Can computers think like humans? Reviewing Erik Larson’s “The Myth of Artificial Intelligence”
In his recent book The Myth of Artificial Intelligence: Why Computers Can’t Think the Way We Do, AI researcher Erik J. Larson defends the claim that, as things stand today, there’s no plausible approach in AI research that can lead to generalized, human-like intelligence.
It’s important to understand what the author is claiming- and what he’s not claiming. He’s not claiming that computers can never think like humans, as some philosophers of mind have claimed. Rather, his position is- if there’s indeed a way to make computers think like humans, we haven’t the foggiest what that is. Our current approaches- no matter how promising they might seem- are all dead ends. He contrasts this with the prevailing optimism about AI: the perception that current approaches are on the path to generalized intelligence, and the problems of this approach are, at least in theory, solvable. If there’s a need for radical rethinking, it’s at the level of details, not fundamental principles. Thought this way, human-like computers seem just a matter of time. Larson, on the other hand, argues that even the fundamental theoretical principles of current AI approaches are non-starters.
If we were to imagine a spectrum of opinions about optimism/skepticism about AI, this is where the book would fall:
The author develops two arguments in defense of his case:
1. The inadequacy of inductive inference.
All of the current approaches in AI (or at least the most promising ones) are based on a certain model of thinking: inductive inference. However, generalized human intelligence can never be developed just on inductive inference. Meaning, so long as AI approaches share this principle, they are all doomed to fail.
One reason this argument is interesting is, the question the book is investigating- whether computers can think like humans or not- might strike one as too “philosophical”: wouldn’t this depend on how we define what constitutes human intelligence? With that comes the worry of subjective goalpost-setting. However, this argument sidesteps this issue altogether, because the author attempts to show that any inductive inference-based AI would keep making mistakes a human would never make. Regardless of how we define “intelligent”, it’s clear that an AI making those mistakes isn’t that (Compare: we might not have a precise definition of what it means to be “healthy”, but we can agree that a person hospitalized from COVID clearly doesn’t qualify).
2. Human intelligence is ineffable.
Human intelligence is primarily based on a mode of inference called abductive inference. This type of inference cannot be reduced to other forms (like induction or deduction), and nor do we know how to code or program it. Put differently, the mode of inference most fundamental to human intelligence is an utter mystery. Since none of the current AI approaches are based on (genuine) abductive inference, and since we don’t know how we could even start writing such a program- there’s really no path for current AI to reach human-like intelligence.
As a case study of sorts, the author talks about the inability of extant AI to make progress in natural language understanding. This didn’t strike me as a standalone argument, however, just a demonstration of the concepts already developed in the foregoing pages.
I’ll now reconstruct these arguments and offer my evaluations, but first:
Disclaimers
One, I’m coming to this book as a complete layman. Not only do I have no relevant expertise, but this is also in fact the first book-length treatment that I’ve read on the topic. As such, I bring no background knowledge to this analysis: I can only evaluate the validity of the book’s arguments, not fact-check the premises. Also, my objections might have fairly simple answers in the literature- but I wouldn’t know about them since I’m only reviewing the case as it’s contained in the book.
Two, in addition to this AI-to-human intelligence discussion, there are two other sections of the book on the history and sociology of the current AI culture, but I’ll not be reviewing them. This is definitely not to say these sections aren’t interesting: the first two chapters, for example, recount how Alan Turing went from essentially being an AI-skeptic to an AI-hopeful due to his involvement in the war effort; and the last section discusses the role of theorizing in science and how an overly data-obsessed AI culture might be jeopardizing it. However, since they are narrative accounts and not arguments, there’s not much I can provide in terms of evaluation. Also, I’m more interested in the prospects of AI becoming human-like, which I believe forms the core of the book anyway.
With that said…
Argument 1: The problem with inductive inferences
Dumb, narrow, and brittle computers
A useful starting point is to first get a grasp on the problems with AI as it stands today. Computer intelligence tends to be very “narrow” in scope, and this is by design: a chess-playing AI, because of its high degree of specialization, can’t also play checkers. An extreme case of this is what the author calls the “brittleness” problem: not only can a narrow AI not accomplish other tasks, but even slight deviations in the setup- which wouldn’t even register to humans- messes up computer output completely. Consider an AI that can play the game Breakout perfectly, which requires moving a paddle back and forth to bounce a ball back to the bricks. Moving the paddle a few pixels closer to the bricks wouldn’t drastically affect the performance of a human player, yet do the same for an AI and its “entire system falls apart”. The same goes with image detection software: they usually have a very high rate of success, but just changing a few pixels here and there messes up the system completely.
It’s important to know what these problems are telling us about AI capacities, especially since it’s easy to be misled by the successes of modern AI in a large variety of important tasks (especially for a layman like me). To this end, consider this analogy. Let’s say I, a human, am given ten high school math problems to solve. I’ve been away from high school math for a while, so I have no clue as to where to even start. I turn in my answer sheets virtually empty. Realizing I require training, I acquire a high school math book and start studying it diligently. This turns out to be quite an ordeal, but the next time I appear on the test, my scores do improve dramatically- I get a 9 out of 10, in fact. This clearly demonstrates my proficiency in high school math.
“But hold on”, says my professor, “Let’s actually look at the questions to see which one you got wrong”. As it turns out, all questions on the test were copied almost exactly from the book, but the one I got wrong had slightly different wording. This is curious: it’s essentially identical to what’s in the book, with a relatively minor change in wording or values. How could I have gotten this wrong?
As it turns out, my way of preparing for the test was not to actually learn math but to commit the entire book to memory. That way, if a question from the book appeared verbatim on the test- I would recognize it, and perfectly regurgitate the answer. However, since my learning is entirely memory-based, even a slight deviation from the way the questions appear on the test renders me helpless.
Given this new fact in place, how would my math proficiency be evaluated? I think people would agree that I have made no progress since I took the test the first time. My high score in the second test is in fact a very misleading indicator of success since the test questions couldn’t tell a purely memory-based performance from an understanding/learning-based one. Indeed, when I’m administered another ten-question test, with each one having a slightly different wording from their counterparts in the book- I fail to answer any of them. Clearly, my math proficiency- if it can be called that- is incredibly narrow and brittle, going only as far as memorizing and reproducing answers from the book.
When an AI fails at playing Breakout or image recognition after a slight deviation in the “test”, that mistake is demonstrative of an overall lack of “intelligence”- regardless of how high their success rate is otherwise. Whatever human intelligence might be, it’s not that. This is why it’s important to pay attention to these problems and ask- why do these problems persist in current AI? Are the results of lack of fine-tuning in the code’s details, which will be solved as technology progresses? Or is it that they result from fundamental premises on which current approaches are based?
The author argues in favor of the latter. All current AI is based on a certain model of thinking or inference known as induction, and the system-exposing mistakes are direct results of the constraints inherent in such an inference.
Inductive inference
Very simply put, the inductive inference is learning from experience. If we see the sunrise in the east every day, we naturally conclude that the sun would also do the same tomorrow. What gives us confidence in this assertion is the fact that we’ve always seen this happen. The same goes with there being no white ravens or blue emeralds: we haven’t seen any such examples in all our observations, meaning they probably aren’t there. More generally, induction is generalizing a conclusion to a population from a sample, to extrapolate our findings from the known to the unknown. Philosophers have often pointed out that this mode of inference is fundamentally unjustified: the fact that we’ve seen something happen, no matter how frequently or consistently, shouldn’t give us any confidence of continuing to observe it in the future. This “problem of induction”, while still a matter of debate among philosophers of science and epistemology, isn’t strictly relevant to the AI debate. What we’re concerned with is whether computers can display human-like intelligence, not perfect intelligence. The question of justifying human inferences is a different issue altogether, and we’ll let the philosophers mull over this.
However, the author argues, current AI suffers from a more practical problem of induction that is, in fact, relevant. The success of inductive inferences relies on getting enough sample sizes. Pronouncing a judgment on vaccine or drug efficacy just based on a ten-people trial, for example, is bad inference. To make good inferences, we need to take enough data into account, and often various kinds of it. This is where AI comes up short: while it can make successful inferences in very narrow contexts- like playing games with set rules- the real world is way too dynamic and complex for it to make accurate predictions. In other words, the problem here is failing to reach large and varied enough sample sizes. Sure, AI systems might excel at a high degree in a narrow enough task, but it has a non-negligible fraction of mistakes, and those mistakes expose it as lacking intelligence altogether (consider the math problem analogy above). Following the writer and statistician Nassim Nicholas Taleb, the author characterizes these failed inferences as involving complex (non-binary) choices and a high degree of randomness. As it happens, the real world is filled with problems with just these two features, and often these decisions lead to high-impact consequences.
More specifically, the author identifies two constraints that limit even our most promising AI approaches (involving machine learning and big data). The first he calls the empirical constraint: the system is only as smart as we teach it to be. For instance, if we tell an AI to identify all references to a company called “Blue Box Inc.”, it can only do so based on explicit evidence- like occurrences of the words “blue” and “box” in order. However, consider a case where someone posts a status on Facebook about the company, and a commenter below uses the pronoun “it” to refer to it. This comment is a reference to the company, but the AI wouldn’t be able to recognize it- because that requires parsing context, which cannot be done just with the initial information input. Alternatively, someone might be talking about a product of the company Blue Box and not the company itself- which shows up in the system as a false positive. These are cases where a human can easily pick up the references, unlike automated AI.
The second constraint inherent to induction-based AI approaches is the frequency assumption: things are classified in a certain way based on their frequency. For example, if the computer notices that whenever “Inc.” follows a pair of words (like “Blue Box”) it’s most frequently referring to a company, the next time the computer sees “X Y Inc.” it will slap the “company” label on it with high confidence. Of course, in the real world, the “Inc.” label might be used in other contexts as well: perhaps as an abbreviation. A telling example of this is sarcastic or humorous pieces of news. A parody article, for example, might have references to violence and handguns- and the computer will dutifully classify this article as having a “serious” subject matter. Again, these clues are easily recognizable by humans, but not so much AI who can’t process contextual information. Examples of mistakes like this might be low compared to the high success rate, but like the one question I got wrong in my math test- these are the mistakes that expose AI as having no intelligence, to begin with.
Taking stock
At this point, it would be useful to evaluate the case thus far and see if the author is really succeeding in what he wants to achieve. In the first argument, the author wants to convince us that any inductive inference-based AI- just by virtue of its fundamental premises- will never be intelligent. Inductive inference systems can’t succeed in principle. For this to be successful, the author would need to show that these problems cannot be solved by just adding more data, which we can always hope to do as technology progresses (at least conceivably).
Thus far in the book though, I don’t think the author has been able to show this. The examples he gives about constraints of AI- not recognizing a pronoun referent or sarcastic, low-frequency instances- don’t strike me as insurmountable problems. Perhaps, with more data, computers can be trained to pick up on clues in an article that tells it that this is meant to be humorous. In such a case, the computer would ignore its higher frequency-based assumption (e.g. “news pieces with references to handgun and violence are serious”), and choose a different category for the article. Sure, an AI researcher might give reasons as to why these things are impossible for technical reasons, but my question is- do those reasons follow from the fact that induction is taken as a premise to this entire approach? Because that’s what the author has been trying to show- AI systems are doomed to failure for a specific reason: because they’re induction-based. Not because of lack of data, which seems fixable. I don’t think the author has been able to convincingly demonstrate this just yet.
He has one additional piece to his case though: adding more data might be a problem unto itself.
Just add data?
To show that induction-based approaches can’t be improved by adding more data, the author draws from his own research experience.
There are two issues, one more serious than the other, to just a data-centric solution. First, there’s the question of how much data to add for the AI to start making accurate predictions. Apparently, even when asked to answer a simple question like why is the street wet, computers need to be fed just an astonishing amount of data. This is because a lot of what we take to be “common sense” actually includes a universe of implicit knowledge that we don’t really even consider to be information. When feeding data to a computer, one has to carefully include groundbreaking insights such as pouring a liquid into a glass container with no cracks and only one opening will fill it up.
The second and perhaps more fundamental- problem is what the author calls the selection problem. To have the AI answer questions, it would need access to a body of knowledge arranged by topic. For example, to train a computer to answer the question why is the street wet, we want to include all possible causes about the street being wet: rain, kids using super soakers, fire hydrants, fire departments, flood, malfunctioning washing machines, overflowing sewers, and so on. However, with so much information at its disposal, the computer now struggles to select the pieces relevant to the problem at hand. Put differently, to be comprehensive, you’d need to include “odd” information (e.g. malfunctioning washing machines) into your street-being-wet explanation list- which defeats the purpose of organizing the computer’s knowledge by “topic”. It seems an increasing amount of data would have to be at a trade-off with usefully categorizing the data, so the computer doesn’t struggle with relevance.
Taking stock again
A lot rides on how one interprets this problem of adding more data. For the author, this shows that one can’t infer their way through the world if they’re relying exclusively on performing induction from available data, regardless of how much or how varied that dataset is. Indeed, selecting the right sorts of information from your knowledge base itself requires inference. And whatever this form of inference is, it’s not induction. If you’re convinced by this line of argument- then I think you should be convinced by the author’s overall case, the idea that current induction-based AI approaches are all doomed to fail. If, however, you can think of a way out of this big data-relevance mess (or you’re familiar enough with the AI literature to know a solution), then you’d probably disagree with the author. You might also be undecided on the issue. I think I’m in this camp: I do see the seriousness of the selection problem, but I’m not totally convinced if that’s fundamental to the mode of inference.
Not to undersell his case but, this is the decision that the author’s entire first argument hinges on. One could read his previous examples about AI mistakes, but not be convinced of their intractability: it doesn’t seem impossible that more data could solve them. The only way to block that suspicion is to show adding more data is intrinsically problematic, and in fact processing, it requires a distinct sort of inference.
Argument 2- Human intelligence is ineffable
Abductive inferences
According to the author (and the philosopher Charles Sanders Peirce, whom he cites extensively), the mode of inference most fundamental to human intelligence is abductive inference. The abductive inference is essentially guessing- how, after having observed a phenomenon, we can identify its cause without spending a lot of time inductively crunching data. So after having seen the road as being wet, we can almost immediately come to a conclusion about what might be the case from context- whether it be rain, or the sewers, or something else entirely. This is somewhat uncanny, as this amounts to immediately choosing the right hypothesis- or at least a plausible one- among countless others. As a first attempt to formalize this mode of inference, Peirce observes that it’s a reaction to surprise:
· The surprising fact, C, is observed.
· But if A were true, C would be a matter of course.
· Hence, there is reason to suspect that A is true.
This mode of inference is most clearly seen in detective novels and scientific investigations, where plausible theories are selected despite severe underdetermination by the data at hand. But abduction is carried out at the most mundane decision-making junctures as well- even when it comes to observing a flower in the garden and concluding that it is indeed a flower.
And yet, the author argues, we have no idea how to teach a program to abduce. More fundamentally, we don’t know how to precisely formalize abduction. Without crossing that bridge, AI can never reach human intelligence.
This argument is much simpler than the previous one, but it stands on two crucial premises- one, abduction is fundamental to human intelligence; and two- we haven’t made any progress with abduction in the realm of AI.
Is abduction fundamental to human intelligence?
The author explains this mode of inference and with a lot of examples, but only gives one argument for its fundamentality- abductive inference cannot be reduced to deduction or induction (or both). In fact, when written out, the formula for abduction looks like bad reasoning, “broken deduction” to use Peirce’s terminology. The author substantiates this point with these representative examples:
Deduction: All the beans from this bag are white. These beans are from this bag. Therefore, these beans are white. (A -> B; A; therefore B)
Induction: These beans are from this bag. These beans are white. Therefore, all the beans from this bag are white. (A; B; therefore A -> B)
Abduction: All the beans from this bag are white. These beans are white. Therefore, these beans are from this bag. (A -> B; B; therefore A)
Especially when one contrasts abduction with deduction, it seems obvious that the former cannot be reduced to the latter. If this argument is successful, that would indeed prove the irreducibility of abduction. However, it seems to me like the abductive inference isn’t telling us the full story: the example can be unpacked to yield some hidden premises. Maybe something like:
All the beans from this bag are white. These beans are white. The beans are close to the bag. The source of beans would need to be close to it and contain white beans. Therefore, these beans are from this bag.
This sounds really clumsy, and a logician can probably straighten out the formulation further, but I think it shows that abduction if made sufficiently explicit, can possibly be translated to deduction and/or induction. The penultimate premise in this new scheme, for example, can gain support from induction.
Without adequately ruling out the reducibility of abduction to deduction and/or induction, however, I don’t think the first premise of this argument- abduction being fundamental to human intelligence- can be established. Maybe we’re incredibly adept deduction-induction machines, but the conclusions come so fast- reflexively- that it appears “magical”, or even “guess-like”. This is just a possibility, and I personally do feel partial to our intelligence having a component in addition to deduction and induction- just that I wasn’t convinced by the author’s case to that effect.
Abduction-based AI?
This leads us to what I think is a rather glaring omission from the book: the author doesn’t discuss AI systems that at least purport to incorporate abductive inference. In the notes section of the book, he mentions such attempts have indeed been made, but they are inductive in disguise. I have two comments in response to that: first, whether an abductive-AI can be reducible to an inductive-AI seems like a very substantive discussion, and I think it would’ve been better to let the readers decide this after providing relevant details. Second, saying that “abductive AI” is flawed because they’re in fact inductive, is to say- you need to be convinced of my other argument for this one to make sense. But as I mentioned in my evaluation of the first argument, readers can differ on whether the author has indeed identified fundamental, in-principle problems with induction-based approaches. This is not a problem with the argument per se, just that the two arguments are collapsing into one, and the case is no longer cumulative. As such, I think an additional chapter on abductive inference-based AI- or even a long appendix- could’ve been very beneficial to the author’s case.
A comment on style
Since this is a book review, I wanted to offer some good faith criticism of the book’s organization as well. To reconstruct the first entire argument, I had to pull material from three chapters: 10, 11, 12. Chapter 10 introduces the general, “practical” problem of induction: life is way too complex to induce successfully, hence computer predictions tend to be so narrow and brittle. Chapter 11 talks about the two constraints that the author says are associated with inductive inference. Finally, the second half of chapter 12 argues that these constraints can’t be overcome by just adding more data. As my first “taking stock” section shows, I was wholly unconvinced of the argument until I read the section on selection problems (at which point I understood the argument, but remained somewhat undecided). However, the author strongly rules out induction at the ends of chapters 10 and 11, even before completing his case. Chapter 10 ends with “[T]o make progress in AI, we must look past induction”, and yet the author hasn’t even introduced the two constraints yet. Likewise, chapter 11 has this to say near the end:
Fundamentally, the underlying theory of inference is at the heart of the problem. Induction requires intelligence to arise from data analysis, but intelligence is brought to the analysis of data as a prior and necessary step. We can always hope that advances in feature engineering or algorithm design will lead to a more complete theory of computational inference in the future. But we should be profoundly skeptical. It is precisely the empirical constraint and the frequency assumption that limit the scope and effectiveness of detectable features — which are, after all, in the data to be syntactically analyzed. This is another way of saying what philosophers and scientists of every stripe have learned long ago: Induction is not enough.
Again, the author might be right in all of this- but he can’t make this pronouncement just yet, without having explained the problems of adding more data. Reading the concluding sections of chapters 10 and 11, I couldn’t help but think “is that all there is to his case?”
In addition, the first half of chapter 12 is taken up by discussions of abductive inference, before finishing the case for the inadequacy of induction. When the author discusses the aforementioned selection problem, he does it in the context of abductive inference- even though this should be a part of his case against induction-only approaches. Either that or I’ve fundamentally misunderstood the author’s arguments somewhere down the line.
I think a better structure would’ve been to make a continuous case against induction-only approaches (general problems, two constraints, how adding more data is of no help), and then making a separate case on the irreducibility of abduction. Of course, I’d also have liked another chapter on “abductive AI” as well.
These comments notwithstanding, the book is very accessible, and I can definitely recommend it as a first read to anyone looking to know more about AI and its limits.
