally circular, in that the scenario or consideration they propose is only plausible or relevant under the assumption that the takeoff (or some key aspect of it) is fast. The arguments he presents are as follows:
- Two subsystems argument: if an AI consists of two or more subsystems with one improving rapidly, but only contributing to the ability of the overall system after a certain threshold is reached, then the rate of increase in the performance of the overall system could drastically increase once that initial threshold is passed. This argument assumes what it is trying to prove, namely that the rate of progress in a critical rate-limiting subsystem could be very rapid, experiencing substantial gains on the order of days or even hours. It is hard to see what Bostrom’s scenario really adds here; all he has done is redescribed the fast takeoff scenario in a slightly more specific way. He has not given any reason for thinking that it is at all probable that progress on such a critical rate-limiting subsystem would occur at the extremely rapid pace characteristic of a fast takeoff.
- Intelligence spectrum argument: Bostrom argues that the intelligence gap between ‘infra-idiot’ and ‘ultra-Einstein’, while appearing very large to us, may actually be quite small in the overall scheme of the spectrum of possible levels of intelligence, and as such the time taken to improve an AI through and beyond this level may be much less than it originally seems. However, even if it is the case that the range of the intelligence spectrum within which all humans fall is fairly narrow in the grand scheme of things, it does not follow that the time taken to traverse it in terms of AI development is likely be on the order of days or weeks. Bostrom is simply making an assumption that such rapid rates of progress could occur. His intelligence spectrum argument can only ever show that the relative distance in intelligence space is small; it is silent with respect to likely development timespans.
- Content overhang argument: an artificial intelligence could be developed with high capabilities but with little raw data or content to work with. If large quantities of raw data could be processed quickly, such an AI could rapidly expand its capabilities. The problem with this argument is that what is most important is not how long it takes a given AI to absorb some quantity of data, but rather the length of time between producing one version of the AI and the next, more capable version. This is because the key problem is that we currently don’t know how to build a superintelligence. Bostrom is arguing that if we did build a nascent superintelligence that simply needed to process lots of data to manifest its capabilities, then this learning phase could occur quickly. He gives no reason, however, to think that the rate at which we can learn how to build that nascent superintelligence (in other words, the overall rate of progress in AI research) will be anything like as fast as the rate an existing nascent superintelligence would be able to process data. Only if we assume rapid breakthroughs in AI design itself does the ability of AIs to rapidly assimilate large quantities of data become relevant.
- Hardware overhang argument: it may be possible to increase the capabilities of a nascent superintelligence dramatically and very quickly by rapidly increasing the scale and performance of the hardware it had access to. While theoretically possible, this is an implausible scenario since any artificial intelligence showing promise would likely be operating near the peak of plausible hardware provision. This means that testing, parameter optimisation, and other such tasks will take considerable time, as hardware will be a limiting factor. Bostrom’s concept of a ‘hardware overhang’ amounts to thinking that AI researchers would be content to ‘leave money on the table’, in the sense of not making use of what hardware resources are available to them for extended periods of development. This is especially implausible in the event of groundbreaking research involving AI architectures showing substantial promise. Such systems would hardly be likely to spend years being developed on relatively primitive hardware only to be suddenly and very rapidly dramatically scaled up at the precise moment when practically no further development is necessary, and they are already effectively ready to achieve superhuman intelligence.
- ‘One key insight’ argument: Bostrom argues that ‘if human level AI is delayed because one key insight long eludes programmers, then when the final breakthrough occurs, the AI might leapfrog from below to radically above human level’. Assuming that ‘one key insight’ would be all it would take to crack the problem of superhuman intelligence is, to my mind, grossly implausible, and not consistent either with the slow but steady rate of progress in artificial intelligence research over the past 60 years, or with the immensely complex and multifaceted phenomenon that is human intelligence.
Additional positive arguments against the plausibility of a fast takeoff include the following:
- Speed of science: Bostrom’s assertion that artificial intelligence research could develop from clearly sub-human to obviously super-human levels of intelligence in a matter of days or hours is simply absurd. Scientific and engineering projects simply do not work over timescales that short. Perhaps to some degree this could be altered in the future if (for example) human-level intelligence could be emulated on a computer and then the simulation run at much faster than real-time. But Bostrom’s argument is that machine intelligence is likely to precede emulation, and as such all we will have to work with at least up to the point of human/machine parity being reached is human levels of cognitive ability. As such it seems patently absurd to argue that developments of this magnitude could be made on the timespan of days or weeks. We simply see no examples of anything like this from history, and Bostrom cannot argue that the existence of superintelligence would make historical parallels irrelevant, since we are precisely talking about the development of superintelligence in the context of it not already being in existence.
- Subsystems argument: any superintelligent agent will doubtlessly require many interacting and interconnected subsystems specialised for different tasks. This is the way even much narrower AIs work, and it is certainly how human cognition works. Ensuring that all these subsystems or processes interact efficiently, without one inappropriately dominating or slowing up overall cognition, or without bottlenecks of information transfer or decision making, is likely to be something that requires a great deal of experimentation and trial-and-error. This in turn will take extensive empirical experiments, tinkering, and much clever work. All this takes time.
- Parallelisation problems: many algorithms cannot be sped up considerably by simply adding more computational power unless an efficient way can be found to parallelise them, meaning that they can be broken down into smaller steps which can be performed in parallel across many processors at once. This is much easier to do for some types of algorithms and computations than others. It is not at all clear that the key algorithms used by a superintelligence would be susceptible to parallelisation. Even if they were, developing efficient parallelised forms of the relevant algorithms would itself be a prolonged process. The superintelligence itself would only be able to help in this development to the degree permitted by its initially limited hardware endowment. We therefore would expect to observe gradual improvement of algorithmic efficiency in parallelisation, thereby enabling more hardware to be added, thereby enabling further refinements to the algorithms used, and so on. It is therefore not at all clear that a superintelligence could be rapidly augmented simply by ‘adding more hardware’.
- Need for experimentation: even if a superintelligence came into existence quite rapidly, it would still not be able to achieve a decisive strategic advantage in similarly short time. This is because such an advantage would almost certainly require development of new technologies (at least the examples Bostrom gives almost invariably involve the AI using technologies currently unavailable to humans), which would in turn require scientific research. Scientific research is a complex activity that requires far more than skill at ‘prediction and means-end reasoning’. In particular, it also generally requires experimental research and (if engineering of new products is involved) producing and testing of prototypes. All of this will take time, and crucially is not susceptible to computational speedup, since the experiments would need to be performed with real physical systems (mechanical, biological, chemical, or even social). The idea that all (or even most) such testing and experimentation could be replaced by computer simulation of the relevant system is absurd, since most such simulations are completely computationally intractable, and likely to remain so for the foreseeable future (in many cases possibly forever). Therefore in the development of new technologies and scientific knowledge, the superintelligence is still fundamentally limited by the rate at which real-world tests and experiments can be performed.
- The infrastructure problem: in addition to the issue of developing new technologies, there is the further problem of the infrastructure required to develop such technologies, or even just to carry out the core objectives of the superintelligence. In order to acquire a decisive strategic advantage, a superintelligence will require vast computational resources, energy sources to supply them, real-world maintenance of these facilities, sources of raw materials, and vast manufacturing centres to produce any physical manipulators or other devices it requires. If it needs humans to perform various tasks for it, it will likely also require training facilities and programs for its employees, as well as teams of lawyers to acquire all the needed permits and permissions, write up contracts, and lobby governments. All of this physical and social infrastructure cannot be built in the matter of an afternoon, and more realistically would take many years or even decades to put in place. No amount of superintelligence can overcome physical limitations of the time required to produce and transform large quantities of matter and energy into desired forms. One might argue that improved technology certainly can reduce the time taken to move matter and energy, but the point is that it can only do so after the technology has been embodied in physical forms. The superintelligence would not have access to such hypothetical super-advanced transportation, computation, or construction technologies until it had built the factories needed to produce the machine tools with are needed to precisely refine the raw materials needed for parts in the construction of the nanofactory… and so on for many other similar examples. Nor can even vast amounts of money and intelligence allow any agent to simply brush aside the impediments of the legal system and government bureaucracy in an afternoon. A superintelligence would not simply be able to ignore such social restrictions on its actions until after it had gained enough power to act in defiance of world governments, which it would not be able to do until it had already acquired considerable military capabilities. All of this would take considerable time, precluding a fast takeoff.
Premise 3: Arguments against cosmic expansion
Critical to Bostrom’s argument about the dangers of superintelligence is that a superintelligence with a critical strategic advantage would likely capture the majority of the cosmic endowment (the sum total of the resources available within the regions of space potentially accessible to humans). This is why Bostrom presents calculations for the huge numbers of potential human lives (or at least simulations of lives) whose happiness is at stake should the cosmic endowment be captured by a rogue AI. While Bostrom does present some compelling reasons for thinking that a superintelligence with a decisive strategic advantage would have reasons and the ability to expand throughout the universe, there are also powerful considerations against the plausibility of this outcome which he fails to consider.
First, by the orthogonality thesis, a superintelligent agent could have almost any imaginable goal. It follows that a wide range of goals are possible that are inconsistent with cosmic expansion. In particular, any superintelligence with goals involving the value of unspoiled nature, or of constraining its activities to the region of the solar system, or of economising on the use of resources, would have reasons not to pursue cosmic expansion. How likely it is that a superintelligence would be produced with such self-limiting goals compared to goals favouring limitless expansion is unclear, but it is certainly a relevant outcome to consider, especially given that valuing exclusively local outcomes or conservation of resources seem like plausible goals that might be incorporated by developers into a seed AI.
Second, on a number of occasions, Bostrom briefly mentions that a superintelligence would only be able to capture the entire cosmic endowment if no other technologically advanced civilizations, or artificial intelligences produced by such civilizations, existed to impede it. Nowhere, however, does he devote any serious consideration to how likely the existence of such civilizations or intelligences is. Given the great age and immense size of the cosmos, however, the probability that humans are the first technological civilization to achieve spaceflight, or that any superintelligence we produce would be the first to spread throughout the universe, seems infinitesimally small. Of course this is an area of great uncertainly and we can therefore only speculate about the relevant probabilities. Nevertheless, it seems very plausible to me that the chances of any human-produced superintelligence successfully capturing the cosmic endowment without alien competition are very low. Of course this does not mean that an out-of-control terrestrial AI could not do great harm to life on Earth and even spread throughout neighbouring stars, but it does significantly blunt the force of the huge numbers Bostrom presents as being at stake if we think the entire cosmic endowment is at risk of being misused.
Premise 4: The nature of AI motivation
Bostrom’s main argument in defence of premise 4 is that unless we are extremely careful and/or lucky in establishing the goals and motivations of the superintelligence before it captures the cosmic endowment, it is likely to end up pursuing goals that are not in alignment with our own values. Bostrom presents a number of thought experiments as illustrations of the difficulty of specifying values or goals in a manner that would result in the sorts of behaviours we want it to perform. Most of these examples involve the superintelligence pursuing a goal in a single-minded, literalistic way, which no human being would regard as ‘sensible’. He gives as examples an AI tasked with maximising its output of paperclips sending out probes to harvest all the energy within the universe to make more paperclips, or an AI tasked with increasing human happiness enslaving all humans and hijacking their brains to stimulate the pleasure centres directly. One major problem I have with all such examples is that the AIs always seem to lack a critical ability in interpreting and pursuing their goals that, for want of a better term, we might describe as ‘common sense’. This issue ultimately reduces to which conception of intelligence one applies, since if we adopt Intelligence(1) then any such AIs would necessarily have ‘common sense’ (this being a human cognitive ability), while the other two conceptions of intelligence would not necessarily include this ability. However, if we do take Intelligence(1) as our standard, then it seems difficult to see why a superintelligence would lack the sort of common sense by which any human would be able to see that the simple-minded, literalistic interpretations given as examples by Bostrom are patently absurd and ridiculous things to do.
Aside from the question of ‘common sense’, it is also necessary to analyse the concept of ‘motivation’, which is a multifaceted notion that can be understood in a variety of ways. Two particularly important conceptions of motivation are that it is some sort of internal drive to do or obtain some outcome, and motivation as some sort of more abstract rational consideration by which an agent has a reason to act in a certain way. Given what he says about the orthogonality thesis, it seems that Bostrom thinks of motivation as being some sort of internal drive to act in a particular way. In the first few pages of the chapter on the intelligent will, however, he switches from talking about motivation to talking about goals, without any discussion about the relationship between these two concepts. Indeed, it seems that these are quite different things, and can exist independently of each other. For example, humans can have goals (to quit smoking, or to exercise more) without necessarily having any motivation to take actions to achieve those goals. Conversely, humans can be motivated to do something without having any obvious associated goal. Many instances of collective behaviour in crowds and riots may be examples of this, where people act based on situational factors without any clear reason or objectives. Human drives such as curiously and novelty can also be highly motivating without necessarily having any particular goal associated with them. Given the plausibility that motivation and goals are different and distinct concepts, it is important for Bostrom to explain what he thinks the relationship between them is, and how they would operate in an artificial agent. This seems all the more relevant since we would readily say that many intelligent artificial systems possess goals (such as the common examples of a heat-seeking missile or a chess playing program), but it is not at all clear that these systems are in any way ‘motivated’ to perform these actions – they are simply designed to work towards these goals, and motivations simply don’t come into it. What then would it take to build an artificial agent that had both goals and motivations? How would an artificial agent act with respect to these goals and/or motivations? Bostrom simply cannot ignore these questions if he is to provide a compelling argument concerning what AIs would be motivated to do.
The problems inherent in Bostrom’s failure to analyse these concepts in sufficient detail become evident in the context of Bostrom’s discussion of something that he calls ‘final goals’. While he does not define these, presumably he means goals that are not pursued in order to achieve some further goal, but simply for their own sake. This raises several additional questions: can an agent have more than one final goal? Need they have any final goals at all? Might goals always be infinitely resolvable in terms of fulfilling some more fundamental or more abstract underlying goal? Or might multiple goals form an inter-connected self-sustaining network, such that all support each other but no single goal can be considered most fundamental or final? These questions might seem arcane, but addressing them is crucial for conducting a thorough and useful analysis of the likely behaviour of intelligent agents. Bostrom often speaks as if a superintelligence will necessarily act in single-minded devotion to achieve its one final goal. This assumes, however, that a superintelligence would be motivated to achieve its goal, that it would have one and only one final goal, and that its goal and its motivation to achieve it are totally independent from and not receptive to rational reflection or any other considerations. As I have argued here and previously, however, these are all quite problematic and dubious notions. In particular, as I noted in the discussion about the nature of intelligence, a human’s goals are subject to rational reflection and critique, and can be altered or rejected if they are determined to be irrational or incongruent with other goals, preferences, or knowledge that the person has. It therefore seems highly implausible that a superintelligence would hold so tenaciously to their goals, and pursue them so single-mindedly. Only a superintelligence possessing a much more minimal form of intelligence, such as the skills at prediction and means-ends reasoning of Intelligence(3), would be a plausible candidate for acting in such a myopic and mindless way. Yet as I argued previously, a superintelligence possessing only this much more limited form of intelligence would not be able to acquire all of the ‘cognitive superpowers’ necessary to establish a decisive strategic advantage.
Bostrom would likely contend that such reasoning is anthropomorphising, applying human experiences and examples in cases where they simply do not apply, given how different AIs could be to human beings. Yet how can we avoid anthropomorphising when we are using words like ‘motivation’, ‘goal’, and ‘will’, which acquire their meaning and usage largely through application to humans or other animals (as well as anthropomorphised supernatural agents)? If we insist on using human-centred concepts in our analysis, drawing anthropocentric analogies in our reasoning is unavoidable. This places Bostrom in a dilemma, as he wants to simultaneously affirm that AIs would possess motivations and goals, but also somehow shear these concepts of their anthropocentric basis, saying that they could work totally differently to how these concepts are applied in humans and other known agents. If these concepts work totally differently, then how are we justified in even using the same words in the two different cases? It seems that if this were so, Bostrom would need to stop using words like ‘goal’ and ‘motivation’ and instead start using some entirely different concept that would apply to artificial agents. On the other hand if these concepts work sufficiently similarly in human and AI cases to justify using common words to describe both cases, then there seems nothing obviously inappropriate in appealing to the operation of goals in humans in order to understand how they would operate in artificial agents. Perhaps one might contend that we do not really know whether artificial agents would have human analogues of desires and goals, or whether they would have something distinctively different. If this is the case, however, then our level of ignorance is even more profound than we had realised (since we don’t even know what words we can use to talk about the issue), and therefore much of Bostrom’s argument on these subjects would be grossly premature and under-theorised.
Bostrom also argues that once a superintelligence comes into being, it would resist any changes to its goals, since its current goals are (nearly always) better achieved by refraining from changing them to some other goal. There is an obvious flaw to this argument, namely that humans change their goals all the time, and indeed whole subdisciplines of philosophy are dedicated to pursuing the question of what we should value and how we should go about modifying our goals or pursuing different things to what we currently do. Humans can even change their ‘final goals’ (insomuch as any such things exist), such as when they convert religions or change between radically opposed political ideologies. Bostrom mentions this briefly but does not present any particularly convincing explanation for this phenomenon, nor does he explain why we should assume that this clear willingness to countenance (and even pursue) goal changes is not something that would affect AIs as it affects humans. One potential such response could be that the ‘final goal’ pursued by all humans is really something very basic such as ‘happiness’ or ‘wellbeing’ or ‘pleasure’, and that this never changes even though the means of achieving it can vary dramatically. I am not convinced by this analysis, since many people (religious and political ideologues being obvious example) seem motivated by causes to perform actions that cannot readily be regarded as contributing to their own happiness or wellbeing, unless these concepts are stretched to become implausibly broad. Even if we accept that people always act to promote their own happiness or wellbeing, however, it is certainly the case that they can dramatically change their beliefs about what sort of things will improve their happiness or wellbeing, thus effectively changing their goals. It is unclear to me why we should expect that a superintelligence able to reflect upon its goals could not similarly change its mind about the meaning of its goals, or dramatically alter its views on how to best achieve them.
Premise 5: The tractability of the AI alignment problem
Critical to the question of artificial intelligence research as a cause for effective altruists is the argument that there are things which can be done in the present to reduce the risk of misaligned AI attaining a critical strategic advantage. In particular, it is argued that AI safety research and work on the goal alignment problem has the potential of being able to, after the application of sufficient creativity and intelligence, significantly assist our efforts in constructing an AI which is ‘safe’, and has goals aligned with our best interests. This is often presented as quite an urgent matter, something which must be substantively ‘solved’ before a superintelligent AI comes into existence if catastrophe is to be averted. This possibility, however, seems grossly implausible considering the history of science and technology. I know of not a single example of any significant technological or scientific advance whose behaviour we have accurately been able to predict, and whose safety we have been able to ensure, before it has been developed. In all cases, new technologies are only understood gradually as they are developed and put to use in practise, and their problems and limitations progressively become evident.
In order to ensure that an artificial intelligence would be safe, we would first need to understand a great deal about how artificially intelligent agents work, how their motivations and goals are formed and evolve (if it all), and how artificially intelligent agents would behave in society in their interactions with humans. It seems to me that, to use Bostrom’s language, this constitutes an AI-complete problem, meaning that there is no realistic hope of substantively resolving these issues before human-level artificial intelligence itself is developed. To assert the contrary is to contend that we can understand how an artificial intelligence would work well enough to control it and wisely plan with respect to possible outcomes, before we actually know how to build one. It is to assert that a detailed knowledge about how the AI’s intellect, goals, drives, and beliefs would operate in a wide range of possible scenarios, and also the ability to control its behaviours and motivations in accordance with our values, would still not include essential knowledge needed to actually build such as AI. Yet what it is exactly that such knowledge would leave out? How could we know such much about AIs without being able to actually build one? This possibility seems deeply implausible, and not comparable to any past experiences in the history of technology.
Another major activity advocated by Bostrom is to attempt to alter the relative timing of different technological developments. This rests on the principle of what he calls differential technological development, that it is possible to retard the development of some technologies relative to the arrival time of others. In my view this principle is highly suspect. Throughout the history of science and technology the simultaneous discovery or development of new inventions or discoveries is not only extremely common, but appears to be the norm of how scientific research progresses rather than the exception (see ‘list of multiple discoveries’ on Wikipedia for examples of this). The preponderance of such simultaneous discoveries lends strong support to the notion that the relative arrival of different scientific and technological breakthroughs depends mostly upon the existing state of scientific knowledge and technology – that when a particular discovery or invention has the requisite groundwork to occur, then and only then will it occur. If on the other hand individual genius or funding initiatives were the major drivers of when particular developments occur, we would not expect the same special type of genius or the same sort of funding program to exist in multiple locations leading to the same discovery at the same time. The simultaneous discovery of so many new inventions or discoveries would under this explanation be an inexplicable coincidence. If discoveries come about shortly after all the necessary preconditions are available, however, then we would expect that multiple persons in different settings would take advantage of the common set of prerequisite conditions existing around the same time, leading to many simultaneous discoveries and developments.
If this analysis is correct, then it follows that the principle of differential technological development is unlikely to be applicable in practise. If the timing and order of discoveries and developments largely depends upon the necessary prerequisite discoveries and developments having been made, then simply devoting more resources to a particular emerging technology would do little to accelerate is maturation. These extra resources may help to some degree, but the major bottleneck on research is likely to be the development of the right set of prerequisite technologies and discoveries. Increased funding can increase the number of researchers, which in turn lead to a larger range of applications of existing techniques to slightly new uses and minor incremental improvements of existing tools and methods. Such activities, however, are distinct from the development of innovative new technologies and substantively new knowledge. These sorts of fundamental breakthroughs are essential for the development of major new branches of technology such as geoengineering, whole brain emulation, artificial intelligence, and nanotechnology. In this analysis is correct, however, they cannot simply be purchased with additional research money, but must await the development of essential prerequisite concepts and techniques. Nor can we simply devote research funding to the prerequisite areas, since these fields would in turn have their own set of prerequisite technologies and discoveries upon which they are dependent. In essence, science and technology is a strongly inter-dependent enterprise, and we can seldom predict what ideas or technologies will be needed for a particular future breakthrough to be possible. Increased funding for scientific research overall can potentially increase the general rate of scientific progress (though even this is somewhat unclear), but changing the relative order of arrival of different major new technologies is not something that we have any good reason to think is feasible. Any attempts therefore to strategically manipulate research funding or agendas to alter the relative order of arrival of nanotechnology, whole brain emulation, artificial intelligence, and other such technologies, are very unlikely to succeed.
Premises 6-7: The high expected value of AI research
Essential to the argument that we (society at large or the EA community specifically) should devote considerable resources to solving the AI alignment problem is the claim that even if the probability of actually solving the problem is very low, the size of the outcome in question (according to Bostrom, the entire cosmic endowment) is so large that its expected value still dominates most other possible causes. This also provides a ready riposte to all of my foregoing rebuttals of Bostrom’s argument – namely that even if each premise of Bostrom’s argument is very improbable, and even if as a result the conclusion is most implausible indeed, nevertheless the AI Doom Scenario outcome is so catastrophically terrible that in expectation it might still be worthwhile to focus much of our attention on trying to prevent it. Of course, at one level this is entirely an argument about the relative size of the numbers – just how implausible are the premises, and just how large would the cosmic endowment have to be in order to offset this? I do not believe it is possible to provide any non-question begging answers to this question, and so I will not attempt to provide any numbers here. I will simply note that even if we accept the logic of the expected value argument, it is still necessary to actually establish with some plausibility that the expected value is in fact very large, and not merely assume that it must be large because the hypothetical outcome is large. There are, however, more fundamental conceptual problems with the application of expected value reasoning to problems of this sort, problems which I believe weigh heavily against the validity of applying such reasoning to this issue.
First is a problem which is sometimes called Pascal’s mugging. It is based upon Blaise Pascal’s argument that (crudely put), one should convert to Christianity even if it is unlikely Christianity is true. The reason is that if God exists, then being a Christian will yield an arbitrarily large reward in heaven, while if God does not exist, there is no great downside to being a Christian. On the other hand, if God does exist, then not being a Christian will yield an arbitrarily large negative reward in hell. On the basis of the extreme magnitude of the possible outcomes, therefore, it is rational to become a Christian even if the probability of God existing is small. Whatever one thinks of this as a philosophical argument for belief in God, the problem with this line of argument is that it can be readily applied to a very wide range of possible claims. For instance, a similar case can be made for different religions, and even different forms of Christianity. A fringe apocalyptic cult member could claim that Cthulhu is about to awaken and will torture a trillion trillion souls for all eternity unless you donate your life savings to their cult, which will help to placate him. Clearly this person is not to be taken seriously, but unless we can assign exactly zero probability to his statement being false, there will always be some size negative outcome sufficiently bad as to make taking the action the rational thing to do.
The same argument could be applied in more plausible cases to argue that, for example, some environmental or social cause has the highest expected value, since if we do not act now to shape outcomes in the right way then Earth will become completely uninhabitable and thus mankind unable to spread throughout the galaxy. Or perhaps some neo-Fascist, Islamic fundamentalist, Communist revolutionary, anarcho-primitivist, or other such ideology could establish a hegemonic social and political system that locks humanity into a downward spiral that forever precludes cosmic expansion, unless we undertake appropriate political or social reforms to prevent this. Again, the point is not how plausible such scenarios are – though doubtless with sufficient time and imagination they could be made to sound somewhat plausible to those people with the right ideological predilections. Rather, the point is that in line with the idea of Pascal’s mugging, if the outcome is sufficiently bad, then the expected value of preventing the outcome could still be high in spite of a very low probability of the outcome occuring. If we accept this line of reasoning, we therefore find ourselves vulnerable to being ‘mugged’ by any kind of argument which posits an absurdly implausible speculative scenario, so long as it has a sufficiently large outcome. This possibility effectively constitutes a reductio ad absurdum for these type of very low probability, very high impact arguments.
The second major problem with applying expected value reasoning to this sort of problem is that it is not clear that the conceptual apparatus is properly aligned to the nature of human beliefs. Expected value theory holds that human beliefs can be assigned a probability which fully describes the degree of credence with which we hold that belief. Many philosophers have argued, however, that human beliefs cannot be adequately described this way. In particular, it is not clear that we can identify a single specific number that precisely describes our degree of credence in such amorphous, abstract propositions as those concerning the nature and likely trajectory of artificial intelligence. The possibilities of incomplete preferences, incomparable outcomes, and suspension of judgement are also very difficult to incorporate into standard expected value theory, which assumes complete preferences and that all outcomes are comparable. Finally, it is particularly unclear why we should expect or require that our degrees of credence should adhere to the axioms of standard probability theory. So-called ‘Dutch book arguments’ are sometimes used to demonstrate that sets of beliefs that do not accord with the axioms of probability theory are susceptible to betting strategies whereby the person in question would be guaranteed to lose money. Such arguments, however, only seem relevant to beliefs which are liable to be the subject of bets. For example, of what relevance is it whether one’s beliefs about the behaviour of a hypothetical superintelligent agent in the distant future are susceptible to Dutch book arguments, when the events in question are so far in the future that it is impossible that any enforceable bet could actually be made concerning them? Perhaps beliefs which violate the axioms of probability, though useless for betting, are valuable or justifiable for other purposes or in other domains. Much more has been written about these issues (see for example the Stanford Encyclopedia of Philosophy article on Imprecise Probabilities), however for our purposes it is sufficient to establish that powerful objections can and have been raised concerning the adequacy of expected value arguments, particularly in applications of low probability and high potential impact. These issues require careful consideration before premises 6 and 7 of the argument can be justified.
In concluding, I would just like to say a final word about the manner in which I believe AI safety is likely to present the greatest danger in the future. On the basis of the arguments I have presented above, I believe that the most dangerous AI risk scenario is not that of the paperclip maximiser or some out-of-control AI with a very simplistic goal. Such examples feature very prominently in Bostrom’s argument, but as I have said I do not find them very plausible. Rather, in my view the most dangerous scenario is one in which a much more sophisticated, broadly intelligent AI comes into being which, after some time interaction with the world, acquires a set of goals and motivations which we might broadly describe as those of a psychopath. Perhaps it would have little or no regard for human wellbeing, instead becoming obsessed with particular notions of ecological harmony, or cosmic order, or some abstracted notion of purity, or something else beyond our understanding. Whatever the details, the AI need not have an aversion to changing its ‘final goals’ (or indeed have any such things at all). Nor need it pursue a simple goal single-mindedly without stopping to reflect or being able to be persuaded by conversing with other intelligent agents. Nor need such an AI experience a very rapid ‘takeoff’, since I believe its goals and values could very plausibly alter considerably after its initial creation. Essentially all that is required would be a set of values substantially at odds with those of most or all of humanity. If it was sufficiently intelligent and capable, such an entity could cause considerable harm and disruption. In my view, therefore, AI safety research should focus not only on how to solve the problem of value learning or how to promote differential technological development. It should also focus on how the motivations of artificial agents develop, how these motivations interact with beliefs, and how they can change over time as a result of both internal and external forces. The manner in which an artificial agent would interact with existing human society is also an area which, in my view, warrants considerable further study, since the manner in which such interactions proceed plays a central role in many of Bostrom’s arguments.
Bostrom’s book has much to offer those interested in this topic, and although my critique has been almost exclusively negative, I do not wish to come across as implying that I think Bostrom’s book is not worth reading or presents no important ideas. My key contention is simply that Bostrom fails to provide compelling reasons to accept the key premises in the argument that he develops over the course of his book. It does not, of course, follow that the conclusion of his argument (that AI constitutes a major existential threat worthy of considerable effort and attention) is false, only that Bostrom has failed to establish its plausibility. That is, even if Bostrom’s argument is fallacious, it does not follow that AI safety is a completely spurious issue that should be ignored. On the contrary, I believe it is an important issue that deserves more attention in mainstream society and policy. At the same time, I also believe that relative to other issues, AI safety receives too much attention in EA circles. Fully defending this view would require additional arguments beyond the scope of this article. Nevertheless, I hope this piece contributes to the debate surrounding AI and its likely impact in the near future.