Thoughts on “Thinking, fast and slow”

I couldn’t resist buying a copy of Daniel Kahneman’s best-seller when returning from holidays. Several friends and colleagues told me it was a great book; it got great reviews; and Kahneman’s journal articles are invariably a good read, so I was curious.

Its general message is simple and intuitively appealing: Kahneman argues that people use two distinct systems to make decisions, a fast one and a slow one. System 1, the fast one, is intuitive and essentially consists of heuristics, such as when we without much thought finish the nursery rhyme ‘Mary had a little…’. The answer ‘lamb’ is what occurs to us from our associative memory. The heuristic to follow that impulse gives the right answer in most cases but can be lead astray by phrases like ‘Ork, ork, ork, soup is eaten with a …’. Less innocuous examples of these heuristics and how they can lead to sub-optimal outcomes are to distrust the unfamiliar, to remember mainly the most intense and the last aspect of an experience (the ‘peak-end rule’), to value something more after possessing it than before possessing it (the ‘endowment effect’) and to judge the probability of an event by how easily examples can come to mind.

System 2, the slow way to make decisions, is more deliberative and involves an individual understanding a situation, involving many different experiences and outside data. System 2 is what many economists would call ‘rational’ whilst System 1 is ‘not so rational’, though Kahneman wants his cake and eat it by saying that System 1 challenges the universality of the rational economic agent model whilst nevertheless not wanting to say that the rational model is wrong. ‘Sort of wrong sometimes’ seems to be his final verdict.

Let me below explore two issues that I have not seen in the reviews of this book. The first is on whether or not his main dichotomy is going to be taken up by economics or social science in the longer-run. The second, related point, is where I think this kind of ‘rationality or not’ debate is leading to. Both issues involve a more careful look at whether the distinction between System 1 and 2 really is all that valid and thus the question of what Kahneman ultimately has achieved, which in turn will center on the usefulness of the rational economic man paradigm.

Will Kahneman’s dichotomy be a winner? I very much doubt it. Mainly, it has to compete with much more intuitive dichotomies already floating around in the social science literature: conscious versus unconscious; gut versus brain; heart versus head; emotional versus analytical; positive versus normative; and individualistic versus communal. ‘System 1 and System 2’ doesn’t have quite the same intuitive appeal to it, does it now? If you like, it requires System 2 to understand and apply the dichotomy, which kills it as a term used by a whole discipline. Too many other things could also be System 1 and 2.

More importantly, Kahneman does not in fact have a very clear idea as to what the difference between those two systems is, even admitting himself that it is short-hand for a continuum of decision-making facilities rather than a straight dichotomy. Now, nearly all dichotomies are simplifications but good ones have only a small range of phenomena that are in between. To see this, take the concept of gender, which is a very popular simplification: yes, there are some humans that are hard to classify as either male or female but for the vast majority of humans the simple male/female dichotomy works perfectly well.

When it comes to decisions, the grey areas are large between these two systems. Mainly, it seems unlikely that most decisions are taken by just one of these two Systems. Someone learning to play piano is for instance partially using motor-reflexes and partially analysing notes, using both ‘Systems’ simultaneously. What does ‘fast and slow’ then mean in that situation? Even reading an equation involves both simple pattern recognition of symbols and conventions (such as x2=x*x= some variable multiplied by some variable) as well as deeper understanding of what is implied by that equation. Both Systems are thus often ‘on’ at the same time, making it hard to really talk about one of them as getting tired too easily or only being used for some problems.

Also, System 2 is not simply on or off: someone can for instance pay various degrees of attention to what is said. Again, that does not lend itself to the simple ‘slow’ label: paying only half attention is what, half slow? Indeed, one can argue that System 2 is never really off. The subconscious (normally labelled as part of the emotional System 1) can sometimes be surprisingly analytical and alert the conscious to found regularities that would normally require some time from a conscious System 2, such as when you suddenly realise that Mexico is not the 53rd state of America even though you didnt make that connection when you heard it earlier that day. Now, one can argue that this simply shows that the activation of System 2 can be subconscious and does not prove it is always ‘on’ but if one takes that excuse then this would mean that an individual would not even know which system is in use, making it even harder to label System 2 as rational and hard to say for any outside observer when it is on or off.

We can think about this issue deeper by pondering out loud how to formally model these Systems, something that Kahneman conspicuously neglects to do.

From an economic point of view, there are several options I can think of to model the distinction between System 1 and 2. If I had to write down a simple model I could solve then it would center around beliefs. The mind would then be involved in the recognition of the true type of a good we come into contact with, whereby System 2 represents a time investment in order to reduce the odds of attaining a wrong belief and System 1 is nothing other than the current set of beliefs arrived at in the past. So individuals could either go with their existing estimate or update it, involving more costly effort. It is easy to see how one can write this down and what the bits and bobs are to make it work as a model. Mainly, it would need some kind of ‘subconscious alerting rule’ based on expected value that determines whether an individual’s System 2 gets activated. That is a kind of meta-rule that itself would be subject to revision.

Whilst this model describes some of Kahneman’s examples such as the evaluation of the likelihood of something or the word that finishes the sentence ‘Mary had a little…’, it doesn’t fit many of Kahneman’s important examples of mental habits such as the peak-end rule or the endowment effect. Those latter habits are more like heuristics about beliefs rather than beliefs themselves. So my first-flush model would neither really describe what System 2 does, nor does the concept of beliefs capture the notion of ‘heuristics’, which are a set of rules as to how to approach a problem rather than a prior belief about a problem.

Suppose we then take the language of heuristics seriously and envisage Kahneman’s Systems as a series of ‘mental checks’ via which a belief gets updated. We then get a more expanded notion of these two Systems. What one then says about System 1 and 2 is that System 2 comes up with new heuristics about how to handle certain situations (like an expanded set of prior examples to compare a situation with) as well as meta-heuristics about which sub-heuristic to have running at what time. System 1 just applies the heuristics learned by System 2 at some point in the past.

Apart from fitting most of Kahneman’s examples, this kind of framework allows for big differences across people, which is useful when considering actual empirical work. Within this kind of framework, one doesn’t need to generalise over a whole population how smart System 1 and 2 are: the unconscious of one person can be very smart if it employs a lot of ‘good’ heuristics thought up by that person previously. Hence it allows individuals to differ in terms of all the heuristics they have up and running ‘normally’, thus allowing for ‘fast and smart’ as well as ‘slow and dumb’. It would for instance allow for people who can do statistics very quickly without much effort or for the whom the peak-end rule is not an unalterable given, something Kahneman seems to think is impossible.

This appealing ‘heuristic maker’ versus ‘heuristic implementer’ story is not quite satisfactory either though because it begs the question of whether there in fact is something like a distinct rational System 2 at all: perhaps deep thought itself is nothing more than applying more meta-heuristics simultaneously rather than less. Alerting the conscious mind to the importance of solving a problem is then rallying the motivation to let loose more available heuristics on an issue in order to re-evaluate and thus over-write existing heuristics. This would mean there is no clean distinction at all but merely a matter of degrees of effort thrown at a problem in which case you could throw away the distinction between System 1 and 2 and replace it with the much simpler labels ‘a little effort’ and ‘a lot of effort’.

Even if we were to stick to the heuristics interpretation and were willing to say there is an ‘on’ switch to the ‘heuristics maker’ so that we can speak of two more or less distinct situations, this lends itself much more to the existing ‘learning versus applying’ dichotomy than a cumbersome System 1 and 2 labelling.

The next step again is to envisage whole sets of heuristics as a ‘mental model’ and that individuals select the mental model for every problem they encounter based on meta-rules of association, whereby System 1 is simply the application of these meta-rules and System 2 is both a re-evaluation of some meta-rules as well as the learning of new basic mental models. This fits in much better with the ‘categorisation’ theory of memory and thinking that is popular in cognitive psychology and neuroscience.

A problem with that kind of interpretation though is that the subconscious too is known to continuously make and apply mental models from a very early age, most notably of other people: humans are very good at very quickly predicting what other people will do and thus system one ‘recognises’ whether someone is hostile or not, something that requires the subconscious to have built up a mental model of another person. The fact that kids aged 2 can read the emotions in others makes it clear that building mental pictures of the motivations and actions of others is a continuous activity in itself that is hardly the domain of some super-rational System 2. At least, it would lead one to the situation of saying that young kids who can’t even talk or add up 2 plus 2 in many ways are more System 2 active than adults!

What the above attempts at thinking of possible formalisations show is that as soon as one tries to come up with a simple distinction between System 1 and System 2, one invariably arrives at other labels to describe the decision-making process. ‘Heuristics’, ‘Categorisation’, ‘Mental models’, ‘Meta-rules’, etc. then come to mind as more descriptive labels for what is going on. So it is those terms that I see having a future. At best, ‘System 1 and 2’ only has a limited future until the literature decides on a deeper interpretation.

In terms of the whole rationality debate, the main contribution that Kahneman makes with this book, and that he in my opinion has made throughout his career, is to keep economists honest. This is a very valuable thing to do and this book once again does a great job of doing it.

Kahneman’s value derives from the great temptation amongst economists, particularly those of a strong theoretical bent, to fall in love with their abstractions and to pretend they are accurate descriptions of how things really work. Whereas the early economists were explicit about how their view of ‘rational economic man’ was merely an abstraction, later generations have far too often taken that abstraction and other closely associated ones (such as the whole notion of stable preferences, discount rates, loss-aversion, risk-aversion, etc.) literally. The number of economists I know who pretend to their students and themselves that things ‘derived’ from these abstractions, like the Welfare theorems, ‘prove’ things about the real world is astoundingly high. That pretence cannot be confronted and belittled often enough. Our models are derivative of an understanding of the real world, not the ultimate source of that understanding.

Good models are like good maps: reflections of a lot of measurement and observation. Some maps might be of what a piece of land would look like if re-shaped in a particular way, like a map of a yet to be built canal, drawing on observations of gravity and rainfall. But no good map of the Manhattan area pretends New York is not really there because it is inconveniently non-convex and displays adaptive preferences! Similarly, to say on the basis of a map of Mars that this proves the earth has big craters is silly beyond belief, yet some of our teaching is precisely of that ilk, with the Welfare Theorems that are rammed down each generation’s throat being a good case in point. It doesn’t matter how beautiful or elegant those maps of Mars are, they still don’t describe or prove anything about Earth and it takes people like Kahneman to remind us that none of our maps are perfect and some distinctly look like Mars.

To philosophers of science or even to any well-read economist, the fact that humans do not behave perfectly rationally in many important situations is nothing new and Kahneman is not telling them anything that they couldn’t find in writings of centuries ago. Unfortunately, it is a point that cannot be made often enough as our vanity urges us on, every generation anew, to pretend that our assumptions are not abstractions that loosely describe some aspect of the real world, but the undoubted truth of how Earth looks or could look. Kahneman in that sense plays the vital role of this generation’s small child who points at the map in front of him and says ‘yes, Earth is also round like the planet on this map, but it is not blue everywhere’. Not only do we need him this generation, but we will need another one like him the next generation and the next. So I recommend this book precisely to all those who hate reading stuff like this: the map-drawers who don’t look out of the window enough.

Will Kahneman affect the maps of the future, i.e. will the mainstream economic view of human decision making be expanded to include some of his pet concerns? I don’t actually think any of the ‘System 1 foibles’ that Kahneman points to will make it because ultimately economics is about how to explain as much as possible with as little mental effort as possible and from that point of view, the question is not which mental foibles we have but rather which important areas of human behaviour are poorly illuminated within the mainstream. It is that ‘potential additional market for economic advice’ which will determine what gets taken up in the standard canon, not any ex ante appraisal of the internal life of the mind. To use the map analogy, the question for those who refine the maps of whole planets is not the details of each shrub, but rather whether there is a market for maps that include continents not yet on it.

The area of most fruitful expansion for economics and hence where I expect map building to venture next is the whole business of the loyalties that underpin families, teams, and even countries. But that discussion should be had another day!

This entry was posted in Economics and public policy, Education, Geeky Musings, Information, Life, Literature, Methodology, Philosophy, Science, Social, Society. Bookmark the permalink.

10 Responses to Thoughts on “Thinking, fast and slow”

  1. I found that a great complement to Thinking, Fast and Slow was Expert Political Judgement by Tetlock. I reviewed it a few weeks ago.

  2. Tel says:

    http://en.wikipedia.org/wiki/Subsumption_architecture

    Except that if you put your hand on a hot plate by accident, then it jerks away very fast (turns around at the spinal column I believe, hence System 1, maybe even System 0 if you allow that)… but if you deliberately put your hand on a hot plate and you are expecting the hit, then you can hold it there longer (System 2, override capability).

    Also, deliberately deciding to repeated an action in order to train yourself to react quickly when you need to, is another way System 2 can override System1.

  3. Paul Frijters says:

    Tel,
    The true reflexes probably would count as not thinking at all. The learning of quick reactions would fit the heuristics story in that System 2 would be creating System 1 responses. It doesn’t fit the notion of fixed systems though.

  4. Paul Frijters says:

    Tel,
    The true reflexes probably would count as not thinking at all. The learning of quick reactions would fit the heuristics story in that System 2 would be creating System 1 responses. It doesn’t fit the notion of fixed systems though.

  5. desipis says:

    I haven’t read the book, but I tend to view this dichotomy of thought a little differently. Rather than two different systems of thought I conceive it as one system of thought that could be described as say ‘heuristic’ in nature. However within this system there would be two different types of heuristic. The first type (‘Type 1’) works basically as the System 1 is described, intuition based on learning though personal experience.

    The second type (‘Type 2’) would be ‘abstract’ or ‘rational’ heuristics. The more abstract the nature the broader range of experiences would be required to develop it and thus the more mental resources it takes to processes. This means the thoughts may need to occur across greater time (as serial/consequential thoughts). I’d also see these heuristics as typically developed through language and communication, representing heuristics socially crystallised across a much greater range of experiences than an individual can directly access. You could break this type into two subtypes, the heuristics of the language itself (grammar, relationship between the meaning of words, accepted truths, etc) and the heuristics of applying the language to real experiences. Sufficient experience with these ‘rational’ heuristics would train the mind to develop its own more efficient Type 1 heuristic that would function the same way (and possibly be flawed).

    In essence I would consider we’re using a combination of the different types of thoughts all the time, although we might use each type to different extents. Or to put it another way, all System 2 thought occurs within the framework of System 1 thought. Given the evidence of subconscious language processing, I’m not sure I’d see this divide as being related to the concious/subconscious divide, other than the way I described how more abstract thought would require more mental resources and thus be more likely to involve concious thought.

    I would still conceive reflexes to be thought though. They’d just be primitive subconscious thoughts that are simple heuristics learnt through limited, but significant experience. The significance of the pain sense would mean the heuristic developed would be dominant over whatever heuristic had caused the hand to (unintentionally) come in contact with the hot plate. If however a more dominant heuristic (developed over a more significant range of experience than the pain reflex) were to cause the hand to intentionally come in contact with the hot plate it would ‘override’ the reflex. Consider how older people are typically able to tolerate the pain compared to younger people, especially children.

  6. Paul Frijters says:

    Hi Desipis,

    clearly, the issue of how thinking actually goes is a very complex one and simple stories describing it must by design be simplifications so the object is to find dichotomies that capture important aspects.
    Your learning story for System 2, involving more communication (and teaching!) than mine, sounds eminently plausible to me. Very close to a ‘heuristic learner’ versus ‘heuristic implementer’ breakdown. There is of course a whole literature on heuristics with many categories of heuristics and a whole raft of papers on when we (can) learn which one.
    On a minor point, true reflexes, like the knee reflex or what happens once you start sneezing, are never learned but truly automatic. The impulses involved in the knee reflex dont even reach the brain as I recall, just the spinal cord. Some cant be unlearned either as anyone who has tried to stop a sneeze in progress can attest to!

  7. Brad says:

    Always amuses me that economists continually stumble across these discoveries years after they’ve been studied in psychological science and declare them new, original, exciting, etc.

  8. Paul Frijters says:

    Brad,

    it always amuses me how poorly many read a post they wish to criticize. Have you missed the sentence “To philosophers of science or even to any well-read economist, the fact that humans do not behave perfectly rationally in many important situations is nothing new and Kahneman is not telling them anything that they couldn’t find in writings of centuries ago.” ??

    Indeed, where in my post do you hear me say that either Kahneman or I was being original?

    If I claim any originality it is in the specific ideas as to how to mathematically model this stuff as I doubt any psychologist will have done it in that way, but I am open to be proven wrong.

  9. desipis says:

    Paul, one of the difficulties in dealing with the complexity is that to produce any meaningful predictions from such a model you need to have access to all the detail that went into the system. That’s not exactly feasible to do in the case of the human mind. In terms of economics it’s likely that the application is limited to statistical data collected about the trends in human behaviour rather than explicit models predicting what the behaviour will be. Models might help identify interesting or useful areas to collect data about though. They can also help explain why naive conclusions based on rational expectations don’t play out in reality.

    I think it’s important to understand more than just the fact that people might not always act rationally, but what factors influence people to act more or act less rationally. It’s particularly important when these factors aren’t just incidental parts of the social environment but are factors deliberately engaged to induce people to act against their own interests. If you’re going to base your conception of an ideal (economic) society as one based on free individuals acting in their own self interest (“free market/capitalism”) then attacks on individual’s ability to play their part does at least in some way appear as an attack on that ideal society.

    As far as reflexes go, I’m sure there are some ‘hardwired’ reflexes; the autonomic nervous system is likely full of them. Reacting to a tap on the knee by kicking out doesn’t exactly seem like the type of behaviour that is learned. I do recall (possibly incorrectly) reading somewhere that learning to walk relies to some extent on communication between the legs/feet and the spinal cord, indicating that there is at least some adaptation/learning that can occur in the spinal cord itself.

  10. desipis,

    yes, the question of ‘what to do if people are too complex to capture with a workable simplification’ is very important to economics. There are three answers, all tried extensively in the literature. The first is to pretend the simplification is unquestionably true in terms of policy prescripts for unknown reasons. This is the ‘as if’ defense, or, as I like to refer to it, the ‘god knows I am right’ defense. The second is to say ‘Well, I might not be able to work out how it all works perfectly but i hope to gain some insight by working out what each simplified set of assumptions gets me and see how much reality I am roughly capturing’. That’s my favourite one that I subscribe to and it includes a lot of economics, not just behavioural but also bits of micro and macro. The third one is to say ‘Give up on individual human models and just model higher level abstractions and their inter-relations (like interest rates and GDP), abandoning the pretense that you can get anywhere from the bottom up’. I find that a very defensible position also but see it as a case-by-case issue whether it does better than number 2 in terms of predictive fit, how easy it is to explain to students, and whether it engenders a useful mindset.

Comments are closed.