A Heuristic Example
At a cost of fifteen million dollars, British Airways presented a suite of printed advertisements in 1996 in newspapers and magazines. Each ad had a big photograph, a small inset photograph, and a few words to help us integrate them. One ad had a big photograph depicting a dove in a bird bath, a small inset photograph depicting the head of a grinning man in a shower, and a textual promise that British Airways provides private showers for its customers at its arrival lounges at Heathrow and Gatwick airports in London.
The inset photograph in these ads is always an intrusion. Like the big photograph, it has sharp rectangular boundaries. The two photographs are distinguished by technical features of photography and have content taken from different conceptual domains. But the ads typically suggest some morphological continuity between key elements of the inset photograph and key elements of the big photograph. For example, the head of the man in the shower rests on the body of the dove.
The ad I want to talk about presents a big black-and-white photograph printed in a technical style taken from the 1940s. It has low contrast, ambient lighting, and a hint of brownish coloring. It depicts an attractive young mother with neck-length hair, a simple string of pearls around her neck. She wears a short-sleeve knit top and a woven skirt, and sits cradling her infant in her arms so it reclines on its back, its head elevated and its knees slightly elevated, its legs sprawled in enviable comfort over her forearm, its outer thigh additionally supported by her left hand, on which she wears a set of wedding bands. She looks at the face of her sleeping child, who wears a cloth diaper in a period style.
But the head of the infant is occluded by an inset photograph, in color, with relatively high contrast and strong directional lighting, depicting a smiling, sleeping businessman who is balding slightly. He is reclined in an airline seat. The scale of the photographs has been chosen to make the shoulders of the businessman fit the shoulders of the infant and to locate his head where the infant's head would be if his weren't in the way. The text above the photograph reads, "The new Club World cradle seat. Lullaby not included." The text beneath reads, "Introducing the unique new business class cradle seat. It doesn't simply recline but tilts as a whole, raising your knees and relieving your body of stress and pressure. Pity you may not be awake to enjoy all the other changes on new Club World."
Obviously, even obtrusively, the ad is a blend. The reader of the ad is expected to unpack the blend, that is, to recognize it as a blend of input scenes, and to reconstruct, mentally, those inputs. The first input has the businessman in the business class cradle seat. The second input has the child cradled in its mother's arms.
The reader of the ad is meant to construct a counterpart mapping between some elements of the two inputs, that is, to find elements in one input that have counterparts in the other input. In this case, the cradle seat is the counterpart of the mother's cradling arms, and the businessman is the counterpart of the infant. These counterparts are projected to the blend: in the blend, the man is the baby and the seat is mother's arms.
There is selective projection from the inputs: some elements but not others are projected from each input to the blended space. From the space of mother-and-child we are to project comfort and happiness but not the baby's incompetence at business, speech, and dining; from the space of the businessman we project his reclining position but not our certainty that he would prefer to be someplace else.
There is emergent structure in the blend, that is, structure in the blend that is not available from either of the inputs. For example, in the British Airways blend, one can have the power, benefits, and responsibilities of advanced professional maturity and yet the comforts of infant dependency and irresponsibility.
There is also projection back to an input, that is, structure and inferences developed in the blend that lead us to reconceive an input. In this case, the unruffled comfort of the man-baby in the blend is to be projected back to the input with the businessman: regardless of our prejudices about airplane travel, the passenger in business class on British Airways is blissfully content.
Considerable work has been done to make it possible for the reader of the ad to establish even more elaborate counterpart connections between the two input spaces, thereby increasing the topological fit between the blended space and its input spaces. The posture of the baby in the mother's arms looks entirely natural but has been carefully posed to emphasize the raising of the knees and the tilting of the body and so to make the baby's posture blend more easily with the posture of the businessman in the cradle seat. The disparity in period styles of the photographs can be read as signifying temporal distance between the two corresponding input spaces, inviting the reader to frame the businessman as the grown-up infant. This makes the woman in the blend his actual mother as she appeared and as she nurtured him when he was an infant. Projecting these identities to the blend gives features of happiness and comfort there that come from being cradled not only like an infant but also by one's own mother, who adores her child and cares for it. The mother-and-infant input can now be further framed as something remembered fondly if unconsciously by the businessman as he reclines. His comfort in the cradle seat brings back to him, effectively if indistinctly, the pleasure of being in his mother's arms.
Actually, in the input with the businessman, his mother may be dead or otherwise incapable of holding him in this way, and if she is living and tried it, the result would be uncomfortable and possibly injurious. The businessman cannot have in reality what he remembers so fondly. But in the blend, the businessman can have again what he can no longer have--he can be cradled as a sleeping infant in his mother's arms. Who would not want to fly with British Airways, on these terms?
The advertising agency which developed this ad presented it to focus groups, whose members reassured the agency that its meaning was unmistakable.
But as The New York Times reported in an article that reproduced the advertisement and analyzed it, actual readers often objected to it (Bryant, 1996). "Many have written the airline to say they think the ad is Freudian, sexist, and even demeaning to flight attendants, who in the eyes of some beholders are represented by the mother in the ads."
The ad agency should not have been surprised. Conceptual blending is a quick, powerful, and inventive cognitive operation. A reader who makes sense of a blend and its inputs typically regards the interpretation as natural and inevitable, hardly an "interpretation" at all. In the case of the British Airways ad, whether we like the fact or not, we all possess the cultural frame in which an attractive young stewardess attends gently to the physical and psychological comfort of the older businessman. In this publicly-shared conceptual frame, it is understood that part of what the airline is selling is the attention of such a woman. Despite the elimination of the most suggestive uses of this conceptual frame in airline advertising ("Coffee, Tea, or Me?"), it is still routinely used in print and video advertisements, in which an attractive stewardess, no longer wearing a short skirt, stockings, and pumps, and additionally sharing the screen in alternation with an attractive steward, nonetheless gazes at the potential customer with an altogether attentive and pleasing look as if to suggest the prospect of her personal service. Once we have a young, attractive woman in the blend taking care of the businessman, it is straightforward for readers to recruit the frame of a stewardess and use it to structure both the blend and the businessman input, with a connection between the stewardess in the businessman input and the woman in the blend. In the blend, the stewardess is a young, attractive, attentive mother, and the inference projected back to the businessman input is that on British Airways, you can expect a stewardess who treats you in a similar fashion.
According to The New York Times, the ad agency tried to dismiss these complaints by explaining that "People thought way too much about this," but that seems unlikely: people did not think they were "interpreting" at all; they thought they were responding to what was "there." In the blend, the woman is attractive, however proper. She caresses the businessman intimately, holding him against her full bosom, gazing down at him lovingly. Readers know the standard cultural frame of Oedipal sexual attachment to the mother, and the father is conveniently out of the picture. Given the ways in which conceptual integration operates, it is natural for readers to interpret these features of the blend as prompts to create corresponding elements in the businessman input.
It is true that the ad reads, "Lullaby not included," which, it can be argued, makes the woman in the ad a mother and not a stewardess, but of course it is just as easy to read the phrase in the opposite direction: on British Airways, the businessman enjoys the woman's intimate physical and personal attention, with full sexual connotations, but not the lullaby because the woman in the businessman input is after all technically a stewardess rather than a mother: mothers sing lullabies but stewardesses must draw the line somewhere.
Conceptual integration is path-dependent: blends are put together and unpacked using what the understander already has in his conceptual structures. Highly charged conceptual structures have a lower threshold for activation. Readers can be expected to bring their politically and ideologically charged conceptual structures to bear whenever they can. When they do, they will not imagine that they are "reading in" to the blend any more than will the mechanical engineer who sees this ad and interprets it as principally displaying the posture achieved in the new cradle seat on British Airways.
For many readers of the ad, recruiting these politically and ideologically charged conceptual structures had a basic effect on their choice: the ad, which was intended to offer them incentives of personal comfort to lead them to reason to a commercial choice, led them to align commercial choice with political choice and so to refuse to patronize a corporation they judged to be sexist.
This introductory heuristic example suggests two points about political persuasion. First, the British Airways ad suggests an answer to the abiding question, why do politicians stick with hackneyed presentations of their positions and policies--new bridge, new deal, rebuilding the infrastructure, land of opportunity, new day, dawn in America, drawing us together, closer to the people, a stronger America--despite pleas in every forum that they offer something new? Perhaps the politicians lack invention, but alternatively, perhaps they know intuitively that offering a new blend runs the risk that members of their audience will bring to its interpretation highly charged conceptual frames, leading some of them to interpret it in a way that will hurt the politician, who will then have to back-peddle. In a climate of negative political attack, where a politician's utterances are read by opponents with an eye toward making them sound evil, the prudent politician will prefer clichés.
Second, it's nice to have your cake and eat it, too. I would not suggest
that the ad agency (or anyone else) behind the British Airways blend thought
this way, but imagine that market research had shown that the important
audience for these ads is senior male businessmen and that they love the
sensual, Oedipal connotations of the ad. In such a situation, a marketing
executive, calculating over payoffs and costs in different audiences, might
conclude that the payoff from stimulating the businessmen far outweighs the
cost of angering those who will find the ad both sexist and offensive
("All the letter writers have been women," reports The New York
Times), provided the accusation of sexism can be plausibly denied.
The marketing executive would then have incentive to approve a blend that
prompts for a persuasive interpretation strongly enough to induce it but tangentially
enough, or with enough distraction, to make it deniable. This ploy is a
standard weapon in the everyday political arsenal: offer a presentation
sufficiently new and sufficiently nuanced to suggest a persuasive
interpretation that the politician can then ignore, neglect, or disavow.
But this is risky business.
The British Airways ad shows us a basic cognitive operation, conceptual integration. Conceptual integration creates conceptual blends. The inventors of the ad used conceptual integration as a tool for guiding readers to reason toward a commercial choice. To the surprise and regret of British Airways, some readers connected that commercial choice to a political choice.
Conceptual integration is related to conceptual framing. We "frame" a situation when we see it as an instance of a general and conventional conceptual frame. For example, we "frame" Jacques Chirac as president of France. Framing is one type, a minimal type, of conceptual blending, in which one of the inputs is a specific situation (with, e.g., Chirac, France) and the other is a general and conventional conceptual frame (e. g., "president of a nation"). There is cross-space mapping of counterparts (Chirac-president, France-nation) and projection to the blend. There is emergent structure (e.g., president of France) not available from either input. These "single-framing" blends are often skeletal, and often provide structures that become new conventional frames: blending "France" and "president of nation" provides "president of France," which can become a frame; blending "secretary" and "treasurer" provides "secretary to the treasurer," which can become a conventional frame. Repeated blending can provide composite frames: " secretary to the president of France."
But frames are only one kind of input to blending. For example, the inputs to the British Airways ad are not just general conceptual frames; they are specific instances of frames, and each has more than one frame. The input with the businessman has frames for both air travel and businessman.
Conceptual framing and conceptual integration are part of the "backstage cognition" needed to interpret the British Airways ad. It provides a specific instance of my general claim for this paper:
When we flesh out the claim that reason and choice depend on backstage cognition, we get the following more specific claims:
This combination of features is partly responsible for the power of basic
cognitive operations. It is also partly responsible for the
difficulty--notorious in cognitive science„of recognizing the existence of
these cognitive operations, or the greater difficulty of noticing them as they
operate, or the yet greater difficulty of analyzing what it is they do when
There are various basic cognitive operations that fit this description. They include:
The automatic ease and power of these basic cognitive operations makes it natural for the social scientist to skip over both their analysis and the analysis of their biases, both in the people analyzed and in the social scientist.
To demonstrate the essential importance and complexity of backstage cognition
in the social sciences, I will consider in this article one basic cognitive
operation--conceptual integration. The theoretical work on conceptual
integration has been done jointly by Gilles Fauconnier and me, and presented
in various publications: Fauconnier & Turner (1994, 1996, in press a and b,
and in preparation), Turner and Fauconnier (1995 and in press), Turner (1996
and 1997), and Fauconnier (1997). The model we offer--the "network model of
conceptual integration"--has additionally served as the basis for Coulson
(1995), Freeman (1997), Grush & Mandelblit (1996) Mandelblit (in press and
manuscript), Oakley (1995), Roberts (1996), Sun (1994), and Veale (1996).
(The website for conceptual integration is
To demonstrate the unexamined importance of a basic cognitive operation
like conceptual integration to the social sciences, I will concentrate on two
kinds of mental events that are fundamental in the social
sciences--counterfactual reasoning and interdependent decision-making.
"If Churchill Had Been Prime Minister in 1938..."
I turn now to a less impressionistic description of conceptual integration and to my first topic in the social sciences--counterfactual reasoning.
The blend in the British Airways ad could have been suggested, approximately, by a counterfactual statement: "Our British Airways cradle seat could not be more comfortable if it were your mother's arms." In the counterfactual blend, the seat is your mother's arms. The input spaces to this counterfactual blend are (1) "you" as an adult in the airplane seat; (2) "you" as an infant in your mother's arms. Actually, discovering the input spaces in this case can be quite complicated. The input spaces could differ even across readers who all achieve essentially the same blend. For example, people never historically cradled by their mothers could still construct the intended counterfactual blend by using a generic "Mother Cradling Child" input space. The blend would in that case take a specific "you" and a specific "mother" from the specific input space, but it would take cradling from the generic input space. In the blend, the specific "you" would be cradled in a way that had no biographical referent. Since few readers can remember being cradled in diapers by Mom, perhaps we all need this generic space as an input to the blend. It is even possible for a reader to interpret the two "you"s in the two input spaces as entirely generic and so to construct a blend that applies generically to most people but not to the reader--perhaps the reader has some medical condition that prevents cradling or reclining in seats. Such a reader could even resent the blend.
A moment's thought shows that even in the relatively uncomplicated interpretation, the projection from the input spaces to the counterfactual blend is highly selective, and that considerable emergent structure develops in the blend. For example, we do not interpret the counterfactual assertion as meaning that the seat is as short as your mother's arms, or that your mother could hold you comfortably for twelve hours. We do not interpret it as meaning that a passenger in a British Airways cradle seat must wear a diaper, or that infants are in fact permitted to use these cradle seats.
Counterfactual reasoning, which arises by virtue of conceptual integration, is an indispensable element of political reasoning. Consider a prototypical counterfactual assertion:
--If Churchill had been prime minister in 1938 instead of Neville Chamberlain, Hitler would have been deposed and World War II averted.This counterfactual assertion asks us to blend conceptual structure from different mental spaces to create a separate, counterfactual mental space. The input spaces include (1) Churchill in 1938 as outspoken opponent of Germany; and (2) Neville Chamberlain in 1938 as prime minister facing the threat from Germany. To construct the blend, we project parts of each of these spaces to it, and develop emergent structure there.
From the first mental space, the blend takes Churchill. From the second mental space, the blend takes the role prime minister. In the blend, Churchill is prime minister by 1938. The blend is contrary-to-fact with respect to both of its input spaces. The antecedent and the consequent exist in the blended space; neither exists in either of the input spaces.
Because the process of blending is largely unconscious, it seems easy, but it is in fact complex. It has many standard features that can be illustrated from the Churchill example.
Blends exploit and develop counterpart connections between input spaces. The space with Churchill and the space with Chamberlain share many identity-counter parts, such as date, England, Germany, Hitler, and tension. Churchill and Chamberlain are additionally frame-counterparts: each is an English political figure, holding a certain political office, with views about Germany.
Counterparts may or may not both be brought into the blend, and may or may not be fused in the blend. Many paired counterparts are brought into the blend as fused units: Hitler in the blend is a single fused entity corresponding to Hitler in each of the inputs but not equal to them--the Hitler in the blend has a different life. Churchill is brought into the blend but not fused with his frame-counterpart, Chamberlain. Chamberlain's political office is brought in but not fused with its frame-counterpart.
The projection from the input spaces is selective. The blend takes from the space with Churchill his opposition to Germany but not his political office or his reputation as having poor judgment of the sort that would prevent him from obtaining a position of leadership. The blend takes from the space with Chamberlain the role prime minister and the situation faced by the prime minister in 1938, but not Chamberlain himself or the default knowledge attached to prime minister that world leaders facing aggression are concerned greatly to avoid unnecessary war. We frame Chamberlain according to this default knowledge but keep it out of the blend, where we need a prime minister who views conflict as inevitable.
Blends recruit a great range of conceptual structure and knowledge without our recognizing it. Very little of the structure needed for the contrary-to-fact blended space is mentioned. The Churchill blend recruits conceptual frames of world leaders, political aggression, and wars. It recruits the relevant history of Germany and England. These recruitments are needed for the reasoning to work properly in the blend. Academic theories may also be recruited to the blend--game-theoretic interaction during political aggression, or deterrence by "power-maximizing" actors. These recruitments may drive the elaboration of the blend in one direction or another.
Blending is a process that can be applied repeatedly, and blends themselves can be inputs to other blends. Someone might respond to the Churchill counterfactual, "That's only because Hitler was irrational: a more rational Hitler would have seen that his strategic chances were still excellent, and would not have backed down." This new counterfactual blend takes part but not all of the original Churchill blend, and additionally takes part but not all of the characteristics of Hitler from spaces that refer to actual situations. In the new counterfactual hyper-blend, World War II is not averted.
Former prime minister Margaret Thatcher created just such a hyper-blend when she argued that, as leaders of Britain, France, and the United States should have refused to appease Hitler, so Western leaders should refuse to appease aggressors in the war in Bosnia. Thatcher asked members of her audience to take two spaces--the space referring to the situation in Bosnia and the counterfactual space in which Hitler was opposed and the atrocities were averted--as inputs to the construction of a third, blended space in which the Western leaders oppose the aggressors in Bosnia and atrocities are thereby averted. Her policy--"Not Again!"--is anchored in what she takes to be the persuasiveness of the original counterfactual.
Of course, Thatcher implicitly invited her audience to imagine the counterfactual blend in which Margaret Thatcher is still prime minister during the period in which war breaks out in Bosnia, and the further counterfactual blend in which Margaret Thatcher is prime minister in 1938 and opposes Hitler. In both of these counterfactual blends, the aggressors back down and the atrocities are averted or ended. These two counterfactual blends can be made stronger if they receive projections from the space that contains (one view of) the "Falklands" war, in which Margaret Thatcher is prime minister, "The Iron Lady," war victor, staunch in her defense of honor regardless of the considerable practical difficulties, courageous adversary of aggressors, enforcer of Britain's policy over vast geographical distances. Thatcher not need refer to the "Falklands" space; her identity evokes it.
The "Falklands" space and the two counterfactual blends in which Thatcher faces down (a) Hitler and (b) the aggressors in Bosnia are available to serve as reinforcing inputs to the "Never Again!" space, which represents Thatcher's policy toward Bosnia. Projections of this sort demonstrate the remarkable way in which character--once it has been connected to a specific actor in a space that has reference--can be projected to blends in which that specific actor faces past or hypothetical situations. In fact, character can be projected to other counterfactual blends having to do with what other actors might do or might have done if they possessed that character.
Blends develop structure not provided by the inputs. Typically, the blend is not a simple cut-and-paste reassembly of elements to be found in the input spaces but instead resembles what Kahneman (1995) calls a "mental simulation," in which it develops considerable emergent structure. Usually, we focus on this additional emergent structure. For example, in the blend, but not in any of its inputs, Hitler backs down and World War II is averted.
Inferences, arguments, and ideas developed in the blend can have effect in cognition, leading us to modify the initial inputs and to change our view of the corresponding situations. A student of historicist patterns that led to World War II might know Churchill's personality well but not have brought what she knows to bear on her conception of appeasement in 1938. The Churchill blend might challenge her to reconsider the causal weight of personality. Thatcher's blend might lead someone to rethink the situation in Bosnia and even to choose to intervene.
Selectivity of projection and variability of recruitment can lead to different constructions and inferences. We saw in the British Airways blend a classic case in which activating different input spaces for the blend led to different reasoning and differ ent choices. Many people, hearing the "Never Again!" blend proposed by Thatcher, which asks us to compose a scene in which Western powers intervene in a distant country, will complete that blend with structure from a "Vietnam" frame. Thatcher's blend will then include disaster for the intervening Western powers.2 Thatcher's blend is meant to lead people to reason to a political choice, but it can lead them to reason to the opposite political choice if it is completed and elaborated differently.
How does structure develop in a counterfactual blend? How does structure developed in a blend lead us to reconsider input spaces?
Blends develop by three mechanisms: composition, completion, and elaboration. We selectively compose structure from input spaces into the blend. To do so, we exploit counterpart connections between the input spaces. Partial composition provides a working space for further composition. Completion provides additional structure once a few elements have been brought in. A minimal framing of Churchill and Hitler as adversarial heads of state invites us to complete that structure by recruiting any amount of specific or general knowledge we have about personal opposition, international relations, negotiation, and so on. Elaboration develops the blend through imaginative mental simulation according to principles and logic in the blend. Some of these principles will have been brought to the blend by completion. Continued dynamic completion can recruit new principles and logic during elaboration. But new principles and logic may also arise through elaboration itself.
Composition and completion often draw together conceptual structures usually kept apart. As a consequence, the blend can reveal latent contradictions and coherences between previously separated elements. It can show us problems and lacunae in what we had previously taken for granted. It can equally show us unrecognized strengths and complementarity. In this way, blends yield insight into the conceptual structures from which they arise.
Composition, completion, and elaboration all recruit selectively from our most favored patterns of knowing and thinking. Consequently, blending is very powerful, but also highly subject to bias. It is hard to evaluate bias in blends, for two reasons. First, composition, completion, and elaboration operate for the most part automatically and below the horizon of conscious observation. Therefore, we rarely detect consciously the infrastructure in the blend that makes it effective. Second, since the emergent structure in the blend comes from our favored patterns of knowing and thinking, we are likely to regard biased infrastructure in the blend as unobjectionable even if we somehow manage to detect it.
For example, in trying to reason about a blend only on the basis of its
proper historical structure, we may unwittingly complete the blend with
evidence from a later historical moment. In the Churchill counterfactual,
we use what we know of 1938. But once we have Churchill as prime minister
in the blend, it is impossible to prevent completion from another (covert)
input space--Churchill as prime minister later in time. The
counterfactual blend in which Churchill opposes Hitler in 1938 is plausible
only because we can recruit to it Churchill's great fighting spirit in
opposing Hitler during World War II, and we can know of that spirit only
because World War II was not averted. Our reasoning in the blend--that
World War II might have been averted--depends, therefore, on the
non-occurrence of World War II: the blend makes sense to us only because it
did not happen. In this way, our ex post knowledge can affect our supposed
ex ante reasoning in ways detectable only on analysis. Even the selection of
objects of ex ante reasoning can be influenced by ex post knowledge: Had
Churchill never been prime minister, it unlikely that we would think of
constructing a blend in which he was prime minister in
Ex post input
spaces seep into ex ante counterfactual blended spaces, and in fact
prompt us covertly to construct them.
The Fundamental Problem of Causal Inference in the Social Sciences is Conceptual Integration
As we have seen, counterfactual reasoning arises by means of the essential cognitive operation of conceptual integration, which is much more complicated and systematic than is suspected, and whose operation occurs for the most part below the horizon of conscious observation. Counterfactual reasoning is part of the routine mental operation of political agents and political scientists. Its scope is much wider than might be apparent--counterfactual reasoning as a basis for choice is only very rarely flagged by an "if-then" linguistic form whose antecedent has a verb in the past perfect. Grammatical hints of counterfactual thinking are often subtle. As Gilles Fauconnier (1994) observes, they include subjunctivity, main verbs like wish or prevent, modal verbs like could, and adverbs like not. Linguistic form often provides no test at all, since the same form can express counterfactuality in one case and not in another--"The President could hold your opinion, instead of the one he takes" evokes a counterfactual blend; "The President could hold your opinion. Have you asked him?" does not.
In fact, there seems to be no form of causal inference in the social sciences that does not depend upon counterfactual reasoning and hence upon conceptual integration and blending. Gary King, Robert O. Keohane, and Sidney Verba, in their highly influential Designing Social Inquiry: Scientific Inference in Qualitative Research (1994), provide a definition of causality in the social sciences in a section titled "Defining Causality" (pages 77-79). King, Keohane, and Verba instruct us to imagine an election in a specific congressional district: the Democratic incumbent runs against one Republican (nonincumbent) challenger and receives fraction x of the vote. We are to "imagine that we go back in time to the start of the election campaign and everything remains the same, except that the Democratic incumbent decides not to run for re-election and the Democratic Party nominates another candidate," who receives fraction y of the vote. King, Keohane, and Verba define the causal effect (in this case, of the incumbency of the Democratic nominee) as the quantity x-y. They write:
This counterfactual condition is the essence behind this definition of causality, and the difference between the actual vote [x] and the likely vote in the counterfactual situation [y] is the causal effect . . .
Of course, this effect is defined only in theory since in any one real election we might observe either [x] or [y] or neither, but never both. Thus, this simple definition of causality demonstrates that we can never hope to know a causal effect for certain. Holland (1986) refers to this as the fundamental problem of causal inference, and it is indeed a fundamental problem since no matter how perfect the research design, no matter how much data we collect, no matter how perceptive the observers, no matter how diligent the research assistants, and no matter how much experimental control we have, we will never know a causal inference for certain.
In some sciences, it is possible to design two situations that vary only on the independent variable, to run the two situations experimentally, and to contrast the effects in the two cases. For example, one lab rat (the control) may receive a normal diet while a genetically identical lab rat receives a diet in which fructose has been replaced by glucose. In the social sciences, and especially in reasoning about large political, historical, or economic events, it is nearly always impossible to run two such situations as an experiment. The methodological substitute for doing so is to contrast the actual situation with an imagined counterfactual situation.
To prevent imagination from running wild in the construction of these counterfactual situations, methodologists have proposed, for example, that counterfactual spaces must be thoroughly specified in their details, especially the details of their causal antecedents and consequents; that they must be consistent with well-established historical facts (e.g., they should require minimal rewriting of the original input spaces; the antecedent should have been recognized historically as an option and should additionally have been possible to bring about; and the antecedents and consequences should be close in time); that they must be consistent with well-established theoretical laws and statistical generalizations; that they must be parsimonious (explain as much as possible with as few assumptions as possible); and so on. A survey can be found in Philip Tetlock and Aaron Belkin's introduction to their edited volume, Counterfactual Thought Experiments in World Politics (1996). These criteria have been developed in the hope of using counterfactual spaces as a substitute for controlled laboratory experiments.
The great danger for the social scientist here is to assume that we already understand how counterfactual reasoning works merely because we are so good at it, or, worse, to assume that the mechanisms of counterfactual reasoning are visible and open to conscious manipulation and control. The danger is easily overlooked: King, Keohane, and Verba, for example, nonchalantly slip into their description of the counterfactual blend a requirement that "everything remains the same, except" for a single explicit change. But research on conceptual integration shows that in general the counterfactual space is not a copy of one of its inputs (in this case, the real election) with one explicit change. On the contrary, the projection to the counterfactual space is typically highly selective. Moreover, the counterfactual space develops emergent structure. Even if it is argued that some counterfactual spaces can be found that are copies of an input space with one explicit change, this is a shaky criterion to propose for counterfactual reasoning and hence causal reasoning in general.
Conceptual integration in reasoning and choice is useful, indispensable,
systematic, and intricate, but it does not in general accord with the
criteria summarized by Tetlock and Belkin or required by King, Keohane, and
Verba. It is largely invisible to the kind of conscious management implied
by such criteria. For example, the cognitive scientist will be greatly
suspicious that strong biases influence which counterfactual blends spring
to life to begin with. We invent one counterfactual blend but not another.
These biases on invention are of the first order of importance but typically
unconscious. They may easily swamp the biases that arise during the
construction of the details inside any particular counterfactual blend.
But these biases on invention are missed by methodological controls dedicated
to regulating the internal structure of the blend once it has arisen. I
offer this oversight as an example of the kind of blindness that may arise
in social science from inattention to the nature of cognitive operations.
The study of interdependent decision-making has various names--game theory, rational choice theory, theory of rational expectations, positive political theory, the new institutionalism, economic theory of politics, and so on. It attempts to extract formal aspects of how actors make reasoned, strategic decisions on the goal of maximizing their profit, or, more accurately, their subjective expected utility.
A standard example of a game is "The Battle of the Bismarck Sea." An admiral must transport troops by boat. He can select the shorter northern route or the longer southern route. An enemy admiral must try to bomb them. He can send his planes either north or south to look for the targets. (Many restrictions are quietly assumed; we are not to ask whether the enemy admiral might send half of his planes in each direction, and so on.) There are exactly four sets of possible choices, since each player has two choices. Values are assigned for each player for each of the four outcomes. We then inspect the mathematics of the game tree to see which choice is better for each player.
This extremely simple example already has most of the features of game-theoretic analysis. The decision-making is interdependent and the decision-makers act on their decisions. Various hypothetical sequences of actions have various outcomes and these outcomes are assigned values for each player. In simple examples of this sort, these values might be given in dollar amounts or percentages of market share or years in jail or number of fatalities or some other unit, but in principle all such values are really subjective values assigned by the players to the outcomes, relative to other outcomes. Players are always trying to enhance what they perceive to be their welfare by achieving higher subjective relative utilities. The values are arrayed into a matrix of payoffs of outcomes for players. Game theory is the study of arithmetic conditions that arise in the tree of nodes leading to the outcome numbers, on the view that the arithmetic drives or should drive decision-making at each node.
Inevitably, game theory has sustained attack for making simplistic assumptions. For the most part, game theorists appear to be well aware that these games are cartoons of reality--"parsimonious models" rather than "thick descriptions." Not without reason, they are proud that their mathematically encrypted but conceptually simple analyses nonetheless show something about interdependent decision-making. Something is already a lot in the analysis of reason and choice.
What I offer is not an attack on game theory but a demonstration of how it might be made less implausible and more adequate by being integrated with cognitive science. In 1986, Herbert A. Simon published an essay--"Rationality in Psychology and Economics"--whose spirit I will adopt. Simon describes my principal assumption: "economic actors use the same basic processes in making their decisions as have been observed in other human cognitive activities . . ." (page 39). Simon describes my basic view: "The emerging laws of procedural rationality have much more the complexity of molecular biology than the simplicity of classical mechanics. As a consequence, they call for a very high ratio of empirical investigation to theory building. They require painstaking factual study of the decision-making process itself" (page 39). Simon states my prescription:
Economics without psychological and sociological research to determine the givens of the decision-making situation, the focus of attention, the problem representation [i.e., the framing of the problem in context], and the processes used to identify alternatives, estimate consequences, and choose among possibilities--such economics is a one-bladed scissors. Let us replace it with an instrument capable of cutting through our ignorance about rational human behavior. (pages 39-40).
Eleven years after he offered it, Simon's prescription is apt for all "positive" social science. I hope to follow his recommendation by offering an example of what this new instrument would look like. There are many avenues that could be taken to a "cognitive behavioral game theory." I begin with a few that I will not pursue:
Actual human choice typically requires elaborate and complex cognitive work of sorts that game theory does not at present deal with. Attempts to introduce cognitive research to economics and political science include not only Simon's work but also Kahneman and Tversky (1979), Kahneman, Slovic, and Tversky (1982), and Richard Thaler (1991). Reviews of the divide between economics and cognitive science can be found in Hogarth and Reder (1987), Smith (1991), and Lewin (1996). Game theory has made rudimentary attempts to acknowledge some of this complexity. For example, a preliminary attempt has been made to recognize some gross distinctions in the extent and kinds of knowledge about the game possessed by the players (games of perfect, certain, symmetric,or complete information). An attempt has also been made to recognize the role of conceptual focal points (e.g., the number 100 rather than 99) in decision-making. Nonetheless, as Lewin (1996) summarizes, "the 'declaration of independence' from psychology remain[s], and it haunts economics to this day" (page 1295).
It turns out that conceptual integration is a basic instrument of decision-making, as off-the-equilibrium-path nodes demonstrate. Suppose that actor A believes that, for the past, present, or future, if A does P, actor B responds by doing Q; desiring to avoid Q, A has never done P and plans never to do P; the node S at which A does P and B does Q is put together cognitively by conceptual blending of knowledge and frames. It must be assembled in this fashion; it does not refer to any actual situation, and no one has any memory of it. To arrive at a conception of that space, we must blend together concepts of real actors (A and B), real characteristics of those actors, unreal actions (P), hypothetical responses (Q), and models of behavior. Importantly, this blended space is causal for the space of A's actual behavior: under rationality, this space causes its own counterfactuality.4
In fact, all nodes in a game must be constructed cognitively by conceptual integration, which is easiest to do for very simple games with two players, each of whom has a single move, and where the outcomes are relatively clear since some external agency imposes or provides payoffs in a single, monolithic, publicly recognized unit (yen, years in jail) and at a level that overshadows all other payoffs. It becomes harder to do the requisite conceptual integration as the game becomes more complicated, as in, for example, a highly iterated non-constant sum game without adominant strategy equilibrium but with many Nash equilibria.
Values must be assigned to these outcomes. This assignment of values also arises through conceptual integration. Game-theoretic analysis begins by taking those values as given. In this way, the game theorist takes the payoff matrix as an oracle, in the technical sense: in mathematics, one way to attack a problem is to try to show that if there were an oracle that could supply the answer to some part of the problem, we could solve the entire problem; this shows that the larger problem reduces to the problem solved by the oracle.
But in reality, the payoff matrix does not come from an oracle. Values attached to outcomes are not something outside of decision-making; they come exactly from decision-making, and the assignment of values is typically dynamic and shifting during action.
The player must assign the comprehensive value of an outcome by conceptual integration over all the relevant values, and this integration is unlikely to be linear. Suppose, for example, that Sue believes that if she has a chance to tell a certain story in casual conversation in a group that includes Max, it will prompt Max to begin to court her. She believes the same of Joe. She welcomes the courtship in either case. The chance arises, but Sue stays silent, because Max and Joe are both present, and she prefers not to induce known rivalry between them. The value of having Max and Joe court her is then not the sum of having each court her. There is essential emergent structure in this blend of Max-and-Joe-as-rivals, and it includes an emergent value that is not the linear sum of the values of its inputs.
But now suppose that Todd is also present. Sue thinks Todd's reaction to seeing others court her will be, first, painful brooding, and, second, a determination to court her, which she welcomes highly. Then the high positive value of being courted by Max and Joe is unrelated to the positive values of being courted by Max or Joe.
The game theorist takes the conceptual content of the nodes and the matrix of values as given, and begins the game-theoretic analysis after this point. But from the cognitive perspective, these things are not given; they are instead constructed mentally by conceptual integration and other basic cognitive operations.
The game theorist proceeds in the sequence: (1) chart, or if that is impractical, characterize formally the tree of decisions and actions; (2) assign values; (3) analyze the resulting arithmetic of the game tree. There may be a few cases of decision-making, like the Battle of the Bismarck Sea, where this procedure fits actual decision-making. But in actual decision-making, conceptual integration to assemble the nodes and values is always active. The contents and values of all the nodes are always under construction.
Conceptual integration is additionally indispensable to game theory in a quite different way. It is the basis of deciding, at the specific level, what game one is playing, and, at the general level, whether or not one is playing a game at all. First, the specific level. We use conceptual frames (e.g., seduction, debate) in understanding specific situations. The conceptual frame is part of the integration network that constitutes our understanding of the specific situation. Since conceptual frames carry default inferences and principles of inference, it is natural to assume that decision-making in any specific situation will depend on what frames are used by the decision-maker as conceptual inputs. Yet Von Neumann and Morgenstern's Theory of Games and Economic Behavior (1944 ), which launched the application of game theory to economic behavior, offered as one of its great insights that certain essential mathematical regularities can be found in the formal structure of decision-making that are independent of specific conceptual frames. Some cognitive scientists have argued that at least in certain specific cases this assumption is demonstrably wrong (Simon 1978 and 1982, Kahneman and Tversky 1979, Tversky and Kahneman 1986). If it is wrong as a general principle, then it follows that conceptual integration (of frames and specific situations) is indispensable to decision-making. "Positive" social science depends on cognitive mechanisms of framing.
Second, the general level. Conceptual integration is involved in deciding whether or not one is playing a game. Interdependent decision-making on the goal of maximizing my utility is itself a (very general) conceptual frame, which brings with it pressures on decision-making. It can be used as an input to an integration network that constitutes the (not necessarily conscious) understanding of a specific situation, but there are alternative general frames. Contra Bentham, even the more general conceptual frame of seeking pleasure and avoiding pain does not seem to be obligatory as the governing input to all integration networks that serve as the basis of decisions. Certainly, people seem to try to reason toward welfare-enhancing decisions when they frame themselves that way--especially if they use a specific frame that carries maximizing utility as part of its stipulated goal structure (as in "Matching Pennies" or "poker," two of Von Neumann and Morgenstern's examples). But it is not easy for the cognitive scientist to swallow the assumption that people have no choice but to blend that abstract frame of game-playing with their current specific situation, all of the time, and always to act on the central inference of that blend.
There appears to be a feeling that evolutionary pressure by itself (succeed rather than fail) must ensure that actors try to enhance their welfare by maximizing their expected utility, but the mechanisms of evolutionary pressure are not simple. For example, curiosity and routine action to satisfy it are apparently adaptive for our species, but curiosity killed the cat. Acting on curiosity may be unrelated to any local or recognizable game or utility, and the details of its downstream utility may be for the most part unimaginable to the actor and beyond his evaluation. The benefit of curious behavior may be absolute but not resident within anything that looks like a strip of interdependent decision-making. Acting on curiosity does not have to be reasoned, it does not have to involve decision, and it does not have to be connected to any utility that is recognizable to the actor or that is imaginable by the actor. It can be impulsive. Yet much of our action--even our action in explicitly political, financial, or legislative contests--could in principle be driven by curiosity. It would be inappropriate to respond that in this case the actor has placed a high value on satisfying curiosity and is making reasoned decisions accordingly. That claim merely reasserts that the curious actor is reasoning--to both values and decisions--, is making decisions, and understands the connections between the decisions, the actions, and the values, when in fact it is unclear that the curious actor is reasoning to values, is making reasoned decisions at all, or has any model relating these decision to these utilities.
The assumption of rational play is, like the assumption that outcome values
are supplied by oracle, a black box in game theory, one that a cognitive
scientist, at least, would like to see filled.
Blending is an element of political reason which typically escapes notice because it works in the backstage of cognition, as we have seen for counterfactual argument and interdependent decision-making. The introduction given here to the operation and effects of conceptual integration in political reason inevitabley raises advanced issues having to do with the specific mechanisms of conceptual integration, issues which, although too large and complicated to be addressed here, have been explored extensively in Turner (1996), Fauconnier & Turner (in press, a and b), and especially Fauconier and Turner (in preparation). They include:
Cross-space mapping. All blending depends upon cross-space mapping, but absent some constraints, the combinatoric possibilities on cross-space mapping are very large. Guidance in cross-space mapping comes from various sources, including existing mental frames available for the inputs, existing conceptual metaphors, topological principles, and pragmatic influences of goals and context.
Optimality constraints. Conceptual blending is not compositional, algorithmic, or deterministic. Accordingly, it cannot be modeled as such, and blends cannot be predicted exclusively from their inputs. However, there are several definite but competing optimality constraints on blending, with names like topology, integration, web, unpacking, good reason, and metonymy projection. Each of these requires a long discussion, with many examples, to explain, and their interaction is complicated. There are additionally pragmatic constraints and cognitive constraints on blending, having to do with working memory, influence from background knowledge, influence of prior activity, and so on. Although it is in the nature of conceptual integration that deterministic predictions cannot be made about its operation over inputs, especially given that the optimality constraints often conflict, nonetheless, falsifiable "ceterisparibus" claims can and have been made about its operation, including claims about what counts as a good or a bad blend.
Standard Network Types. A blend and its inputs are nodes in a "conceptual integration network." Standard general types of conceptual integration network have developed to provide guidance in solving the specific cognitive difficulties of any particular case of blending. A problem of blending is rarely attacked from scratch; rather, attempts are typically made to take it as an instance of a familiar type of conceptual integration network.
Naturally, each of these operational aspects of conceptual integration comes
with its own further biases and influences for the backstage cognition of
Bryant, Adam. "Advertising," The New York Times, Thursday, 6 June 1996, page D7.
Coulson, Seana. 1995. Analogic and metaphoric mapping in blended spaces. Center for Research in Language Newsletter, 9, 1: 2-12.
Fauconnier, Gilles. 1997. Mappings in Thought and Language. Cambridge: Cambridge University Press.
Fauconnier, Gilles. 1994. Mental Spaces: Aspects of Meaning Construction in Natural Language. Second edition. Cambridge: Cambridge University Press, 1994. [First edition. Cambridge: MIT, 1985.]
Fauconnier, Gilles and Mark Turner. 1994. "Conceptual projection and middle spaces," UCSD Cognitive Science Technical Report 9401. San Diego. [Available from http://markturner.org]
____________. 1996. "Blending as a Central Process of Grammar" in Conceptual Structure, Discourse, and Language. Edited by Adele Goldberg. Stanford: Center for the Study of Language and Information.
____________. In press, a. "Conceptual Integration Networks." Cognitive Science.
____________. In press, b. "Principles of Conceptual Integration" in Conceptual Structure, Discourse, and Language, II. Edited by Jean-Pierre Koenig. Stanford: Center for the Study of Language and Information.
____________. In preparation. Conceptual Integration.
Fearon, James D. 1991. "Counterfactuals and Hypothesis Testing in Political Science."World Politics 43:2, 169-195
Freeman, Margaret. 1997. "Grounded spaces: Deictic -self anaphors in the poetry of Emily Dickinson," Language and Literature, 6:1, 7-28.
Grush, Rick & Nili Mandelblit. 1996. "Blending in language, conceptual structure, and the cerebral cortex." Second Conference on Conceptual Structure, Discourse, and Language (Buffalo).
Hogarth, Robin M. and Melvin Reder. 1987. Rational choice: The contrast between economics and psychology. Chicago: University of Chicago Press.
Holland, Paul W. 1986. "Statistics and Causal Inference." Journal of the American Statistical Association 81:396 (December 1986); pages 945-960.
Kahneman, Daniel. 1995. "Varieties of counterfactual thinking" in What Might Have Been: The Social Psychology of Counterfactual Thinking. Edited by Neal J. Roese and James M. Olson. Erlbaum: Hillsdale, NJ.
Kahneman, Daniel and Amos Tversky. 1979. "Prospect Theory: An Analysis of Decision under Risk." Econometrica 47:2 (March), 263-291.
Kahneman, Daniel, Paul Slovic, and Amos Tversky, editors. 1982. Judgment under uncertainty: Heuristics and biases. Cambridge: Cambridge University Press.
King, Gary, Robert O. Keohane, and Sidney Verba. 1994. Designing Social Inquiry: Scientific Inference in Qualitative Research. Princeton University Press.
Lewin, Shira. 1996. "Economics and Psychology: Lessons For Our Own Day From the Early Twentieth Century." Journal of Economic Literature 34 (September), 1293-1323.
Mandelblit, Nili. In press. Blends in Hebrew Causatives. Proceedings of the Fourth Conference of the International Cognitive Linguistics Association. Albuquerque.
_____________. 1995. The theory of blending as part of the general epistemological developments in cognitive science. Manuscript.
Oakley, Todd. 1995. "Presence: the conceptual basis of rhetorical effect." University of Maryland dissertation.
Robert, Adrian. 1996. Blending in the interpretation of mathematical proofs. Second Conference on Conceptual Structure, Discourse, and Language Buffalo.
Simon, Herbert A. 1978. "Rationality as process and product of thought." American Economic Review: Proceedings 68:1-16.Reprinted in Simon (1982).
Simon, Herbert A. 1982. Models of Bounded Rationality. 2 volumes. Cambridge: MIT Press.
Simon, Herbert A. 1986. "Rationality in Economics and Psychology." The Journal of Business 59: 4,2; reprinted as Hogarth and Reder (1997), pages 25-40.
Smith, Vernon L. 1991. "Rational Choice: The Contrast between Economics and Psychology." Journal of Political Economy 99:4 (August), 877-897.
Sun, Douglas. 1994. "Thurber's Fables for our Time: A Case Study in Satirical Use of the Great Chain Metaphor" in Studies in American Humor, new series volume 3, number 1, pages 51-61.
Tetlock, Philip and Aaron Belkin, editors. 1996. Counterfactual Thought Experiments in World Politics. Princeton: Princeton University Press.
Thaler, Richard. 1991. Quasi rational economics. New York: Russel Sage Foundation.
Turner, Mark. 1996a. "Conceptual Blending and Counterfactual Argument in the Social and Behavioral Sciences," Philip Tetlock and Aaron Belkin, editors, Counterfactual Thought Experiments in World Politics. Princeton, N.J.: Princeton University Press. pages 291-295.
Turner, Mark. 1996b. The Literary Mind. New York: Oxford University Press.
Turner, Mark. 1991. Reading Minds: The Study of English in the Age of Cognitive Science. Princeton: Princeton University Press.
Turner, Mark. 1987. Death is the Mother of Beauty: Mind, Metaphor, Criticism. Chicago: University of Chicago Press.
Turner, Mark. 1989. "Categories and Analogies" in Analogical Reasoning: Perspectives of Artificial Intelligence, Cognitive Science, and Philosophy. Edited by David Helman. Dordrecht: Kluwer, 3-24.
Turner, Mark and Gilles Fauconnier. 1995. "Conceptual Integration and Formal Expression." Metaphor and Symbolic Activity. 10:3, 183-203.
____________. In press. "Conceptual Integration in Counterfactuals" in Conceptual Structure, Discourse, and Language, II. Edited by Jean-Pierre Koenig. Stanford: Center for the Study of Language and Information.
Tversky, Amos and Daniel Kahneman. 1986. "Rational Choice and the Framing of Decisions." The Journal of Business 59: 4,2; reprinted as Hogarth and Reder (1987), pages 67-94.
Veale, Tony. 1996. "Pastiche: A Metaphor-centred Computational
Model of Conceptual Blending, with special reference to
Cinematic Borrowing." Manuscript. Available from the Conceptual
Integration Web Page:
This study was conducted while the author was Agnes Gund
Member of the School of Social Science, Institute for Advanced Study,
Princeton, in 1996 and 1997. He is grateful for support provided by
the Institute, the School, and Agnes Gund.
I thank Mat McCubbins for this observation.
3 I thank Bruce Bueno de Mesquita for conversation on thispoint.
I thank Bruce Bueno de Mesquita and Bary Weingast for
illumination on this point.