Moving Beyond System 1 and System 2

Pavlovian Conditioning

As an effect, Pavlovian conditioning refers to changes in behavior that are due to the spatio-temporal pairing of events (Bouton, 2016; De Houwer, Barnes-Holmes, & Moors, 2013). For instance, in fear conditioning, pairing an originally neutral light with an aversive electric shock results in the light evoking fear (Craske, Hermans, & Vansteenwegen, 2006). Likewise, evaluative conditioning is a change in liking of an originally neutral stimulus that occurs as the result of pairing this stimulus with a valenced (i.e., positive or negative) stimulus (De Houwer, 2007). Within cognitive psychology, it is most often assumed that Pavlovian conditioning effects are mediated by the formation of simple associations between representations in memory (Bouton, 2016; see De Houwer, 2018a, for a historical analysis). Although different association formation models have been put forward that differ with regard to their assumptions about the conditions under which associations are formed and influence behavior (see Bouton, 2016, for a review), they all share the assumption that conditioning occurs only if pairings result in the formation of associations. Those associations are hypothetical structures in memory via which activation can spread from one representation to another, thereby allowing stimuli to evoke responses they did not evoke before the stimulus pairings.

During the past decade, however, my colleagues and I have argued that Pavlovian conditioning (in humans) is mediated by the formation of propositional representations (De Houwer, 2009, 2018b; Mitchell, De Houwer, & Lovibond, 2009). In essence, propositions are informational units that can specify how things are related. For instance, the propositions “light predicts shock” and “light sometimes co-occurs with shock” both involve the concepts “light” and “shock” but differ in how those concepts are related. Propositions can be vague about the way in which events are related (e.g., “light is somehow related to the shock”) and can be irrational (e.g., “light produces the shock”) but all propositions are inherently relational in nature. Propositional models postulate that the formation of propositions in Pavlovian conditioning is akin to higher order reasoning: it involves problem solving that is directed at discovering the way in which events in the world are related. Hence, these models predict that Pavlovian conditioning and reasoning are moderated by the same variables, including verbal instructions, awareness, and mental load. In the following paragraphs, we highlight some of the evidence supporting this prediction.

First, just like higher order cognitive phenomena, conditioning in humans is highly sensitive to verbal information. Already since the 1930s, we know that a simple instruction stating that a light will be followed by a shock, results in fear of the light even if the light and shock are never paired (Cook & Harris, 1937). Subsequent research revealed striking parallels between the moderators of fear conditioning via instructions and the moderators of fear conditioning via the actual pairing of stimuli (see Mertens, Boddez, Sevenster, Engelhard, & De Houwer, 2018, for a review, and De Houwer, 2018b, for related evidence on evaluative conditioning via instructions). Moreover, instructions and actual pairings can jointly influence behavior. For instance, during a first phase, Lovibond (2003, Experiment 3) delivered a shock to participants whenever a compound of two colored squares appeared on a computer screen. As a result, each color separately evoked fear. Afterward, he informed participants that one of the colors is safe (i.e., never followed by the shock), thus implying that the shock on the preceding compound trials was related to the other color. Importantly, such a simple instruction (“this color is safe”) increased fear for the other color, demonstrating that actual pairings (i.e., on compound trials) and instructions can jointly determine behavior. Finally, also instructions about the nature of the relation between stimuli (e.g., whether one stimulus is a predictor vs. a cause of the other) have a profound impact on Pavlovian conditioning (e.g., Hughes, Ye, Van Dessel, & De Houwer, 2019; Waldmann & Holyoak, 1992). The fact that Pavlovian conditioning is highly sensitive to verbal instructions fits well with the idea that both pairings and instructions result in propositional (and thus relational) knowledge that provides the basis for changes in behavior.

Second, Pavlovian conditioning depends heavily on awareness of the relations between stimuli, as well as the availability of working memory resources (see Lovibond & Shanks, 2002; Mitchell et al., 2009, for reviews). Since the 1970s, numerous studies have been published in which fear conditioning was observed only in participants who were aware of the stimulus contingencies (e.g., Dawson & Biferno, 1973). On top of this correlational evidence, experimental studies showed that variables that influence contingency awareness (e.g., the presence of an attention demanding secondary task) also influence fear conditioning. For some time, researchers considered the possibility that some types of Pavlovian conditioning such as evaluative conditioning might not depend on contingency awareness and working memory resources, but recent evidence strongly argues against this possibility (see Corneille & Stahl, 2019, for a review). For instance, a meta-analysis showed that contingency awareness is by far the most important moderator of evaluative conditioning, accounting for no less than 36% of the variance (Hofmann, De Houwer, Perugini, Baeyens, & Crombez, 2010). This heavy reliance of various Pavlovian conditioning effects on contingency awareness and working memory resources parallels findings in research on higher cognition and fits very well with the idea that the construction of relational knowledge relies on working memory (Halford et al., 2010). It is, of course, difficult to exclude the possibility that some instances of Pavlovian conditioning in humans can occur in the absence of contingency awareness and working memory resources. However, after many years of research on this topic, I believe that it is safe to conclude that, if those instances exist in humans, they occur only when very strict boundary conditions are met (e.g., Greenwald & De Houwer, 2017).¹ While it definitely remains worthwhile to look for and document the boundary conditions of those instances of non-relational Pavlovian conditioning (see McLaren et al., 2014, for a discussion of possible candidates), it can no longer be assumed by default that conditioning effects are non-relational in nature. On the contrary, I would argue that the burden of proof is now on those who wish to claim that specific instances of conditioning effects in humans are non-relational.

Implicit Evaluation

Implicit evaluation can be defined as the automatic effect of stimuli on evaluative responses (De Houwer, Gawronski, & Barnes-Holmes, 2013). In less technical terms, it refers to the spontaneous “gut” feelings that people regularly experience. Most often, implicit evaluations are captured using indirect measures such as the Implicit Association Test (IAT; Greenwald, McGhee, & Schwartz, 1998). As is the case with Pavlovian conditioning, the most popular account of implicit evaluation relies on the notion of simple associations in memory (e.g., Fazio, 2007). More specifically, it is assumed that the representations of valenced stimuli in memory are associated with the representations of the concepts “good” or “bad”. Upon presentation of a valenced stimulus, activation of the stimulus representation will spread automatically to those evaluative representations and thus result in an automatic evaluative response.

Although the idea of automatic spreading of activation via associations provides a simple account of the automatic nature of implicit evaluation, it is not the only possible account. Recently, I put forward the idea that implicit evaluation is mediated by propositions (and thus relational knowledge; De Houwer, 2014; see also Mandelbaum, 2016). Whereas the formation of propositions might rely on working memory (and thus instructions, awareness, and resources), the retrieval of propositions could well be similarity-based and largely independent of working memory as is, for instance, postulated by episodic memory models (e.g., Hintzman, 1986; Logan, 1988; Schmidt, De Houwer, & Rothermund, 2016).

Several interesting predictions have been derived from a propositional account of implicit evaluation. A first line of research was based on the idea that propositions can result from instructions. Hence, if implicit evaluation reflects propositions, it should be sensitive to instructions. In line with this prediction, simply providing instructions about the evaluative properties of novel stimuli (e.g., Bob helps old ladies cross the street) is enough to produce changes in the implicit evaluation of those stimuli (e.g., automatic positive reactions to Bob; e.g., Cone, Mann, & Ferguson, 2017; De Houwer, 2006; Gregg, Seibt, & Banaji, 2006). One might object that these effects (a) arise only for process-impure measures that capture not only implicit but also explicit (i.e., non-automatic) evaluations, (b) are mediated by changes in explicit evaluations that produce changes in associations, or (c) are limited to novel stimuli. Recent evidence, however, argues against these possible objections. First, it has been shown that instruction-based implicit evaluations can be found on a wide variety of indirect measures (e.g., Van Dessel, Ye, & De Houwer, 2019), as well as on parameters in multinomial processing tree models designed to capture automatic processes (e.g., Smith, Calanchini, Hughes, Van Dessel, & De Houwer, in press). Second, research revealed that instructions can influence implicit evaluations even in the absence of changes in explicit evaluations (Van Dessel, De Houwer, Gast, Smith, & De Schryver, 2016). Third, to the extent that instructions provide diagnostic and credible information, they can alter even long-standing, deeply rooted implicit evaluations. For instance, Van Dessel, Ye, and De Houwer (2019) informed participants about an actual event in the life of Mahatma Gandhi during which he forbade doctors from giving a medicine to his wife, resulting in her death, but later took the medicine himself when he contracted the disease, resulting in his recovery. Simply providing this single piece of verbal information drastically reduced implicit liking of Gandhi on a variety of measures.

A second line of studies examined the impact of relational information on implicit evaluations. For instance, in a study by Peters and Gawronski (2011), participants saw on each trial the name of a previously unknown person together with a positive or negative trait. Some persons were paired most often with a negative trait whereas other persons were paired most often with a positive trait. Importantly, in certain cases, participants were informed that a person had the traits opposite to the ones shown on the screen. This relational information had a significant impact on implicit evaluations but did not reverse them completely (also see Moran & Bar-Anan, 2013; Moran, Bar-Anan, & Nosek, 2016; Zanon, De Houwer, Gast, & Smith, 2014). This pattern of results could reflect the joint impact of propositions and associations on implicit evaluations but it could also reflect the joint impact of multiple propositions on implicit evaluations, some based on pairings (e.g., “this person co-occurred with positive traits”) and some based on relational information (e.g., “this person is good because he was paired with negative traits”; De Houwer, 2018b; Moran et al., 2016). Regardless of whether there is a significant contribution of associative (non-relational) processes, it is now widely accepted that propositions (and thus relational knowledge) have a profound impact on implicit evaluations. Hence, it is time to discard the assumption that implicit evaluation is by default non-relational in nature.

Habitual Responding

Habit research is another stronghold of non-relational, associative theories. In fact, often habitual responding is even defined in terms of a non-relational mechanism, more specifically, the activation of responses via S-R associations. On the basis of this definition, habitual behavior is typically contrasted with goal-directed behavior, that is, behavior that is a function of its consequences (see Wood & Rünger, 2016, for a review). Whereas goal-directed behavior can be based on relational knowledge (i.e., on how a response is related to realizing a goal), habitual responding in this technical sense is banned from the realm of relational cognition on an a priori basis.

Recently, however, serious questions have been raised about the quality of the evidence for habitual responding in humans. In addition to genuine failures to find evidence for habitual responding (de Wit et al., 2018), alleged demonstrations of habitual behavior were criticized on the basis of inadequate controls for goal-directed processes. As my colleagues and I pointed out in a recent paper (De Houwer, Tanaka, Moors, & Tibboel, 2018), tests for goal-directed processes need to be both sensitive enough (sensitivity criterion) and look for all the goals at which the behavior could be directed (information criterion). Several studies that claimed to have found evidence for habitual responding in humans did not meet these criteria. Consider the well-known study of Neal, Wood, Wu, and Kurlander (2011) showing that people who often eat popcorn in movie theaters continue to eat popcorn in movie theaters when it is stale, whereas control participants do not eat the stale popcorn. The behavior of the first group was considered to be habitual because a devaluation of the tastiness of the popcorn did not lead to a reduction of popcorn consumption, suggesting that eating popcorn was not directed at the goal of eating tasty food. It is possible, however, that people who often eat popcorn when watching movies consume popcorn for other reasons than taste, for instance, to have a fuller cinematic experience. Eating stale popcorn might also serve this goal. Hence, the behavior of eating (stale) popcorn could well be goal-directed, be it directed at a goal different than the one that Neal et al. controlled for. Although we do not exclude the possibility that humans sometimes behave in purely stimulus-driven (i.e., habitual) ways, the picture that is emerging now is that goal-directed behavior is the default mode of behaving (Moors, Boddez, & De Houwer, 2017).

When considering this conclusion, it is important to realize that it only concerns habitual responding in the technical sense (i.e., behavior that is a function of the evoking stimulus but not of its current consequences). The term “habit” is often also used in a different sense, namely as behavior that is emitted frequently and/or automatically. There can be no doubt that humans often repeat the same behavior in similar situations and that behavior can have features of automaticity (i.e., occur in the absence of conscious intentions, fast, or without awareness of the controlling variables; Moors & De Houwer, 2006b). These types of “habitual” behaviors, however, could well be goal-directed (see Aarts & Dijksterhuis, 2000; Bargh, 1990). In fact, more and more researchers are considering the possibility that impulsive behavior is strategic in nature (i.e., directed at goals; e.g., Kopetz, Woerner, & Briskin, 2018), including seemingly irrational behavior as seen in addiction (Baumeister, 2017; Hogarth, 2018).

Within the context of the present paper, it is important to note that associative (i.e., non-relational) models dominate also research on automatic goal-directed behavior (e.g., Bargh, 1990; Moskowitz, 2012). However, it is also possible that automatic goal-directed behavior depends on relational knowledge. In fact, one could argue that goal-representations are inherently relational in that they need to specify how someone relates to an end-state (i.e., the fact I desire the end-state) as well as the role that elements have in that relation (e.g., the fact that it is me who desires). Also goal-pursuit seems to necessitate the involvement of relational knowledge, more specifically, knowledge about whether and how an action promotes or hinders the realization of the desired end-state. Hence, it would be worthwhile to start exploring the role of relational knowledge in automatic goal-directed behavior.

To recapitulate: Whereas research in humans provided little evidence for habitual responding in a technical sense, human behavior is often habitual in a broad sense. In contrast to what is often argued, however, the latter behaviors might well depend on relational knowledge.

Converging Arguments From Behavioral Research

Unbeknownst to many cognitive psychologists, a large number of psychological scientists still operate within the behavioral research traditions that originated from the work of behaviorists such as Skinner (1953). In this section, I draw attention to the fact that recent developments in behavioral research are compatible with the idea that Pavlovian conditioning, implicit evaluation, and seemingly habitual behavior are grounded in relational knowledge.

To fully appreciate the merits of this argument, I first need to point out that behavioral and cognitive approaches in psychology have different goals and are therefore not competitors (De Houwer, 2011; Hughes, De Houwer, & Perugini, 2016). Behavioral researchers operate at the functional level of explanation that aims to explain behavior in terms of elements in the environment. Cognitive researchers, on the other hand, operate at the mental level of explanation that documents the mental mechanisms via which elements in the environment influence behavior (Bechtel, 2005; De Houwer, 2011). For instance, whereas behavioral researchers would explain increases in fear of a light in terms of parings of conditioned and unconditioned stimuli (i.e., Pavlovian conditioning as an abstract explanatory behavioral principle), cognitive researchers explain the impact of light-shock pairings on fear for the light in terms of the formation of associations or propositions (i.e., Pavlovian conditioning as a to-be-explained phenomenon). Rather than competitors, both approaches are mutually reinforcing in that knowing more about (the moderators of) environment–behavior relations constrains theories at the mental level, whereas theories at the mental level can facilitate the discovery of (moderators of) environment–behavior relations (see Hughes, De Houwer, & Perugini, 2016, for a more detailed treatment of these ideas).

Within behavioral psychology, it has been proposed that much of human behavior is relational in nature (e.g., Hayes, Barnes-Holmes, & Roche, 2001). Behavioral researchers consider responding to be relational when it is a function not of a single event (e.g., the presence of a single light) but of the relation between events. For instance, people can learn to press a left key whenever two identical stimuli are presented on the screen and to press a right key whenever two non-identical stimuli are presented, regardless of what those stimuli are. Importantly, behavioral researchers have recognized that humans can also respond relationally in arbitrary ways, a phenomenon they refer to as arbitrarily applicable relational responding (Hayes et al., 2001). For instance, the fact that people can respond as if the word “glass” has some functions in common with the object “glass” is considered to be an instance of relational responding that occurs even though the relation between the word and the object is arbitrary (as indicated by the fact that the same object is referred to in French as “verre”). Behavioral researchers have developed functional concepts to describe and analyze this type of relational behavior, paradigms to examine it, and coherent ideas about the learning history that is necessary to establish it (Hayes et al., 2001). Interestingly, despite extensive efforts, until now evidence for (flexible) arbitrarily applicable relational responding has not been found in non-human animals (e.g., Hughes & Barnes-Holmes, 2014; Lionello-DeNolf, 2009). This striking difference between species is consistent with the idea that arbitrarily applicable relational responding (present only in humans) lies as the heart of language (also present only in humans).

In line with the ideas put forward in this paper, it has been argued that also phenomena such as Pavlovian conditioning and implicit evaluation are instances of arbitrarily applicable relational responding. From this perspective, Pavlovian conditioning has much in common with symbolic phenomena like language. More specifically, the pairing of two stimuli is assumed to function as a symbolic cue that signals that the paired stimuli are equivalent in some ways, much like the word “SAME” signals the equivalence of stimuli. For instance, pairing a neutral word with a positive word is assumed to cue people to respond as if the two are equivalent, which includes responding to the originally neutral word in positive ways (De Houwer & Hughes, 2016). Likewise, implicit evaluation has been conceived of as a brief and immediate relational response to stimuli (Hughes, Barnes-Holmes, & Vahey, 2012). Although it is beyond the scope of this paper to provide a detailed discussion of these ideas and findings (see Hughes & Barnes-Holmes, 2016; Hughes, De Houwer, & Barnes-Holmes, 2016; Törneke, 2010, for accessible introductions), for the present purposes, it is important to note that these developments in behavioral research fit very well the idea that also Pavlovian conditioning and implicit evaluation are mediated by relational knowledge. Indeed, any mental mechanism that can produce relational behavior must somehow draw on knowledge about how events are related (see De Houwer et al., 2016, for a more detailed discussion).