Implicit Bias and the Unconscious

Ege Yumusak
Harvard University


There is a desperation in all certainty.
William Kentridge


Implicit biases are largely automatic and unconscious evaluations that people possess. Implicit biases regarding race, for example, influence what their possessor will think when they see a Black man inside their neighbor’s house, or how they will position their body when there is an empty seat next to a Black stranger on the New York subway.[1] Contemporary philosophical attention has focused on implicit biases that are cognitive manifestations of prejudices against social groups like racism, sexism, or ageism. Attention to this subset is well-deserved, as relevant empirical data have been profoundly disconcerting. Studies have shown, in scenarios ranging from evaluation of job applications to treatment recommendations by physicians, that explicit views against such prejudices neither exempt subjects from possessing implicit biases, nor from acting on them. [2]

In this paper, I will argue that current prevalent views of implicit bias do not satisfyingly explain the pervasive and striking manifestations of implicit biases in ordinary life. This will be our starting point. The overall goal of this paper is to lay out the shortcomings of two existing accounts of implicit bias, and attempt an account of implicit biases’ contents that is commensurate with its conscious manifestations. For the sake of consistency, I will exclusively draw on psychological studies that target implicit biases regarding race.

Here is our roadmap: In §1 I will propose an analogy to illuminate the conscious manifestations of implicit biases. In §2 I will argue that the two prevalent understandings of implicit bias are misguided; in §3–4 I will identify the culprit, and in §5 I will propose an alternative. In the final section I will explore the moral implications of this analysis.


Target presentation by Ege Yumusak (Harvard)

1 Implicit bias in the conscious realm

The term ‘implicit’ in ‘implicit bias’ identifies a characteristic of the target mental state. Scientists disagree about whether the relevant characteristic is lack of control or lack of introspective awareness.[3] Recently it has been suggested that ‘implicit’ tracks neither, but the further characteristic of arationality.[4] Invariably, however, scientists have employed empirical measures that block the subject’s explicit reasoning processes either by a cognitive load or by time pressure. In the most popular measure of implicit bias, the Implicit Association Test (IAT), subjects are asked, under time pressure, to sort words (or images) into paired categories on either side of a computer screen.[5] A target word can be the racialized name ‘Darnell’, with the racial categories ‘African American’ and ‘European American’ on either side of the screen, and evaluative categories ‘Good’ and ‘Bad’ on either side, below the racial categories. The subject presses one of the two designated buttons on their keyboard to match the target to the category. Their reaction time and error rate are then used to determine the strength of association between the paired concepts (those concepts that appeared on the same side of the screen). When ‘African American’ and ‘Good’ are paired, a subject with a racial implicit bias against Blacks is expected to react slower and make more errors.

People are typically unaware of the source of their implicit biases, but this does not mean that implicit biases have no conscious life in their mind.[6] To the contrary, possessors of such biases are conscious of implicit biases’ effects on their behaviors, perceptions, and judgments, and they can become reflectively aware of them. Jesse Jackson expresses the shock of such a moment of reflective awareness:

There is nothing more painful to me at this stage in my life than to walk down the street and hear footsteps and start thinking about robbery. Then look around and see somebody White and feel relieved.[7]

As evinced by Jackson’s anecdote, the contents of implicit bias might be lodged in the unconscious; nevertheless, they impact on the subject’s conscious experience. The identifying characteristic of implicit bias is its possible divergence from self-reported explicit views and activity despite opposing conscious attitudes of the subject. In my definition of ‘implicit’, I will understand ‘lack of control’ and ‘lack of awareness’ in a connected way. For our purposes, ‘implicit’ in ‘implicit bias’ will refer to mental states that give rise to behaviors whose causal origins subject lack awareness of, and as such, under the subject’s direct control.[8]

If the empirical results from IAT and other measures merely occurred in the experiment room, we would not be so worried, but these biases are pervasive. Their varied effects shape the conscious reality of the subject in the social world. To elucidate these effects and motivate the revisionist project I am after, I invite the reader to consider implicit biases as a kind of mood.[9]

The folk psychological category of moods picks out a heterogeneous phenomenon. Moods are deemed positive or negative, but it is very difficult to precisify them further. Instead, we describe moods by their effects and their potential causal stories, which are often inferred from these effects. These identifying characteristics are reminiscent of my characterization of ‘implicit’ in ‘implicit bias’. Remember, what makes a bias implicit in my definition is the subject’s unawareness of the source, and their inability to control its effects. Moods, likewise, are consciously experienced even though their source is often unknown. And, although they are defeasible to a certain extent, they largely escape the subject’s control.

Consider this example of a mood:

Effi, newly married, moves into a tiny town in Prussia, called Kessin. Her much older husband travels days at a time, and in her isolation Effi sinks into a mood marked by depression and fear. The house looks dimmer to her than the day she entered it. Shadows come to look like ghosts. She loses interest in writing to her mother and longs to go back to her hometown, Hohen-Cremmen.

Effi’s mood affects her experience in two main ways: by changing the salience of certain elements in the environment (outside or mental) and by imbuing these elements with a valence. For example, the salience of shadows increases in her experience, while the salience of her mother decreases. The mood imbues Kessin with a negative valence and Hohen-Cremmen with a positive one. Also note that Effi’s mood is not constant: she can briefly escape her mood when her friend Roswitha makes a joke. The effects of her mood are, as such, dependent on the context she is in.

Compare this to implicit bias. Similarly, Jackson’s implicit bias modifies the salience and valence of elements in his environment. It colors his experience by making him alert to certain features of the environment and facilitating certain transitions of thought. He becomes attuned to the sounds of footsteps and is gripped by the thought of robbery. This bias also displays the relevant sensitivity to context. The anecdote takes place on an empty street, a setting in which one might plausibly be confronted with their implicit bias, as opposed to, say, a Black Lives Matter rally.

Empirical measures of implicit bias study the saliencing and valencing carried out by implicit biases. Studies have shown that subjects look into the eyes of a Black interviewer for shorter periods of time, and sit further from their Black study partner.[10] When primed with pictures of Blacks, subjects become more prone to misidentifying hand-tools as handguns.[11] These studies make clear that implicit biases affect how things feel and seem to the subject.

Implicit biases’ context-sensitivity has also been documented. Researchers found that subjects’ IAT scores were differentially affected by Black primes depending on whether, in the pictures presented, Black individuals were shown to be in a church environment or an urban environment.[12] The church prime alleviated implicit bias, whereas the urban prime hardened the subjects’ implicit biases. Contextual variation includes variation of subjects’ own emotional states. Different drugs have been found to boost in-group favoritism or increase negative bias towards certain groups as reflected on IAT results.[13]

These diverse measures have been shown to be internally reliable. Yet, many studies have failed to establish cross-correlations. Importantly, affective measures (e.g. seating distance) do not correlate with semantic measures (e.g. peer competence evaluation).[14] In response to this peculiar finding, some theorists have recently argued for a radical heterogeneity of both function and content of implicit biases.[15] They suggest that we cannot give a unified account of implicit bias. In this paper, I take an alternative route by proposing a new account. But before I get there, I will review two prevalent accounts of implicit bias and their shortcomings.


2 The associative content view and the propositional attitude view

Recent views of implicit bias take them to be either unconscious associations or unconscious propositional attitudes. These two prevalent views posit different entities to explain the contents of implicit biases:

The associative content view: The unconscious mind responds to regularities in the world by creating associative links between concepts; these associative structures make up the contents of implicit bias.[16]

The propositional attitude view: Implicit biases are belief-like propositional attitudes in the subject’s unconscious.[17]

These hypotheses disagree about what kinds of mental states undergird the behaviors observed in these experiments. The associative content view states that associations are responsible for bringing about the biased behaviors of the subjects, while the propositional attitude view states that unconscious propositional attitudes produce such behaviors. Importantly, these hypotheses are not about how implicit biases are learned. The acquisition of bias is a separate issue—and orthogonal to the matter at hand.

The propositional attitude view is motivated by compelling evidence that implicit biases do not behave like associations. In particular, the propositional attitude proponent holds that these biases are responsive to reasons: that they partake in inferential processes and are revised by evidence. This is contrasted with associations, which enter associative transitions where the activation of one element conjures the other without being facilitated by an inference.

These results have been compiled and discussed by Eric Mandelbaum, whose account I will be taking as representative.[18] Mandelbaum’s account is a strong brand of the propositional attitude view that identifies implicit biases with bona fide beliefs; other accounts have nominated other belief-like propositional attitudes.[19] Although my criticisms will be most vivid when applied to the strong version of the propositional attitude view, they apply to weaker versions as well.

The propositional attitude proponent maintains: for one, procedures that disrupt perceived regularities do not efficaciously overcome manifestations of implicit bias.[20] The associative content view would require that implicit biases are extinguishable by such procedures, unless biases were permanent once learned. Associations, after all, depend on the spatiotemporal contiguity of stimuli, and are stored in memory when a spatiotemporal contiguity is perceived by the subject.

Moreover, implicit biases are implicated by argumentation, suggesting that the contents of implicit bias undergo propositional processing. I will focus on the last one of these findings—that the contents of implicit bias are sensitive to argumentation—to elucidate the propositional attitude proponent’s case for the reasons-responsiveness of implicit bias.

Evidence for the reasons-responsiveness comes from an influential study, where subjects were given arguments regarding a new university policy to employ more Black professors.[21] Subjects were then tested using two variables: weak/strong arguments and high/low elaboration conditions. A group of subjects were given a weak argument for the new policy, e.g. suggesting that this is a national trend; whereas another group of subjects received a strong argument, e.g. that there will be a subsequent increase in the quality of teaching.[22] These two groups were further divided into two elaboration conditions. In the high elaboration condition, subjects were told that they were part of a small group of students asked to participate in a survey that was going to lead to a policy change at their own university; while in the low elaboration condition, the subjects were told that they were part of a large group, and the policy change concerned a remote university. These four groups were administered the race IAT. Researchers found that strong arguments most notably affected the subjects’ IAT scores in making them less biased in the high elaboration condition. The IAT scores were not significantly different between the argument groups in the low elaboration condition, leading the researchers to conclude that argument strength only influenced the automatic evaluation of the racial categories when the message was carefully processed by the subject.[23]

Interestingly, Briñol et al. are proponents of the associative content view. In accordance with this allegiance, they explain the above findings by speculating that strong messages trigger “many favorable thoughts associated with the [policy] and Blacks”, whereas the weak argument conjures “many unfavorable thoughts associated with the [policy] and Blacks”.[24] They did not ask their subjects to report their explicit thoughts in this experiment, so this speculation is unsupported by data. Moreover, I find it highly implausible that subjects would react to a weak argument by becoming more biased be it through having unconscious or conscious thoughts about the policy. Prima facie, the propositional attitude view provides a much better explanation for the results, since propositional attitudes can be the constituents of arguments and are sensitive to argument strength. If the propositional attitude view is true, both arguments, weak and strong, will have participated in unconscious inferential processes, and influenced the contents of implicit bias differentially.

Skeptics about implicit biases’ reasons-responsiveness emphasize their recalcitrance to reasons.[25] To them, the recalcitrance lends support to the associative content view. They argue that strong arguments (and the high elaboration condition) merely solicit more motivational resources by making the subject alert during the test. Supposedly, this enables the subjects to exert more control over their actions, while the biases are unperturbed by argumentation.

The proponents of the propositional attitude view confront the problem of recalcitrance repeatedly. In another study, researchers attempted to program an implicit bias regarding two fictional social groups, Niffites and Luupites, by asking their subjects to suppose that one of the groups had negative traits. [26] The problem of recalcitrance emerged in the next stage of the experiment, when the researchers attempted to reverse the biases. In one condition, they told the subjects that there was an experimental mix-up resulting in the reversal of the group identifications. In the other condition, they gave the subjects a vivid narrative revealing that the bad guys were, actually, the good guys. While the subjects’ explicit views on the two groups reversed, subsequent IAT showed that their implicit biases had not reversed. Implicit biases weakened in the first condition, but they were as strong as before in the narrative condition. Their weakening is a promising result for the propositional attitude view, but the non-reversal leaves the proponent of the propositional attitude view with a question: why are implicit biases simultaneously responsive and unresponsive to reasons?

The recalcitrance of implicit biases does not rule out the propositional attitude view.[27] This problem, however, casts doubt on what exactly implicit biases are sensitive to. Some options in the vicinity are linguistic tokens, motivating factors, and normative reasons.[28] More research is needed to expound on the relevant properties of implicit biases like evidence sensitivity and inferential integration. While any semantically evaluable factor lends support to the propositional attitude view as opposed to the associative content view, these findings are used to dispute Mandelbaum’s version of the propositional attitude view that identifies implicit biases with bona fide beliefs.[29]


3 The equivalence assumption

The propositional attitude view fares better in the light of evidence that implicit biases participate in unconscious logical processes. As I have reviewed, the propositional attitude proponent grounds the thesis about propositional content in propositional processing. This view’s weaknesses so far have been explaining the lack of correlation between different measures of implicit bias and its distinctive pattern of recalcitrance and reasons-responsiveness.

Identifying implicit biases with propositional attitudes entails that these biases possess many of the paraphernalia of conscious beliefs: that they are mental representations with propositional contents, and can enter inferences as premises.[30] I suggest that the propositional attitude view, in virtue of this identification, is committed to an unjustified assumption that is quite common in the philosophy of mind. This assumption posits an equivalence between the contents of conscious mental states and the contents of unconscious mental states. I will call it the equivalence assumption for short.

The equivalence assumption: The contents of conscious mental states and the contents of unconscious mental states are equivalent.

Speaking figuratively, the equivalence assumption renders consciousness into a light that shines on a mental representation and makes the representation a conscious mental state. According to this view, mental contents are the same when the light shines on them and when it does not.

In theorizing about implicit bias, we are interested in unconscious mental representations that can be at odds with the subject’s avowed beliefs such that sometimes beliefs associated with these representations are not consciously endorsed. The implicit bias BLACK MALES ARE DANGEROUS, in Mandelbaum’s example, evades the light somehow and when the same propositional attitude is brought to the subject’s consciousness, it is not endorsed by the subject, or even recognized as theirs.[31] Still, according to the propositional attitude view, the implicit bias is realized as an unconscious mental state equivalent to the explicit belief BLACK MALES ARE DANGEROUS. To say that implicit biases are unconscious propositional attitudes is to assume equivalence by way of treating consciousness as a light. I should emphasize, though, that the equivalence assumption does not entail this view of consciousness. And, this view of consciousness is not the only view available. Even so, this view and the equivalence assumption are typically associated.

I maintain that the equivalence assumption is not innocuous. To explicate the problem, let us entertain a particular brand of this assumption: the once-popular language of thought hypothesis (LOTH). LOTH states that the content of thought is embodied in a universal mental language (often referred to as mentalese), and that thought processes have a linguistic structure. Daniel Dennett famously delivered a critique of LOTH in a series of essays that problematize propositional attitude ascriptions. In his paper “A Cure for the Common Code?”, Dennett employs a thought experiment about a chess-playing program. The designer of the program says, “[The program] thinks it should get its queen out early”.[32] Dennett questions this ascription, “But for all the many levels of explicit representation to be found in that program, nowhere is anything roughly synonymous with ‘I should get my queen out early’ explicitly tokened”. Observation of behavior, Dennett contends, does not entail that there be an explicit code for the belief that would describe that behavior. With his remark, the designer is talking about emergent properties of the program that emerge out of computational processes. Dennett concludes, “I see no reason to believe that the relation between belief-talk and psychological-process talk will be any more direct”. According to Dennett, belief-talk delivers predictions of behavior. This is exactly what the designer aims: the designer ascribes a belief to the program in order to predict the program’s behavior. Dennett thinks, however, that it would be naïve to assume that the content of the ascribed belief will map onto the engineering reality. Because, belief-talk alone does not entail that there be a representation that encodes the content of this belief. Something can be true of a system without being explicitly represented to be so.

Dennett’s argument concerns the case of artificial intelligence. Consider other cases of belief ascription. Surely when I hold a belief, I SHOULD GET MY QUEEN OUT EARLY, I can legitimately ascribe this belief to myself. I have this belief in virtue of having a consciously available representation that can be expressed linguistically by the sentence ‘I should get my queen out early’.

Consider now that you are playing a game of chess against your friend. By observing their behavior, you attribute the same belief to them. You might think this ascription is warranted, because it’s just the same for them as it is for you. Perhaps they held a different belief that produced the same behavior, e.g. I OUGHT TO PLAY A HIGH RISK/HIGH REWARD GAME. Still, you would not have been wrong in ascribing to them a belief with propositional content, and in that sense, it would still be just the same for them as it was for you. Here, what warranted you to posit an explicit representation was, again, conscious experience. After all, your inference depended on your assumptions about your friend’s conscious experience—something that a computer program lacks.

Now, consider Dijon the dog. Dijon can’t play chess, but we can imagine another game he can play: fetch. When Dijon fetches the ball, you might think he has the belief I SHOULD BRING THE BALL BACK TO MY OWNER. It is harder to say that it would be just the same for Dijon as it is for you. For one, Dijon’s concepts are likely to differ from yours. Insofar as Dijon seems to function like a less sophisticated version of yourself, you might be willing to assert that his behavior is brought about by a code in propositional form. A non-dog-owner-scientist might also find the belief ascription plausible on the grounds that the engineering reality of a dog is comparable to the engineering reality of a human. Both of these empirical grounds are susceptible to skeptical concerns.

Dennett’s thought experiment elucidates what is at stake in ascribing propositional attitudes to representational systems by observing their behaviors. Even if we reject LOTH, by ascribing propositional content to a system, we affirm that the system explicitly encodes a proposition. Like Dennett, I insist that propositional attitude ascription should not come gratuitously. Contra Dennett, I think that this affirmation is not unfounded in situations where we have epistemic access to the representations in question via conscious awareness or intersubjectivity. I argue that, for precisely the reason that we lack epistemic access to the unconscious, we cannot assume equivalence between conscious and unconscious mental representations.


4 Extradetermination

The equivalence assumption is not only unfounded; but, as will be shown, it also paints a misleading picture of the contents of implicit bias.

Many theorists today reject LOTH. Still, they resort to it as a rhetorical device. In doing so, I think they are compelled by a tendency to make a vague assertion about mental contents more determinate.[33] Here is an example from an earlier paper by Mandelbaum, where he appeals to the LOTH rhetoric to motivate his thesis that “in tokening an idea the subject believes that idea”:

If you are having trouble envisioning the thesis, assume that there is a language of thought (LOT). My thesis is that every time a truth-apt sentence is tokened in one’s LOT, one believes that sentence.[34]

Here, Mandelbaum uses the LOTH rhetoric to visualize what a ‘token of an idea’ would look like. He thinks, rightly, that LOTH makes the insufficiently clear notion of a ‘token of an idea’ more concrete for the reader. Mandelbaum does not hold that ideas have linguistic structure, rather he thinks they are structured propositions. By appealing to LOTH, he means to assuage the confused reader who might wonder just what the theorist means, by giving them the concrete notion of ‘a sentence in one’s LOT’.

I call the act of making content that is deemed vague into something determinate extradetermination. The determinacy invoked here relates to how distinct we want our terms to be. Think of a Seurat painting, A Sunday on La Grande Jatte (You can find a study of the painting below). Looking at the painting from a distance, you will confirm that the color of the grass is green and you will be right insofar as the grass is no other single color. But the magic of a Seurat painting is that there are dots of paint in varied colors in every expanse. For the purposes of the conservator of this painting, ‘green’ extradetermines the color of the field. To her, every point of color in the ‘green’ expanse is very important.

As theorists of implicit bias, we want our theory to capture the dots of paint. I think the propositional attitude view cannot achieve this, and like LOTH, it generates the problem of extradetermination.

Apply Mandelbaum’s thesis to implicit bias: Assume that there is a LOT. Ex hypothesi every time an implicit racist tokens the sentence “Black males are dangerous” in their LOT, they believe it. Thereby, the token of this belief is the translation of “Black males are dangerous” to mentalese. In saying this, we would be committed to implicit bias’s having a very particular structure, that of mentalese. LOTH would thus extradetermine the contents of implicit bias by assuming equivalence between the structure of language and the structure of mental representations.

The propositional attitude proponents (including Mandelbaum) do not hold LOTH regarding conscious or unconscious attitudes. But, they hold that the contents of conscious mental states are propositions. When they encounter data suggesting that implicit biases are reasons-responsive, they reach to conscious contents to find what undergirds implicit biases. The standard picture of propositional attitudes, however, runs into the same problem: like LOTH, it extradetermines the contents of implicit bias.

Consider an analogous claim about the contents of moods: that the underpinnings of moods are propositions. In response to this, one would ask, what proposition inflicts the subject with a bad mood, or underwrites the joyous mood of another? For Effi, would the proposition MY OLD LIFE WAS SO MUCH BETTER make do? Or could it be MY HUSBAND IS HAVING AN AFFAIR? This skeptical tack would be motivated by doubts regarding the ascription of a particular propositional attitude on the basis of a diverse set of behaviors. Propositional attitudes are limited in their explanatory powers. Being determinate, they cannot meet the demands of psychological phenomena that have diverse and context-dependent manifestations. They pose the threat of extradetermination for implicit bias—as they do for moods.


5 The indeterminate content view

In this section, I propose a view of implicit bias that does not rely on the equivalence assumption, and avoids extradetermination by modifying one of the pieces of the puzzle: propositions.[35]

Remember the designer’s comment. In Dennett’s thought experiment, the ascription ‘It thinks it should get its queen out early’ predicted the behavior of the program. Such a prediction, in the form of an ascription, delivers an interpretation for the behavior of the program. Likewise, I suggest that we treat ascriptions of propositional attitudes regarding implicit bias as interpretations.  By doing so, we will regard propositions as models of contents, and not contents themselves. According to this proposal, then, propositions model unconscious contents and thereby deliver an interpretation of these underlying contents. By ascribing a proposition to an implicit bias, we don’t explain what it is, we explain how it will behave.

Note here the key characteristic of models: models deliver interpretations, not translations, which are made between equivalent languages.[36] Even a high-fidelity model is one step removed from reality, because it is an indirect representation of a real-world phenomenon. A translation, on the other hand, is a direct mapping of items in one form to another equivalent form.

I borrow this approach from Tim Crane, who assigns this role to propositions as part of his theory of the unconscious. Crane explains the relation between unconscious contents and the behaviors they give rise to by the following:

[The subject’s unconscious belief system] embodies all the dispositions associated with what [they] believe. So instead of thinking of a single disposition being associated with a distinct propositional content, the propositions which we employ in belief ascriptions are used to model aspects of [the subject’s belief system].[37]

In Crane’s picture, unconscious contents embody dispositions. These dispositions give rise to behavior that is associated with the subject’s unconscious belief system. Propositions are not contents. Instead, they are models for aspects of the subject’s belief-system as a whole. Propositional attitude ascriptions predict the subject’s behavior using these models.

Crane’s suggestion pertains to the unconscious as a whole. Here, my sole focus is on the success of this approach in explaining the unconscious underpinnings of implicit bias. Applying his set-up to implicit bias, I arrive at:

The indeterminate content view: Implicit biases are aspects of the subject’s unconscious belief-system that have particular dispositional profiles. They can be modeled by propositions but their contents are not determined by propositions. [38]

According to the indeterminate content view, propositions are models that interpret the subject’s implicit biases. These biases are aspects of the subject’s unconscious belief-system which embody the relevant dispositions that are pinned as manifestations of implicit bias. Implicit bias’s content is not equivalent to any single proposition; therefore, it is indeterminate.

Unlike the propositional attitude view, in this picture, two propositions might act as equally good models for an implicit bias, or a single, complex proposition can model its content exhaustively. For example, BLACK MALES ARE DANGEROUS is a model for an aspect of the subject’s unconscious belief system. UNKNOWN BLACK MALES ARE DANGEROUS can also be a model for the same aspect. After all, subjects don’t encounter familiar faces in implicit bias studies. The success of either proposition is an empirical question, to be determined by studying the dispositional profile of an implicit bias. The only points of access we have to the unconscious are dispositions that manifest behaviorally and in the subject’s consciousness. Propositional attitude ascriptions, for this view, are not limiting; therefore, they do not extradetermine. The indeterminate content view is superior to the propositional attitude view in this regard.

Furthermore, the simultaneous reasons-responsiveness and recalcitrance is not a problem for this view, because reasons-responsiveness is not a condition that needs to be met by indeterminate contents. It is a property that an aspect of the subject’s unconscious belief-system can have. Propositions are assessable for truth conditions, which adhere to models of the implicit bias and not implicit bias’s contents as a whole.

Lastly, this view accommodates the heterogeneity and lack of constancy of implicit bias—a feat that neither the associative content view nor the propositional attitude view has accomplished. Under the indeterminate content view, different aspects of the subject’s unconscious belief-system can be causally activated in different contexts, giving rise to the context-dependence of implicit bias.

The indeterminate content view gets the order of explanation right. It starts with a whole, the subject’s unconscious belief-system, and then studies behavior generated by this belief-system using the propositions that adhere to its parts as models. The propositional attitude view, on the other hand, asserts that propositions constitute unconscious contents. It starts with these propositions and tries to explain implicit bias as a whole. This is why the lack of correlation between studies poses a problem for this view. The strong propositional attitude theorist, especially, needs to explain how the selfsame belief, BLACK MALES ARE DANGEROUS, undergirds affective and semantic dispositions that do not correlate.[39] The false predictions that this proposition yields motivate skepticism regarding implicit biases; whereas the indeterminate content view is congruous with low predictive success. Often, one size does not fit all: under different contexts, different models are needed to predict behavior.

The indeterminate content view is successful because it explains implicit bias. It achieves this by dispensing with the simplicities inherited from the equivalence assumption: Neither the unconscious contents nor the conscious manifestations of implicit bias are commensurate with the determinate proposition BLACK MALES ARE DANGEROUS. This proposition is an approximation through and through.

Of course, more should be said with regards to conscious states to maintain the rejection of the equivalence assumption. By rejecting this assumption, I do not mean to foreclose the possibility that conscious states are indeterminate just like the unconscious states that underlie implicit bias. I grant that it might turn out that they are equivalent (both indeterminate), but we should not take their equivalence as a premise in our inquiry into the metaphysis of implicit bias. For our purposes, rejecting the equivalence assumption carried us to indeterminate contents and this is where we will end.


6 Moral considerations

In closing, I want to raise the question of responsibility with regards to the three abovementioned theories: Are implicit biases beyond reproach?

The associative content view treats the unconscious mind like a machine that pairs concepts and valences by picking up regularities in the world. If this is all that the mind is doing, implicit biases would be quite reasonable, as they would be the direct products of the racism, sexism, ageism, you name it, in our society. They would be indicative of the qualities of the environment around the bearers of implicit bias and not the qualities of the bearers themselves. Under the associative content view, you might regard implicit biases as unwitting facts about the human mind.[40] The associative content view carries with it this chilling outcome.

The propositional attitude view provides a more palatable alternative: it carries implicit biases into the realm of rationality without making them ipso facto reasonable. According to the propositional attitude view, implicit biases can be modified by logical processes. To eliminate implicit bias, in this view, would be to undertake a laborious process to extinguish all the token representations of the propositional attitudes that encode implicit bias.

The propositional attitude view, however, falls short of acknowledging the complexity of implicit biases and the culprit, I have argued, is its extradetermination of content. The indeterminate content view recognizes the complexity of implicit bias, and following the mood analogy, gives a more substantial role to implicit bias in shaping our conscious experiences. But, compared to the associative content view and the propositional attitude view, the indeterminate content view leaves us with a more challenging task. In this view, there aren’t any isolated propositional attitudes to be targeted. Consequently, the indeterminate content view cannot deliver a strategy for extinction.

One resource that the indeterminate content view has is the notion of interpretation. Crane suggests, in this vein, “Discovering what you believe can resolve indeterminacy and unclarity in your [unconscious], and produce a conscious judgment that settles things as far as you are concerned”.[41] We do not yet know the power of interpretive processes in modifying indeterminate contents that are housed in the unconscious.[42]

In the implicit bias literature, previous studies have looked at the effects of delivering interpretations of the subjects’ IAT scores before they were administered an explicit attitude test.[43] When the subjects thought there was no correlation between their IAT scores and their racial attitudes, their explicit views diverged from their implicit biases. When they trusted the IAT, however, they also complied with its results. Namely, their views demonstrated more prejudice than they would have, had the subjects mistrusted the IAT. This study gives us some insight into the effects of authoritative third-personal interpretation. Future studies should test self-interpretation and its effects. Under the indeterminate content view, then, strategies for the extinction of implicit biases will involve interpretive strategies. Moods can also provide a starting point for this avenue of research.


Current views of implicit bias fail to explain its diverse manifestations that are underwritten by its diverse contents. The modest goal of this paper was to rouse suspicion in the reader when philosophers all too quickly embrace the equivalence of content across the conscious and unconscious domains. This tendency to assume equivalence gives rise to the extradetermination of implicit bias’s contents. Motivated by an analogy to moods, I argued that we should reject overly simple accounts of this attitude that has remarkable ramifications for our mental and moral lives.



Seurat, Georges-Pierre (1884–1886). Study for a Sunday on the Island of La Grand Jatte: couple walking. Fitzwilliam Museum, Cambridge



Amodio, D. M., & Devine, P. G. (2006). Stereotyping and evaluation in implicit race bias: evidence for independent constructs and unique effects on behavior. Journal of Personality and Social Psychology, 91(4), 652–661.

Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg more employable than Lakisha and Jamal? A field experiment on labor market discrimination. The American Economic Review, 94(4), 991–1013.

Briñol, P., Petty, R., and McCaslin, M. (2009). “Changing Attitudes on Implicit versus Explicit Measures: What is the Difference?” In Attitudes: Insights from the New Implicit Measures (pp. 285–326), R. Petty, R. Fazio, and P. Briñol (Eds.). New York: Psychology Press.

Brownstein, M. (forthcoming). “Implicit Bias and Race”. To be included in Routledge Companion to the Philosophy of Race.

Brownstein, M., & Saul, J. M. (2016). Implicit bias and philosophy. Oxford, United Kingdom: Oxford University Press.

Crane, T. (2016). “The Unity of Unconsciousness”. Proceedings of the Aristotelean Society, CXVII(1).

Dasgupta, N., & Greenwald, A. G. (2001). On the malleability of automatic attitudes: Combating automatic prejudice with images of admired and disliked individuals. Journal of Personality and Social Psychology, 81(5), 800–814.

De Dreu, C. K. W., Greer, L. L., Handgraaf, M. J. J., Shalvi, S., Van Kleef, G. A., Baas, M., … Feith, S. W. W. (2010). The Neuropeptide Oxytocin Regulates Parochial Altruism in Intergroup Conflict Among Humans. Science, 328(5984), 1408–1411.

Dennett, D. C. (1981). Brainstorms: Philosophical Essays on Mind and Psychology. Montgomery, VT: Bradford Books.

Fazio, R.H. & Towles-Schwen T. (1999). “The MODE model of attitude-behavior processes.” In Chaiken, S. and Trope, Y. (eds.), Dual-Process Theories in Social Psychology. New York, NY: Guilford Press: 97–116.

Gawronski, B., & Bodenhausen, G. V. (2011). The associative-propositional evaluation model: Theory, evidence, and open questions. Advances in Experimental Social Psychology, 44: 59–127.

Gendler, T. (2008). Alief and belief. Journal of Philosophy, 105: 634–663.

Greenwald, A., McGhee, D., and Schwartz, J. (1998). Measuring Individual Differences in Implicit Cognition: The Implicit Association Test. Journal of Personality and Social Psychology, 74 (6): 1464–1480.

Gregg, A. P., Seibt, B., & Banaji, M. R. (2006). Easier done than undone: asymmetry in the malleability of implicit preferences. Journal of Personality and Social Psychology, 90(1), 1–20.

Hahn, A., Judd, C. M., Hirsh, H. K., & Blair, I. V. (2014). Awareness of implicit attitudes. Journal of Experimental Psychology: General, 143(3), 1369–1392.

 Holroyd, J. & Sweetman, J. “The Heterogeneity of Implicit Bias”. In Brownstein, S. and Saul, J. M. (eds.), Implicit bias and philosophy. Oxford, United Kingdom: Oxford University Press, pp. 80–103.

Huebner, B. (2009). Troubles with Stereotypes for Spinozan minds. Philosophy of the Social Sciences, 39(1), 63–92.

Laughland, O., Swaine, J., and McGraw, D. “Cleveland Officer Who Fatally Shot Tamir Rice Will Not Face Criminal Charges.” The Guardian. Guardian News and Media, 28 Dec. 2015. Web. 12 Nov. 2016.

Levy, N. (2015). “Neither Fish nor Fowl: Implicit Attitudes as Patchy Endorsements”. Noûs, 49(4), 800–823.

Madva, A. (2016). “Why implicit attitudes are (probably) not beliefs”. Synthese, 193(8), 2659–2684.

Mandelbaum, E. (2014). “Thinking is believing”. Inquiry, 57 (1): 55–96.

Mandelbaum, E. (2015). “Attitude, Inference, Association: On the Propositional Structure of Implicit Bias”. Noûs, 50(3), 629–658.

McGrath, Matthew, “Propositions”, The Stanford Encyclopedia of Philosophy (Spring 2014 Edition), Edward N. Zalta (ed.).

Mitchell, C.J., De Houwer, J., & Lovibund, P.F. (2009). “The propositional nature of human associative learning”. Behavioral and Brain Sciences, 32:182–246. 6

Nier, J. a. (2005). “How Dissociated Are Implicit and Explicit Racial Attitudes? A Bogus Pipeline Approach”. Group Processes & Intergroup Relations, 8(1), 39–52.

Payne, B. K. (2001). “Prejudice and perception: the role of automatic and controlled processes in misperceiving a weapon”. Journal of Personality and Social Psychology, 81(2), 181–192.

Payne, B. K. & Gawronski, B. (2010). “A history of implicit social cognition: Where is it coming from? Where is it now? Where is it going?”. In Handbook of implicit social cognition: Measurement, theory, and applications, B. Gawronski, and B. Payne (eds.), New York, NY: Guilford Press, pp. 1-17.

Rankine, C. (2014). Citizen: An American Lyric. Minneapolis, MN: Graywolf.

Sabin, J. A., & Greenwald, A. G. (2012). “The Influence of Implicit Bias on Treatment Recommendations for 4 Common Pediatric Conditions: Pain, Urinary Tract Infection, Attention Deficit Hyperactivity Disorder, and Asthma”. American Journal of Public Health, 102(5), 988–995.

Hahn, A., Judd, C. M., Hirsh, H. K., & Blair, I. V. (2014). Awareness of implicit attitudes. Journal of Experimental Psychology: General, 143(3), 1369–1392.

Weisberg, M. (2007). “Who Is a Modeler?”. The British Society for the Philosophy of Science, 58(2), 207–233.

Williamson, T. (2000). Knowledge and Its Limits, Oxford: Oxford University Press.

Wittenbrink, B., Judd, C. M., & Park, B. (2001). “Spontaneous prejudice in context: Variability in automatically activated attitudes”. Journal of Personality and Social Psychology, 81: 815–827.

Zawidzki, T. W. (2008). The function of folk psychology: mind reading or mind shaping? Philosophical Explorations, 11(3), 193–210.


[1] These examples come from Claudia Rankine’s Citizen which provides a remarkable study that cuts to the core of our racist dispositions.

[2] Bertrand & Mullainathan 2004; Sabin & Greenwald 2012

[3] For a good review of this disagreement consult Payne & Gawronski 2010, pp.1–12

[4] Brownstein forthcoming

[5] Greenwald et al. 1998

[6] Hahn et al. 2014 shows that some individuals can successfully predict their implicit attitudes and points out that previous research has mistakenly focused on correlations between implicit and explicit attitudes.

[7] Gendler 2008, p. 44

[8] It is possible for some subjects to have racial implicit biases and explicit racist attitudes. In fact, we can say that most (if not all) racists will have racial implicit biases.

[9] This is merely an analogy. There are strong prima facie disanalogies between moods and implicit biases. For example, moods have a distinct phenomenology and moods are often characterized as not being directed at objects (as opposed to emotions which are directed towards particular objects). I think both of these disanalogies are interesting but I will not be able to entertain them here.

[10] Wittenbrink et al. 2001, Amodio & Devine 2006

[11] Payne et al. 2001. This study is especially interesting because it raises a question about whether there is a false percept involved in causing this behavior or if the error stems from a false inference about the identity of a veridical percept.

[12] Wittenbrink et al. 2001

[13] De Dreu et al. 2011, Terbeck et al. 2011

[14] Amodio & Devine 2006

[15] Holroyd & Sweetman 2016

[16] Proponents of this view include: Gendler 2008, Fazio 1999, Gawronski & Bodenhausen 2011

[17] Proponents of this view include: Mitchell et. al 2009, Levy 2015, Mandelbaum 2015

[18] Mandelbaum 2015, pp. 639–646

[19] e.g. Levy 2015’s “patchy endorsements”

[20] e.g. extinction and counterconditioning.

[21] Briñol et al 2009

[22] IThe power of this view lies in explanation and not prediction. the lack of predictive power of the propositional attitude view. bid., p. 294. In both argument conditions, the words “African American” occurred the same number of times to disallow mere activation effects.

[23] Mandelbaum reports the study results as follows, “strong arguments alleviated the bias, weak arguments did not have any effect” (2015, p. 641). Nevertheless, the researchers make no such claim; they merely state that “under high elaboration conditions, automatic evaluations were found to be more positive toward Blacks for the strong than the weak message. In contrast, for low elaboration conditions, we did not find as much attitudinal responsiveness to the manipulation of argument quality” (Briñol et al. 2009, p. 295, emphasis added).

[24] Ibid., p. 295

[25] Correspondingly, moods can be defeated by reason but are usually recalcitrant to such interventions.

[26] Gregg et al. 2006. Not only were they successful in inducing bias, but they also found that biases were transferred when they introduced two new groups, and told the subjects that these groups were equivalent to Niffites and Luupites. These findings most directly support propositional learning, which, as mentioned before, is a separate issue. The effectiveness of the equivalence relation in generating a bias regarding the secondary group, however, can be viewed as congenial to the propositional attitude view. The most plausible explanation of this result is the inferential processing of the bias against the primary group to generate an implicit bias regarding the secondary group.

[27] Levy 2015 deals with this problem by asserting that implicit biases are “patchy endorsements” because of their partial participation in inferential processes, unlike beliefs which participate in such processes systematically.

[28] Suggested by the following authors respectively: Madva 2016, Briñol et al. 2009, Mandelbaum 2015.

[29] Levy 2015

[30] Not all available views about conscious beliefs assign these properties to beliefs. The propositional attitude proponents in this paper do.

[31] Mandelbaum 2015 uses the proposition BLACK MALES ARE DANGEROUS as an example (p. 636). I will retain this example and follow the small caps notation for propositions.

[32] Dennett 1981, p. 107

[33] Caution! I don’t mean to allude to the vagueness problem in philosophical logic here. I rely on a commonplace notion of ‘vague’, meaning not-fixed, indistinct, unsettled. The way I’m using it, the sentence “This is a heap” is not vague, it is determinate in virtue of having a linguistic form. Likewise, I use determinate to mean the ordinary notion of being precise, distinct, fixed, or settled.

[34] Mandelbaum 2014, fn 11.

[35] For the purposes of this paper I am employing this term in a minimally-committed way, as: “the sharable objects of the attitudes and the primary bearers of truth and falsity” (McGrath 2014).

[36] Weisberg 2007 defines modeling as “the indirect theoretical investigation of a real world phenomenon” and compares it to practices that do not employ the mediation of a model.

[37] Crane 2016, p. 11. Here and elsewhere, Crane defends the view that the category belief properly applies only to unconscious representations, and conscious instantiations constitute thought (2013). I will not take a stance with regards to this position.

[38] I want to reiterate that the indeterminacy should be understood as not being ‘fixed’ or ‘precise’. Contrasting this notion with Cartesian “clear and distinct ideas” can be helpful.

[39] In addition to various other worries like how in cases where the individual is explicitly egalitarian, how the conscious egalitarian belief and the unconscious racist belief will cause different behaviors.

[40] This opens the way to the argument of “a perfect storm of human error” causing racist action. In fact, this phrase was used by the prosecutor in the legal case following the murder of Tamir Rice, an innocent black 12-year-old. The prosecutor said that Tamir’s “death was caused by a ‘perfect storm of human error’” (Laughland, et al. 2015).

[41] Mandelbaum 2015, p. 15

[42] Zawidzki’s (2008) mind-shaping account of propositional attitude ascription might also provide more avenues for research in this key.

[43] Nier et al. 2005. Cf. Hahn et al. 2014

Invited Comments from Grace Helton (Princeton)

Comments on Yumusak’s “Implicit Bias and the Unconscious”

Grace Helton
Princeton University


I would like to thank Ege Yumusak for the opportunity to read and comment on her highly intriguing paper, “Implicit Bias and the Unconscious.”

Yumusak develops a new model of implicit bias, on which implicit biases are non-conscious dispositions to form certain conscious attitudes. Implicit biases do not themselves have determinate contents, but they dispose their subjects to form certain conscious attitudes which do have determinate contents. In this way, implicit biases are a bit like moods, which, on Yumusak’s view, lack determinate contents, at the same time that they can shape how their subjects represent the salience and valence of environmental features.

Yumusak claims her account is superior to rival accounts in three ways:

  • The account can explain why there is some evidence that implicit bias is reasons-responsive and other evidence that implicit bias is not reasons-responsive.
  • The account can explain the fact that different measures of implicit bias do not correlate well with each other.
  • The account can explain the context-sensitivity of implicit bias.

I will consider each of these claims in turn. In support of the first claim, Yumusak says: “…the simultaneous reasons-responsiveness and recalcitrance [of implicit bias] is not a problem for [my] view, because reasons-responsiveness is not a condition that needs to be met by indeterminate contents” (§5). Yumusak’s suggestion seems to be that if implicit bias is a kind of non-conscious disposition with indeterminate content, then it is not puzzling that implicit bias should sometimes fail to be reasons-responsive; implicit dispositions just aren’t the sort of things we would expect to change to fit the evidence. What is less clear is how Yumusak’s account is meant to explain results which suggest that, in certain cases, implicit biases are reasons-responsive. Why, for instance, does presenting subjects with a strong argument that we should hire more Black professors reduce their implicit race bias (Briñol et al. 2009)? If implicit biases are dispositions without determinate contents, what is the route by which arguments can rationally impact implicit biases?

In support of the claim that the proposed account can explain the non-correlation of different measures of implicit bias, Yumusak suggests that if implicit bias is a complex disposition, this disposition will likely trigger different conscious attitudes on different occasions. Thus, it is not surprising that performance on (say) affective bias-related tasks poorly predicts how one will perform on (say) semantic bias-relevant tasks.

Yumusak is right that if implicit bias is a disposition with indeterminate content, then it is not surprising that implicit bias might trigger different conscious attitudes on different occasions. What is unclear is how this result is meant to explain the fact that a subject’s performance on (say) an affective bias-related task is not a good predictor of how she will perform on (say) a semantic bias-related task. In order for the former to be a good predictor of the latter, it need not be the case that the former involves the same kinds of conscious mental states as the latter. There need only be some kind of statistical correlation between these performances.

More to the point, the fact that a subject’s behavior on an affective bias-related task is not a good predictor of how she will perform on a semantic bias-elated task seems to suggest a reason to reject Yumusak’s view that the very same disposition underlies both performances.[1] For it seems that if and insofar as we have any reason to think it is one disposition which underlies both of these performances, there must be some statistical correlation between these performances. Consider an analogy with moods, which Yumusak thinks are importantly similar to implicit biases. A depressed mood might cause you to interpret others’ neutral behavior as lukewarm and others’ lukewarm behavior as mildly hostile. Insofar as it is your depressed mood which explains both of these tendencies, these tendencies should be at least somewhat correlated. If there were no such correlation, we should question whether these behaviors are both explained by your depressed mood.

Finally, consider Yumusak’s claim that her proposal can account for the context-sensitivity of implicit bias. While it seems right that Yumusak’s account is well-positioned to explain context-sensitivity, it seems that the competitor accounts can also explain this result.

Consider the result that implicit bias against Black people is reduced when subjects are primed with a picture of a Black person in a church environment (vs. a picture of a Black person in an urban environment) (Wittenbrink et al. 2001). Now suppose that one of the rival accounts of implicit bias is true: implicit biases are beliefs with determinate contents, contents like “Black people are dangerous” (Mandelbaum 2016). This view can easily explain the “church” result. Beliefs work in concert with other beliefs to produce new inferences. Thus, it seems plausible that subjects in this condition might combine an implicit belief like “Black people are dangerous” with the socially pervasive belief “people sitting in church are harmless” and, via the normal inductive-inferential route, come to the conclusion “that person sitting in church is harmless (or at least not dangerous).” Different contexts render other background beliefs relevant, so the belief model of implicit bias can explain the context-sensitivity of bias by appealing to these contextually-relevant background beliefs.


Works Cited

Briñol, P., Petty, R., and McCaslin, M. (2009). “Changing Attitudes on Implicit versus Explicit Measures: What is the Difference?” In Attitudes: Insights from the New Implicit Measures (pp. 285–326), R. Petty, R. Fazio, and P. Briñol (Eds.). New York: Psychology Press.

Holroyd, J. & Sweetman, J. (2016). “The Heterogeneity of Implicit Bias”. In Brownstein, S. and Saul, J. M. (eds.), Implicit bias and philosophy. Oxford, United Kingdom: Oxford University Press, pp. 80–103.

Mandelbaum, E. (2016). “Attitude, Inference, Association: On the Propositional Structure of Implicit Bias”. Noûs, 50(3), 629–658.

Wittenbrink, B., Judd, C. M., & Park, B. (2001). “Spontaneous prejudice in context: Variability in automatically activated attitudes”. Journal of Personality and Social Psychology, 81: 815–827.



[1] This is Holroyd & Sweetman’s “heterogeneous” view of implicit bias (Holroyd & Sweetman 2016). Yumusak dubs this view “radical,” (§1), but in light of the non-correlation results, I struggle to see what is radical about this approach, whether or not it is ultimately correct.

Invited Comments from Katherine Puddifoot (Birmingham)

Response to “Implicit bias and the unconscious” by Ege Yumusak


Katherine Puddifoot
University of Birmingham


Two competing accounts of the nature of the implicit biases have emerged from recent philosophical and psychological work: the unconscious association account, according to which implicit biases are unconscious associations between members of social groups and concepts or affective responses (Gendler 2008; Fazio 1999; Gawronski and Bodenhausen 2011), and the propositional attitude account, according to which they are unconscious beliefs or belief-like propositional attitudes (Mitchell et al 2009; Levy 2015; Mandelbaum 2015). In “Implicit bias and the unconscious” Ege Yumusak’s goal is to address shortcomings of each approach, presenting an alternative account of implicit biases: the indeterminate content view.

The shortcomings of the existing accounts, according to Yumusak, are due to the way that implicit biases appear to be reasons-responsive but also reasons-recalcitrant. The apparent reasons-responsiveness of implicit biases is a problem for the unconscious association account because associative mental states are ordinarily not responsive to reasons. The reasons-recalcitrance of implicit biases is a problem for the propositional attitude account because implicit biases are not responsive to reasons in ways that propositional attitudes ordinarily are, and, as they are not always sensitive to reasons, it is possible that they are not responsive to reasons but rather to linguistic tokens or motivating factors. Due to the ambiguity about what explains the appearance that implicit biases are responsive to reasons, Yumusak concludes that more work must be done to establish that implicit biases are bona fide beliefs. A diagnosis of why existing accounts fall short of adequately accounting for implicit bias is provided. It is argued that two tendencies are found in the literature: the tendency to assume that the content of unconscious mental states are equivalent to the content of conscious mental states, and the tendency to assume that implicit biases are determinate propositional attitudes when they are actually vague and indeterminate.

An alternative approach to implicit bias, which avoids these problematic tendencies, is developed. According to this new approach, ““Implicit biases are aspects of the subject’s unconscious belief-system that have particular dispositional profiles. They can be modelled by propositions but their contents are not determined by propositions.” (page 19) Implicit biases are interpretations of behaviour. By ascribing an implicit bias we are ascribing behaviour rather than saying what mental state is present. There is no determinate proposition that undergirds the behaviour. It is possible for a proposition to act as a model for the implicit bias, but numerous propositions could be used to model any implicit bias, meaning that the content of the bias is indeterminate. For example, take a particular implicit bias that leads to discriminatory behaviour towards Black males, associating them with violence. The propositions BLACK MALES ARE DANGEROUS and UNKNOWN BLACK MALES ARE DANGEROUS can both be models for the same bias.

This new approach to implicit bias is taken to explain how implicit biases can be both reasons-responsive and reasons-recalcitrant because indeterminate contents do not need to be consistently either reasons-responsive or reasons-recalcitrant. The heterogeneity of implicit bias, and the way that different features of implicit biases can be activated in different contexts, is consistent with the indeterminacy of the content.

The paper provides an important contribution to the field of philosophy of implicit bias. Previous discussions have been underwritten by the assumption that unconscious and conscious mental representations have similar functional profiles, and that, as Yumusak argues, consciousness just brings to light mental representations. Discussion has focussed on how implicit biases compare with conscious beliefs, with the assumption that if they are unconscious beliefs they will be similar to conscious beliefs. Moreover, it has been assumed that if implicit biases are propositional attitudes they will be similar to conscious propositional attitudes. It is important to recognise that there is room in the space of logical possibilities for conscious and unconscious mental representations of the same type (e.g. beliefs) to differ significantly in ways other than whether consciousness has brought them to light. This observation significantly complicates existing discussion of implicit bias. Evidence that implicit biases differ from conscious propositional attitudes, has previously been taken to show that they are not propositional attitudes. But once it is recognised that conscious propositional attitudes could have a different dispositional profile to unconscious propositional attitudes, the same evidence would be consistent with implicit biases being unconscious propositional attitudes. In sum, the current discussion changes the parameters of the debate over whether implicit biases are propositional or associative by raising the possibility that implicit biases could be unconscious propositions even if they differ significantly from conscious propositions.

One issue that it is difficult for propositional attitude theorists to address is exactly what the propositional content of any specific implicit bias is. Implicit biases are often implicated in behaviour that could be explained in terms of numerous propositions. For example, a person’s level of implicit bias relating to Black people can predict her seating position in a room occupied by a Black person. This behaviour could be explained in terms of the person unconsciously believing many different things, e.g. they could believe BLACK PEOPLE ARE DANGEROUS or INTERRACIAL INTERACTIONS ARE UNPLEASANT, or numerous other things. Yumusak’s account avoids this problem by arguing that the content of implicit biases is indeterminate. There is no determinate proposition that underwrites the behaviour. By arguing that there is no single proposition that underwrites the implicit bias it is not necessary to explain why the person who is biased might display some behaviours consistent with some belief but not others. This is a serious proposal that deserves scrutiny.

Yumusak’s paper therefore raises important issues to be addressed by philosophers of implicit bias. However, for the purposes of the current response, I focus on a number of potential problems and points in need of further clarification.


1.  Identifying the class of psychological phenomena

In any attempt to provide an account of implicit bias it is necessary to identify the class of psychological phenomena that are being accounted for, before providing the finer details of the account. Yumusak identifies the class by stating that the term implicit bias is used to “refer to mental states that give rise to behaviors whose causal origins subject lack awareness of, and as such, under the subjects direct control” (p. 3). This definition aims to build upon two competing conceptions of the “implicit” in implicit bias, one of which defines implicit in terms of a lack of awareness, and another that defines implicit in terms of a lack of control. For Yumusak, implicit biases have causal origins that are not accessible to the subject and, as such, the subject cannot directly control the biases. There is a problem with this definition of implicit: it is too inclusive. There are mental states that are not implicit but give rise to behaviors whose causal origins are not known and are consequently not under the subject’s direct control. Take an ordinary commonsense belief obtained in childhood. One might believe that p but not remember the causal origin of the belief that p. Knowledge about the causal origin of the belief might be long forgotten. One might not be able to control the belief because one cannot remember the origin of the belief and knowledge of the source would be required to shake one’s conviction. This example raises the possibility that Yumusak has not provided a working characterisation of implicit biases that adequately distinguishes them from other mental states.


2.  A challenge to the associative model?

Yumusak aims to criticise both the unconscious association and propositional attitude accounts of implicit bias, however I would suggest that the move to reject the unconscious association account is too quick. The model is dismissed by the author on the basis of psychological findings, previously discussed by Eric Mandelbaum (2015), that could be taken to show that implicit biases are responsive to argumentation. The author takes the findings to be most plausibly explained by implicit biases entering into inferential relations. Meanwhile, an interpretation of the findings that have been provided by defenders of the association model is dismissed. I here outline a defence of the interpretation that is rejected.

The particular study that is discussed by Yumusak surveyed subjects’ opinions of policies to promote employing more Black professors (Briñol et al 2009). Participants were presented with either strong or weak arguments in favour of the policy. It was found that under specific conditions the strength of the arguments determined the scores that subjects got on the Implicit Association Test (IAT), which measures implicit racial bias. Strong arguments made subjects less biased when the subjects thought that they were a part of a small group of students asked to complete the survey and the policy was going to lead to change in their own university. However, strong arguments did not shift bias when participants thought there were many people who were surveyed and that the policy would impact a remote university. The authors of the study, who were committed to the unconscious association approach, concluded that the stronger arguments led to more favourable thoughts being associated with the policy and Blacks when the information was processed carefully while the weaker arguments led to less favourable thoughts being associated with the same. Yumusak rejects this interpretation of the findings on the basis that it is “highly implausible that subjects would react to a weak argument by becoming more biased be it through having unconscious or conscious thoughts about the policy” (p.9).

I think that this is more plausible than Yumusak suggests. It is easy to imagine a scenario in which one develops an increasingly negative attitude towards p, becoming more biased towards p, as a result of being provided with unconvincing arguments in its favour. Take fracking for natural gas. Imagine that you are naturally disposed to adopt a slightly negative attitude towards fracking, so you have a slight disposition towards consciously and unconsciously responding in a negative way towards fracking, but you have not formed a fully developed view. When you are exposed to arguments in favour of fracking, they are weak. The weakness of the arguments increases the negative associations that you make with fracking. It could do this by changing your propositional attitudes through inference, but it could also simply lead you to associate the topic with negative affect or concepts like dangerous to the environment. It might be, for instance, that the more you think about how weak the arguments are in favour of fracking, the more you think about its negative points, thereby strengthening your negative association with the practice. Meanwhile, positive arguments could imbue positive associations, leading fracking to be associated with positive affective responses or concepts like good for the economy.

The weakness of Yumusak’s response to the unconscious association interpretation of these findings is significant. The author needs to make a strong case against the unconscious association account if the account proposed in the paper is to be plausible. If the unconscious association account is correct then the subsequent discussion of how propositions provide a model for implicit biases is substantially weakened. If implicit biases are unconscious associations then there is little reason to focus on how propositions can model them.


3.  Are propositions a good model for reasons-recalcitrant mental representations?

A related worry is that the author argues that propositions can be a useful model for implicit biases, explaining how people behave when under the influence of implicit bias, although implicit biases are not propositional. However, as discussed in the paper (page 10), implicit biases differ from standard propositional attitudes like beliefs. In cases in which one would expect propositional attitudes like beliefs to be reasons-responsive, implicit biases are reasons-recalcitrant. If implicit biases are not reasons-responsive then one main reason for accepting that they are similar to propositional attitudes like beliefs is lost. Consequently, one reason for thinking that propositions will be useful models for implicit biases, explaining behaviour that occurs under the influence of the biases, is also lost.


4.  Motivation for extra-determination claim

The author focuses on how the propositional attitude view stipulates that implicit biases have a specific structure: that of a propositional attitude. It is argued that propositional attitudes “being determinate, […] cannot meet the demands of psychological phenomena that have diverse and context-dependent manifestations”. This is a surprising claim. Standard propositional attitudes can manifest in diverse ways and contexts. They can also be triggered by a diversity of thoughts and experiences. Take my belief it is hot today. I can have this belief in a diversity of contexts. Whether I believe that it is hot today can depend on a number of factors, e.g. the temperature, but also whether I am on holiday in a hot climate. More needs to be said about what it is about the particular diverse and context dependent manifestations of implicit biases that means that they are not plausibly propositional attitudes.


5.  Why not just dispositions?

Yumusak’s positive proposal is that “Implicit biases are aspects of the subject’s unconscious belief-system that have particular dispositional profiles. They can be modelled by propositions but their contents are not determined by propositions” (page 19). How does this positive proposal improve upon Eric Schwitzgebel’s (2010) claim that implicit biases are in-between beliefs, which have some but not all of the dispositions stereotypically associated with beliefs? Under Schwitzgebel’s account, implicit biases have particular dispositional profiles. Whether or not a certain disposition will manifest, e.g. responsiveness to reasons, will depend upon the context. What seems to distinguish Yumusak’s positive proposal from Schwitzgebel’s is the idea that propositions can model implicit biases. However, I previously noted that propositions might not provide a useful model for implicit biases because implicit biases do not respond to reasons in the same way as other propositional attitudes. Why, then, should the positive proposal of the current paper be preferred to a dispositionalist approach, such as Schwitzgebel’s?


6.  Why indeterminate and not heterogeneous?

The positive proposal of the current paper is proposed as an alternative to the view that implicit biases are heterogeneous (page 6). The heterogeneity view has been proposed in response to findings suggesting that affective measures of implicit bias do not correlate with semantic measures (Holroyd and Sweetman 2016). The same person might have a negative affective response towards Black people, leading them to choose to sit far from a Black peer, but simultaneously display no bias associating Black people with a lack of intelligence. Or vice versa. My question is, how is the indeterminate content approach, the positive proposal of the paper, any better at explaining these results than the heterogeneity account? The indeterminate content view is said to accommodate the lack of consistency across different measures because “different aspects of the subject’s unconscious belief-system can be causally activated in different contexts, giving rise to the context-dependence of implicit bias” (page 20). The suggestion seems to be that different aspects of a person’s unconscious underwrite different responses, and therefore the affective responses (e.g. seating distance) are underwritten by different aspects of the subject’s unconscious to the semantic responses (e.g. association of peers with certain concepts). But if different aspects of a person’s unconscious underwrite the different responses then the things that are currently labelled implicit bias count as a diverse and heterogeneous class. So far, the author might agree. But if the class of implicit bias is diverse and heterogeneous, it is possible that some implicit biases are propositions, with determinate content, which are reasons-responsive. Meanwhile, other things in the class of implicit biases might be affective responses that are non-responsive to reasons, but are not usefully modelled using propositions, in virtue of being affective responses. Evidence of diversity and heterogeneity across the class of items currently classified as implicit biases might count in favour of dividing up the class to reflect the diverse nature of the class, rather than ascribing indeterminacy to the whole class. And if the class is divided in the way that I suggest it might be, it could turn out that no implicit biases are both indeterminate and usefully modelled with propositions. Why should it be thought that there is a distinctive class of “aspects of a subjects unconscious belief-system” that can be modelled by propositions and are indeterminate?



Briñol, P., Petty, R., and McCaslin, M. (2009). “Changing Attitudes on Implicit versus Explicit Measures: What is the Difference?” In Attitudes: Insights from the New Implicit Measures (pp. 285–326), R. Petty, R. Fazio, and P. Briñol (Eds.). New York: Psychology Press.

Fazio, R.H. & Towles-Schwen T. (1999). “The MODE model of attitude-behavior processes.” In Chaiken, S. and Trope, Y. (eds.), Dual-Process Theories in Social Psychology. New York, NY: Guilford Press: 97–116.

Gawronski, B., & Bodenhausen, G. V. (2011). The associative-propositional evaluation model: Theory, evidence, and open questions. Advances in Experimental Social Psychology, 44: 59–127.

Gendler, T. (2008). Alief and belief. Journal of Philosophy, 105: 634–663.

Holroyd, J. & Sweetman, J. “The Heterogeneity of Implicit Bias”. In Brownstein, S. and Saul, J. M. (eds.), Implicit bias and philosophy. Oxford, United Kingdom: Oxford University Press, pp. 80–103.

Levy, N. (2015). “Neither Fish nor Fowl: Implicit Attitudes as Patchy Endorsements”. Noûs, 49(4), 800–823.

Mandelbaum, E. (2015). “Attitude, Inference, Association: On the Propositional Structure of Implicit Bias”. Noûs, 50(3), 629–658.

Mitchell, C.J., De Houwer, J., & Lovibund, P.F. (2009). “The propositional nature of human associative learning”. Behavioral and Brain Sciences, 32:182–246. 6

Schwitzgebel, E. (2010). “Acting contrary to our professed beliefs or the gulf between occurrent judgment and dispositional belief”. Pacific Philosophical Quarterly, 91: 531-553.

9 thoughts on “Implicit Bias and the Unconscious”

  1. It’s a wonderful opportunity to e-present my paper and to receive feedback from thoughtful commentators. I owe thanks to the Minds Online organizers for this great conference, and to Grace Helton and Katherine Puddifoot for their comments. I have benefited enormously from their feedback. I will respond (briefly enough, I hope) to what I take to be the main issues (some objections & pleas for further clarification) raised by their comments.

    Are the competitor accounts more successful than I present? In particular, (i) is the association view better-positioned to explain the responsiveness of implicit bias to argument, and (ii) is the propositional view better-positioned to explain the context-sensitivity of bias (by referencing the interference of background beliefs)?

    To address (i) Puddifoot offers an example from fracking. In her example, an agent who is slightly disposed against this practice hears weak arguments in favor of fracking and becomes even more disposed against it. Their negative associations are strengthened against unconvincing pro-fracking propaganda. In the paper, I had conjectured that the associative view did not have the resources to make sense of a case like this (taken from Briñol et al.s study concerning the employment of Black professors). I welcome the chance to elaborate on what I take to be implausible here. As I see it, the only promising way to accommodate this result fails because it overestimates the rational properties of associations. For these negative associations to be adequately responsive to a weak argument such that they are strengthened by its weakness, there ought to be propositional processing of this argument. This processing would need to evaluate the merit of the argument, as “bad argument!” and this outcome would need to interact with the association that undergirds dispositions against fracking (or Blacks in general, which I think is somewhat more ofa stretch). I grant that just as associative learning can result in propositional structure, propositional processing can result in associative structure. However, given the highy intellectual activity that is required to react to the strength of an argument, I regard the Briñol et al. study to be problem for the associative view. Add to this the fact that the researches took pains to control for the frequency of concepts that might partake in associations and the overall positiveness of the language in the contents of the arguments. (As an aside, having participated in many of these studies, I find it easier to believe that participants’ negative associations regarding the experimenters or psychology studies in general would have been more easily affected by the weak arguments!). I think we would be better off taking associations to be responsive to surface-level properties of language (in which case we will have to trust the experimenters’ that they sufficiently controlled for those variables and think these differences would be the same across weak and strong arguments).

    Helton brings up a study to raise (ii). She proposes an explanation for how subjects’ implicit bias can be diminished by exposure to pictures of Black people in churches under the propositional view. She suggests that the church prompt might trigger the background belief “people sitting in church are harmless”, interfering with the implicit bias “Black people are dangerous”. If this is true, it seems like the propositional view can accommodate context-sensitivity by appealing to background beliefs just as easily, if not better than the indeterminate content view. I think the explanation that Helton uses is exactly the kind of explanation we can use in our modeling practices. In other words, I don’t see a reason to take more seriously our practice of proposing unconscious logical inferences between ‘beliefs’. In the paper, I aim to warn against the inferential jump from rational properties to full-blown propositional content by arguing against LOTH-type reasoning and commending a more modest approach.

    Helton also raises a related concern with regard to the reasons-responsiveness of implicit bias. Helton worries that my approach does not gain any traction regarding this property of implicit bias. This is fair. In response to this worry, I think the best I can do is to fall back on the analogy to moods: even though breathing exercises or lifestyle changes are better interventions for managing moods, talk therapy is not ineffective across the board because after all, we respond to reasons. Just like we have our “good days” and “bad days” colored by various moods, which we can enter by the right cues, our “less biased” or “more biased” selves can be solicited by argumentation. This does not mean that ‘our take on the world’ can be sufficiently altered so as to eliminate bias by argumentation but argumentation can impact on the subject’s unconscious belief-system. It is fair to say that this sensitivity of the unconscious to argumentation does not look like ‘responsiveness to reasons’ in the traditional sense. Still, I think this might be a promising way to go to further develop the indeterminate content view.

    How does the indeterminate content view compare to (i) the in-between beliefs of Schwitzgebel (2010) and (ii) the heterogeneous view of Holroyd & Sweetman (2016)?

    In response to (i), I can say that while there are certainly parallels between Schwitzgebel’s approach and mine, I do not take beliefs to be “vague predicates”. Instead, I take the unconscious as a holistic belief-system and propose to analyze it by propositions, using a practice I call ‘modeling’ following Crane 2016. This approach is in tension with Schwitzgebel’s central claim that in cases where one’s espoused attitudes depart from what their behavior seems to suggest, we ought to suspend ascription of belief, and regard their attitude as an in-between belief instead. I defend the ascription of the proposition—as a model, or as I say in the paper, “as an approximation through and through”. I hold that this propositional attitude ascription has purchase elsewhere: in our hermeneutic, scientific and critical practices. In folk psychology we can use it to understand each other, in social psychology we can use it to predict other behavior, and in our moral lives, we can evaluate the fittingness of various reactive attitudes based on these models. While in Schwitzgebel’s picture, the more reflective moments of the agent and their more spontaneous reactions are evaluated on a par, making the belief in-between, measured on a sliding-scale, according to the picture I present in this paper, there need not exist a sliding scale between egalitarian and racist dispositions of an agent. That said, Schwitzgebel’s starting point is certainly congenial to mine: he is interested in going against “the desire to handle neatly cases like the ones at hand, the desire to support a simple yes-or-no (or numerical degree) answer to questions about what these subjects believe” (2010, p. 545-6). Same here. But I dispense with the neatness in a different way.

    Lastly, Puddifoot asks, why not divide up the class, as Holroyd & Sweetman (2016) recommend? Why ascribe indeterminacy to the whole attitude? Often, in discussions of implicit bias the case of the explicitly egalitarian but implicitly biased individual gets the most attention. There are obviously good reasons for this (my paper privileges this subclass as well). However, in characterizing implicit bias across subjects who might or might not be overtly racist, I think we can cover more ground by focusing on the ways in which implicit bias affects the conscious experience of the subject by modifying the salience and valence of features of the environment. In the paper, I link this phenomenology to the experience of moods. In my view, the most important effects of implicit bias are closely tied with the subject’s worldview and thus are markedly different than those of consciously held beliefs. They operate on a deeper level. For this reason, I take it that whatever undergirds this phenomenology will be a less-fixed, or indeterminate feature of the unconscious belief-system of the subject. In my view, the heterogeneous (but fixed) objects of study that concern Holroyd and Sweetman are proper targets for models of implicit bias. In this paper, I try to lure the reader into seeing the unconscious and our practice of ascribing propositional attitudes under a new light, which I think deserves a fair hearing given the thorniness of this issue.

  2. Hi Ege,

    Thanks for the paper!

    The extradetermination problem seems like it might just be an instance of a general problem in science – namely, that observable evidence underdetermines underlying reality. It is not always clear precisely which mental representations underlie a particular implicit bias, just as it is not always clear precisely which mental representations underlie a particular visual experience (and similarly for any scientific posit). But this doesn’t seem to force us to reject the hypothesis that mental representations underlie these phenomena. Instead we can engage in some degree of idealization (which also seems common in science) and develop ways of testing any particular hypothesis about the contents of an implicit bias, while understanding that the answer will always remain somewhat underdetermined.

    In general, Dennett-style rhetorical questions about precisely which mental representations are stored in a given mind can be given principled answers if we assume some general empirical hypotheses about mental representations. For example, representations that are literally stored will generally be more quickly accessed than representations that have to be inferred. Reaction-time measures can therefore be just one useful part of a solution to the extradetermination problem by pinning down which representations are stored, and which are merely approximations of what’s stored.

    So, for example, on the kind of mental-logic-based view of implicit cognition that Mandelbaum endorses, the hypothesis that the content of an implicit bias is black men are dangerous would generate the prediction that showing subjects with that bias a sentence of the form ‘If black men are dangerous, then X’ would facilitate X (as measured by a lexical decision task, for example) more quickly and easily than it would for subjects without that bias. And more generally, subjects’ speed in correctly discriminating aspects of presented sentences should, all else equal, increase as a function of how close the contents of those sentences are to the content of a stored mental representation such as an implicit bias.

    One reason to go to all this trouble to hang on to representational explanations of implicit bias and other unconscious phenomena over dispositional accounts is that mental representations are causally efficacious particulars that can figure in causal-mechanical explanations of biased behavior. Appealing to dispositions can allow us to construct approximating models, as you say in the paper, and thereby try to capture regularities in behavior. But we don’t seem to get a deep, causal-mechanical explanation of that behavior. Accepting some degree of underdetermination, which can be narrowed with further empirical study, seems to me like a small price to pay for that kind of deeper explanation.

  3. Hi Ege,

    Thank you for your paper! It was very enjoyable, and I found it helpful in not only pinpointing a worry in the implicit bias debate, but perhaps a more general worry about philosophical methodology (although I think JQD response above helped assuage that worry).

    My thought is that by following the particular order or explanation you outline in the paper, the indeterminate view (perhaps) doesn’t address the nature of phenomena that is “implicit bias”. You say that the indeterminate view looks at the entirety of the unconscious belief system, some of which includes beliefs that are pernicious biases. But what is it that distinguishes those beliefs from other beliefs in the unconscious system, and are we really interested in those beliefs, i.e. the non-pernicious ones? You might think that looking at the entire body of beliefs that are unconscious will provide insight into those beliefs that influence biased behavior, but maybe there is a similar equivalence worry of treating all of those beliefs the same. For instance, what it it turned out that negative valence beliefs have stronger influences on behaviors than neutral or positive valence beliefs. My thought is that methodologically, it might be more advantageous to test subjects starting from the assumption that negative valenced propositions influence behaviors, than to observe behaviors and try to determine which kind model of proposition it adheres to. This goes back to your admittance that the indeterminate view falls short of having a strategy for extinction. Ultimately, If pernicious implicit beliefs are indistinguishable on your account from the entire body of unconscious beliefs, then the task to eradicate them may prove to be insurmountable. And this, I take, is a concern that we should consider seriously, since it makes little sense to say simply that we have these beliefs, but to figure out what to do with them to improve (race-based) social relations.

  4. Hi Elís, Thank you for your comment. As I started getting into in my response to Jake, I think yes, we are interested in the non-pernicious ones too. If we take the effects of implicit bias as broadly world-shaping as I take them to be (i.e. by changing the salience and valence of the representation of the whole scene), I think we will have reason to take implicit bias as an aspect of a holistic system just as we would take the conscious manifestations of implicit bias to be a qualitative aspect of the subject’s conscious experience. I don’t think it is a consequence of the view that eradication is impossible, but I think it is a consequence that the extinction methods like counter-conditioning are not going to do anything! These are all important concerns that I think I have to think about more and build in in later drafts of the paper because I certainly agree with you that simply saying “this is an aspect of the unconscious belief system” doesn’t sound surmountable at all.

  5. Just to correct a false (yet often repeated) claim by Mandelbaum (also cited in your paper and in some comments above): it is not correct that counterconditioning is ineffective in changing implicit evaluations. For evidence that counterconditioning does effectively change implicit evaluations, see Olson & Fazio (2006, PSPB) and Hu, Gawronski, & Balas (in press, SPPS). The latter paper actually shows an asymmetry between counterconditioning and counterinstructions in reversing explicit and implict evaluations. Although both procedures reversed previously conditioned explicit evaluations, only counterconditioning reversed previously conditioned implicit evaluations.

  6. Hi Jake, I certainly agree with you that Dennett-style questions have a wide skeptical application. I am trying to use them with caution because I share your conviction that it seems too costly to give up causally efficacious particular altogether. I think my decision to bring this up in the context of implicit bias is a bit unclear or undermotivated in the paper. Thank you for letting me clarify that. I am particularly worried about representational talk (and what Dennett would call ‘brain-writing’) in the case of implicit bias, and especially with regards to implicit biases that concern social groups.

    As I see it, extradetermination is particularly worrying problem where the concepts that are involved with biases of this nature are as rich as they are in this case and as sensitive as they are to contextual cues. In the paper I have laid out several features of what I take to be the conscious manifestations of implicit bias—with an analogy to moods—to motivate the move away from causal particulars. The mental-logic-picture starts falling apart, I think, when we think about the socially embedded subject interacting with individuals of a different race or gender in an environment where they might be negotiating their explicit egalitarian commitments and the features of the environment that keeps grabbing their attention. I am not compelled to think of those cases as involving unconscious inferences, as Grace points out in her example involving church cues but I’m happy for us to model them as such to make scientifically valuable predictions.

    Thanks again for your question. I’m sure this is not satisfying as I need to think more about this matter! I really appreciate the comment.

  7. Just a quick follow-up: more evidence for the effectiveness of counterconditioning in changing implicit evaluations can be found in Karpinski & Hilton (2001), Gawronski & LeBel (2008); Grumm, Nestler, & von Collani (2009), Gibson (2008), etc. etc. All of these studies have been published in highly respected, peer-reviewed journals.

    The “influential” study by Brinol et al. (2009), on the other hand, is still not published in a peer-reviewed journal. The reference is to a book chapter that describes an unpublished study by the authors, and the book in which this chapter appeared was edited by the same authors. I’m not questioning the significance of the findings by Brinol et al. for the debate or the scientific integrity of researchers. However, I am concerned that the debate was inspired to a large extent by a rather selective and highly biased review by Mandelbaum in which he cherry-picked whatever seemed to fit to his argument. On the one hand, he downplayed the effectiveness of counterconditioning in changing implicit evaluations, which has been demonstrated in dozens of studies. But then he is going on to base his arguments on an unpublished study that has not even passed a peer-review. To be fair, Mandelbaum is also citing a few other studies that are consistent with his argument, one being our own paper on balance effects on implicit evaluations (Gawronski, Walther, & Blank, 2005). Yet, even with this citation, he conveniently describes only the first study a series of three experiments, the latter two of which show associative effects of mere co-occurrence regardless of relational qualifiers.

    To be clear: all of this doesn’t mean that Ege’s arguments are ill-founded or wrong, but I do think that Mandelbaum’s paper has been taken a little too serious in the philosophical debate (and that includes Ege’s paper). A more thorough review of the literature reveals a much more complex picture than Mandelbaum has described in his paper; one that poses a challenge to both a purely associative AND a purely propositional account. That’s why Galen Bodenhausen and I have started to focus on mutual interactions between associative and propositional processes (Gawronski & Bodenhausen, 2006, 2011). I’m obviously biased on this point, but I don’t think that the available body on implicit evaluations can be understood without considering the mutual interplay between associative and propositional processes.

  8. Hi Bertram, your point is well-taken. I also mention one instance of Mandelbaum misreporting what the Brinol et al. paper (footnote 23) says about a particular result. I have contacted Brinol et al. with questions for some of their findings because the chapter does not have sufficient information but I never heard from them. Putting that issue aside, I engaged critically with Mandelbaum’s paper because it is concerned with content ascription as opposed to the psychological literature that talks about processes that are involved in implicit bias. The Mandelbaum paper is a good fit against Dennett-style arguments like mine because it demonstrates what is fishy about attribution of propositional content. Thank you for your comment and the references!

  9. Hi Everyone,

    Just some quick responses. My original paper was shock-and-awe length long, and went unread and journals didn’t seem so fond of it. So, to get it to reasonable publication length I narrowed down the citations to what I thought were the few easy to digest studies–not necessarily the ones I thought were the strongest, but the easiest to understand. There are are so, so many more. Since that paper came out one can look at the work of Jeremy Cone & Melissa Ferguson (or Mann & Ferguson), or Kurdi & Banaji (and Banaji was no friend to propositional accounts). I also think one can use other examples than the ones I happen to mention there. I recommend doing so! There are many others to find, and I have no special affection for the ones that happened to not get cut. (Also: I mentioned more studies than seem to get discussed. I’m not complaining about this, just flagging it. Of course I haven’t even mentioned the work of Jan De Houwer, a well-known fellow traveler.)

    I should say I agree that the Brinol study in particular isn’t one of my favorites–it’s just clear. But the fact that I didn’t discuss the other studies (which by the way I do discuss in a footnote and try to explain even though it’s tangential to the discussion) isn’t cherrypicking–to think that is to miss the dialectical situation. If IAs are only associative then propositional interventions should never work. If IAs are propositional and there are associations between propositions then showing any associative effects aren’t particularly interesting. The lack of discussion of those other effects isn’t cherrypicking, it’s ignoring orthogonal data.

    I fully support Bertram’s motivation to find interplay between propositional and associative processes (in fact, I think the model I offer does such a thing, since it builds propositions out of concepts which can have their own associative connections). Of course, one needs to explain how these processes interact and the devil is in the details. But this is where our efforts should lie. I must admit that I’m confused that you’ve read my work as doing anything but that. I argue for three main theses: 1) implicit attitudes aren’t just associatively structured (if they were, then there should be no logical interventions); 2) implicit attitudes are propositionally structured (to explain logical interventions); 3) implicit attitudes are probably beliefs. (The just in (1) is meant to flag associations that come off of propositions.)

    To be clear I also fully reject the idea that there is a “philosophical debate” such as Bertram alludes to. There’s data and theory, and some philosophers and psychologists looks at some, others don’t. One thing that is impressive and heartening is how seriously Gawronski and Bodenhausen take theorizing–they are trying to build an honest-to-god theory and that is rare in this area. To the extent that my theory is read by philosophers its because there are just a dearth of theories offered (especially ones that haven’t already been shown to be descriptively inadequate).

    In sum: denying that there are propositional effects on implicit attitudes is a dead end. It’s not because the three (or however many) studies I cited are sure to replicate; it’s that there are tons of other effects like this out there. The challenge is to show how a mere associative processor (or mere associative structures) can explain them away. Or that’s at least how I see the field.

    Ege, as for your interesting paper I’d just like to add that JQD and I just published a paper critiquing dispositionalism in general (“Against Dispositionalism: Belief in Cognitive Science) which takes on a lot of these arguments.

    And thanks everyone–it’s cool to have these discussions and I’m glad I stepped away from my normal luddite stance to read these.

Comments are closed.