No-loss gambling shows the speed of the unconscious

Abstract

This paper investigates the time it takes unconscious vs. conscious knowledge to form by using an improved “no-loss gambling” method to measure awareness of knowing. Subjects could either bet on a transparently random process or on their grammaticality judgment in an artificial grammar learning task. A conflict in the literature is resolved concerning whether unconscious rather than conscious knowledge is especially fast or slow to form. When guessing (betting on a random process), accuracy was above chance and RTs were longer than when feeling confident (betting on the grammaticality decision). In a second experiment, short response deadlines only interfered with the quality of confident decisions (betting on grammaticality). When people are unaware of their knowledge, externally enforced decisions can be made rapidly with little decline in quality; but if given ample time, they await a metacognitive process to complete. The dissociation validates no-loss gambling as a measure of conscious awareness.

Introduction

How can we tell if someone is aware of their knowledge? Artificial grammar learning (AGL; Reber, 1967) is a particularly useful methodology to address this question as it demonstrably elicits both conscious and unconscious knowledge according to subjective measures of awareness (e.g. Dienes, 2008, Gaillard et al., 2006, Johansson, 2009). Two types of knowledge are involved in sequence classification in AGL: structural knowledge and judgment knowledge (Dienes and Scott, 2005, Scott and Dienes, 2008). During the initial training phase of an AGL experiment, participants are exposed to rule-based sequences generated by the grammar in question. Structural knowledge is (either conscious or unconscious) knowledge of the structural consequences of the grammar (and can consist of rules, patterns of connection weights, chunks, or whole items taken as examples of the structure). During testing, participants classify further novel sequences in terms of their grammaticality (whether they conform to or violate the studied rules). Here, judgment knowledge is the (conscious or unconscious) knowledge constituted by such a judgment (i.e. the knowledge that the test item is or is not grammatical). When both structural and judgment knowledge are conscious, grammaticality decisions are based on hypothesis-driven rule-application or a conscious recollection process of recognised exemplars or bigrams, trigrams or other parts of exemplars encountered during training. Feelings of intuition or familiarity are expressed when structural knowledge is unconscious but judgment knowledge is conscious (e.g.: “I know I’m correct but I don’t know why”) (Norman et al., 2006, Norman et al., 2007). When both knowledge types are unconscious the phenomenology is that grammar judgments are mere guesses; no conscious metaknowledge of what has been learned is expressed. Fig. 1 depicts the relationship between the conscious status of these knowledge types and the associated phenomenology (see also Scott & Dienes, 2010a, for a model of how structural and judgment knowledge develop in AGL; and Scott & Dienes, 2008; Pasquali, Timmermans, & Cleeremans, 2010, for models of how judgment knowledge may become conscious).

Numerous subjective measures of awareness have been used in AGL studies including verbal reports (Reber, 1967, Reber, 1969), confidence ratings made on binary (Tunney & Shanks, 2003) or continuous scales (Dienes, Altmann, Kwan, & Goode, 1995) and structural knowledge attributions (Dienes and Scott, 2005, Scott and Dienes, 2008, Scott and Dienes, 2010a, Scott and Dienes, 2010b, Scott and Dienes, 2010c, Wan et al., 2008; see also Chen et al., 2011, Guo et al., 2011, Rebuschat and Williams, 2009). Recently, wagering has been used to assess conscious awareness. In their AGL study, Persaud, McLeod, and Cowey (2007) asked participants to make high or low wagers using real or imaginary money after making a grammaticality decision. When correct, the wager was added to their total; when incorrect it was deducted. The procedure was presumed to motivate participants to make consistently high wagers whenever they felt more confident than a mere guess in order to maximise financial gain (Koch & Preuschoff, 2007). A tendency to wager high on accurate decisions would then provide an index of subjective awareness which is particularly useful for researchers shy of overestimating unconscious knowledge, a potential pitfall of using verbal reports (Berry & Dienes, 1993). Persaud et al. found that despite a high level of overall performance (81% accuracy) participants made high wagers at a lower than optimal level. This was taken as evidence that participants were unaware of their knowledge (contrast Clifford, Arabzadeh, & Harris, 2008).

However, this post-decision wagering procedure has been criticised due to the potential problem of risk (loss) aversion (Kahneman & Tversky, 1979). For the risk averse participant, losing a certain amount of money is more salient than gaining the same amount. This may encourage consistently low wagers to minimise loss when they do have some awareness of knowledge. In effect, this increases the rate of measured unconscious knowledge as operationalised by Persaud et al. (2007). Conversely, participants showing little risk aversion may be willing to wager large amounts on what, to them, seems like a random process. Thus, wagering without sensitivity to the risk aversion of the individual distorts conclusions drawn about the amount of conscious or unconscious knowledge expressed (Schurger & Sher, 2008). Indeed, Fleming and Dolan (2010) found that economic factors in post-decision wagering systematically influenced measures of perceptual sensitivity. Altering the wager size affected the proportion of low to high wagers which would lead one to change conclusions drawn about low or high levels of awareness. Furthermore, Dienes and Seth (2010) compared a binary verbal confidence scale (‘guess’ vs. ‘sure’) against wagering in an AGL task while measuring risk aversion. They found a greater willingness to indicate confidence in responses using the verbal scale and that risk aversion significantly correlated with the amount of conscious knowledge as measured by wagering, but not as measured by verbal confidence.

In a second experiment, Dienes and Seth (2010) introduced a new methodology to indicate the presence of unconscious knowledge in AGL: No-loss gambling. During the test phase, participants indicated their confidence in each grammaticality decision by either betting on the grammaticality decision (in order to win one sweet if correct) or on a transparently random process. If they chose the latter, they shuffled and then picked one of two face-down cards, one of which had ‘SWEET’ printed on the invisible side, the other had ‘NO SWEET’. Therefore choosing to bet on the cards meant there was a 50:50 chance the participant would add to their winnings. If one chooses to bet on the random process, rather than on the grammaticality decision, clearly no conscious preference for grammaticality or ungrammaticality is shown. This methodology bypasses the potential confound of risk aversion as participants never have the opportunity to lose their winnings but motivation to perform is maintained to maximise gains. When betting on the cards, participants still displayed above chance classification accuracy, satisfying the guessing criterion of unconscious knowledge (Dienes et al., 1995).

In using verbal reports, participants may have their own idiosyncratic definition of ‘guess’ (Gardiner, Ramponi, & Richardson-Klavehn, 1998). In everyday language, ‘guess’ can refer to a range of feelings of confidence. In classifying a test sequence some participants might say ‘guess’ when it felt as if they knew literally nothing relevant (the definition of ‘guess’ we are interested in – the absence of confidence) whereas others may take ‘guess’ to mean merely ‘low confidence’. Merely ‘low’ confidence decisions can involve some awareness of knowing. However, betting on a random process shows a lack of conscious judgment: they are unaware of having any relevant structural knowledge. This is a literal guess as even if confidence in a decision was low, but not absent, it would still be worth betting on the grammar judgment to maximise reward rather than opting for the 50:50 gamble. Furthermore, this plausibly eliminates the problem of bias shown by any participant who says they are guessing but thinks they are not (Dienes, 2008). In other words, no-loss gambling robustly distinguishes conscious from unconscious judgment knowledge (see Fig. 1: No-loss gambling prima facie separates the guess response based on unconscious judgment knowledge from all other response types made with some degree of conscious judgment knowledge).

This paper aims to improve on the methodology of no-loss gambling. In the original study by Dienes and Seth (2010), participants attributed their knowledge (by betting on the grammar decision or on the cards) after the grammar judgment was made. It is possible that awareness of judgment knowledge can be relatively transient. This conscious knowledge could be forgotten, or could degrade, between the two decisions leading one to bet on the cards despite having had conscious judgment knowledge. In effect this would increase the amount of unconscious knowledge as measured by betting on the cards. To address this problem it is simply a matter of ensuring both grammaticality classification and decision strategy are reported simultaneously while the test sequence is available to account for this possibility (cf Tunney & Shanks, 2003, with verbal confidence ratings).

Further, no method can a priori prove itself from the arm chair just because it has good face validity. More broadly, the utility of the no-loss gambling methodology can only be verified if the results it yields are in line with theoretically motivated hypotheses (Dienes, 2004, Dienes, 2008). Thus, a further aim of the paper is to demonstrate the utility of the method – by exploring a contradiction in dual-process theories of recognition memory. Dual-process theories posit that responses based on familiarity are made rapidly and automatically whereas recollection responses are relatively effortful and time-consuming due to strategic retrieval (e.g.: Jacoby, 1991, Yonelinas, 2002; see also the two-stage recollection hypothesis of Moscovitch, 2008). This view is supported by Hintzman and Caulton (1997), who found shorter response times for item recognition than modality judgments requiring conscious recollection of a previous learning episode. Furthermore, Boldini, Russo, and Avons (2004) found that modality matches between learning and test (presumed to influence familiarity) under a strict deadline increased recognition compared to modality mismatches whereas under longer deadlines deep processing enhanced recognition compared to shallow processing (presumed to influence recollection). Such theories relate to AGL in that unconscious structural knowledge can express itself through familiarity and conscious structural knowledge can express itself through recollection. Consistently, in the context of AGL, Turner and Fischler (1993) found strict response deadlines had a greater impact on classification accuracy when participants had been instructed to search for grammar rules during training (thought to maximise explicit learning) compared to those who simply memorised training sequences (thought to minimise explicit learning).

However, studies using the remember-know methodology (R/K; Tulving, 1985) have provided contradictory evidence, finding that in self-paced tests, fully recollective responses are made more rapidly than those based on just familiarity (the R/K method involves subjects reporting on the phenomenology associated with recognition responses, with remember – R – responses indicating recollection, know – K – responses indicating just a feeling of familiarity, and guess – G – indicating no feeling of memory at all). For example, Dewhurst, Holmes, Brandt, and Dean (2006) found that after participants had studied word lists, subsequent ‘remember’ responses to test stimuli were made most rapidly, followed by ‘know’ responses (indicating familiarity without conscious recollection) then ‘guess’ responses (see also Dewhurst and Conway, 1994, Dewhurst et al., 1998, Henson et al., 1999, Konstantinou and Gardiner, 2005). Dewhurst et al. (2006) concluded that RTs reflect the time taken to make a decision based on the recollection or familiarity process and RTs may not reflect the actual retrieval process per se (i.e. the time differences in responses may be based on the information afforded to metaknowledge by R, K or G processes). The R/K method as applied to memory presumably taps similar processes as recollection and familiarity in AGL, though the interpretation is not exactly the same. R and K both involve conscious knowledge that an item was presented before, but familiarity- or rule-based -responses in AGL are not a commitment to an item having been presented before, but to the item being grammatical or ungrammatical. R responses are analogous with rule or recollection responses in AGL; both are dependent on consciously recognising that the item, or parts of the sequence, had been presented previously. K responses reflect a feeling, without conscious recollection, that the item had been presented previously and similarly in AGL familiarity responses reflect feelings of oldness of parts or aspects of a stimulus, without consciously recognising the parts of the sequence leading to that conclusion. However, in the current experiments the focus is on guess responses. A ‘guess’ response in memory occurs in the absence of conscious judgement of whether the stimulus had appeared in training, and in AGL reflects the absence of conscious judgment of whether the sequence follows the grammatical structure from training (see Tunney, 2007, for implementation of the R/K methodology in an AGL task).

Wixted and Mickes (2010) argue that guess responses are made in R/K studies when the memory strength of a particular test item falls on the old/new decision criterion, in a similar manner to guess responses in AGL where the subjective familiarity of a particular test sequence is close to the subjective mean value acquired over the course of the experiment (Scott & Dienes, 2008). Gardiner, Ramponi, and Richardson-Klavehn (2002) concluded from a meta-analysis that the performance of guess responses in R/K studies is typically at chance levels. This means that longer RTs for guess responses could sometimes reflect a lack of knowledge. This is not true of AGL studies where the guessing criterion is often satisfied (e.g.: Dienes and Altmann, 1997, Dienes and Scott, 2005, Dienes et al., 1995, Scott and Dienes, 2010b, Scott and Dienes, 2010c, Tunney and Shanks, 2003), making it an ideal paradigm to investigate the time-course of how knowledge is expressed without conscious awareness of that knowledge. Thus, the primary aim of this paper is to validate the no-loss gambling method by showing it distinguishes responses made with confidence and those made without confidence, by testing the apparent contradiction between hypothesised rapid unconscious responses and the results of standard R/K studies in accordance with dual process theory.