Aha! under pressure: The Aha! experience is not constrained by cognitive load

Highlights

• The (un)conscious nature of insight was assessed by manipulating cognitive load.

• Under high cognitive load, non-insight solutions required more time than insightful ones.

• As cognitive load increased, non-insight solutions became less frequent.

• Insight solutions remained mostly unaffected by cognitive load.

• Insight solutions appeared not to compete for limited cognitive resources.

Abstract

The Aha! moment– the sudden insight sometimes reached when solving a vexing problem– entails a different problem-solving experience than solution retrieval reached by an analytical, multistep strategy (i.e., non-insight). To date, the (un)conscious nature of insight remains debated. We addressed this by studying insight under cognitive load. If insight and non-insight problem solving rely on conscious, effortful processes, they should both be influenced by a concurrent cognitive load. However, if unconscious processes characterize insight, cognitive load might not affect it at all. Using a dual-task paradigm, young, healthy adults (N = 106) solved 70 word puzzles under different cognitive loads. We confirmed that insight solutions were more often correct and received higher solution confidence. Importantly, as cognitive load increased, non-insight solutions became less frequent and required more solution time, whereas insightful ones remained mostly unaffected. This implies that insight problem solving did not compete for limited cognitive resources.

Introduction

The “Aha! experience” — that moment when the solution to a vexing problem suddenly pops into consciousness — has mesmerized scientists and laymen alike (Bowden & Grunewald, 2018; Chein & Weisberg, 2014). Scientifically, this singular subjective experience is referred to as insight (Bowden & Grunewald, 2018). While insight is not rare, most problem-solving involves a multistep analytical strategy (Simon & Newell, 1970; Weisberg, 2015) through which the problem solver searches long-term memory for potential algorithms, schemas, analogies or factual knowledge. This process is referred to as analysis or non-insight (e.g., Bowden, Jung-Beeman, Fleck, & Kounios, 2005; Fleck & Weisberg, 2013). It requires the problem solver to monitor his/her progress while maintaining the mental representation of the problem and avoiding being distracted by irrelevant information (see Shipstead, Harrison, & Engle, 2016; Wiley & Jarosz, 2012, for a review).

In contrast to insight, which feels effortless, non-insight solution-finding relies on continuous, effortful processing (e.g., Metcalfe & Wiebe, 1987). The difference between effortless and effortful processing also forms the core of dual-processing theories (Evans & Stanovich, 2013; Gilhooly, Ball, & Macchi, 2015; Sowden, Pringle, & Gabora, 2015). In such theories, Type 1 processes are assumed to be unconscious, effortless, and not limited by working memory (WM) limits. Type 2 processes, on the other hand, are taken to be conscious, effortful, and constrained by WM limits (Evans, 2019; Evans & Stanovich, 2013). For example, it has been shown that if participants have to make judgments while also performing an unrelated secondary task known to deplete cognitive resources, intuition-based judgments (Type 1) mainly stay unaffected, but deliberate-thought-based judgments are hampered (Type 2; De Neys, 2006; Howarth, Handley, & Walsh, 2016). This observation indicates that Type 1 processes are relatively independent of the cognitive resources needed to consciously manipulate information, whereas Type 2 processes depend more on cognitive resources (see Evans, 2019 for a review). Although it is intuitively appealing to consider insight resulting from Type 1 and non-insight from Type 2 processes, this is not a widely held claim (Benedek & Fink, 2019). For instance, Weisberg (2015) argues that insight, just like non-insight, mostly relies on the same effortful (Type 2) processes to retrieve a solution.

Can the conception of Type 1 and Type 2 processes proposed in dual-processing theories be translated to insight and non-insight problem solving? Dual-processing theories depart from a default-interventionist model (Evans, 2019; but also see De Neys, 2021; Mega & Volz, 2014). Type 1 processes entail default, automatic intuition-based responses resulting from automated processes (e.g., locating sounds), stereotypes, beliefs, or automated skills (e.g., reading; Howarth et al., 2016). Therefore, these intuition-based responses have been regarded as heuristically driven, helping humans to navigate life efficiently (e.g., mindlessly driving to work; Kahneman, 2011). However, sometimes cues in the environment can erroneously trigger an intuition-based response (e.g., stereotype-based judgments), defying the logic of the task at hand (De Neys & Pennycook, 2019). In such cases, it has been found that humans tend to make wrong intuition-based choices by failing to override them by deliberate thought (Type 2) to solve the task (Frey, Johnson, & De Neys, 2018). Humans often opt for the path of least resistance and follow their error-prone intuitions by default, even though these can be wrong (Evans, 2019).

At first sight, the conception that Type 1 processes lead to error-prone intuitions seems perpendicular to how insight solutions have been perceived. For instance, Salvi, Bricolo, Kounios, Bowden, and Jung-Beeman (2016) showed across different types of problems that insight solutions were more often correct than non-insightful ones. Numerous other studies have corroborated this higher solution-accuracy effect of insight (e.g., Danek, Fraps, von Müller, Grothe, & Öllinger, 2014; Hedne, Norman, & Metcalfe, 2016; Webb, Little, & Cropper, 2016). However, the role of intuitive, yet error-prone, Type 1 processes might be different for different phases of the insightful solution search (Zander, Öllinger, & Volz, 2016).

Consistent with representational change theory on insight (Ohlsson, 1992, Ohlsson, 2011), heuristically driven intuitions (Type 1) might initially mislead the problem solver to an ill-defined problem representation, similar to what is proposed in dual-processing theories (Zander et al., 2016). This ill-defined problem representation leads the problem solver astray as content activated from semantic memory will inevitably be insufficient to solve the problem, leading to an impasse (Öllinger & von Müller, 2017). This impasse serves a pivotal function as it propagates negative feedback through the information processing system, hereby decreasing the activation level of the ill-defined problem representation and redistributing the unconscious spreading of activation to more remote, yet unnoticed concepts in semantic memory (Ohlsson, 2011; Öllinger & von Müller, 2017). When this spread of activation converges on the vital concept needed to rectify the initial erroneous problem representation, the solution pops into consciousness as the complete solution path is revealed all at once (i.e., restructuring; Bowden, 1997; Ohlsson, 2011; Öllinger & von Müller, 2017). At this final stage, it is assumed that the problem solver is consciously puzzled but unaware of the unfolding spreading of activation through semantic memory (Ohlsson, 2011). Recently, it has been argued that the problem solver uses the Aha! phenomenology (i.e., positive affect, surprise, obviousness, high confidence) as a heuristic that signals the quality of a solution emerging from the unconscious (Laukkonen, Webb, Salvi, Tangen, & Schooler, 2018). For example, Laukkonen, Kaveladze, Tangen, and Schooler (2020) found that false statements judged on their veracity are considered more true when linked with an Aha! experience than when judged without it (see also Laukkonen, Ingledew, Grimmer, Schooler, & Tangen, 2021). In that sense, Type 1 processes play a double part in insight problem-solving: although they may mislead the problem solver at first, in the end, they signal the quality of the solution surfacing from the unconscious via an Aha-heuristic (Laukkonen et al., 2018; Zander et al., 2016). This idea corresponds with recent research showing that Type 1 processes can culminate into accurate logic-driven intuitions (Bago & De Neys, 2019, see also Bowers, Regehr, Balthazard, & Parker, 1990) and experts' intuitions that can be dead right (see Salas, Rosen, & DiazGranados, 2010, for a review). These observations indicate that the error-proneness of Type 1 intuitions do not need to be irreconcilable with insight's high accuracy, and that Type 1 and Type 2 processes are useful concepts when thinking about insight and non-insight problem solving, respectively.

However, not all insight theorists agree with these assumptions (Benedek & Fink, 2019, Weisberg, 2015, Weisberg, 2018, see also Chater, 2018). Weisberg (2015) assumes that effortful Type 2 processes mainly dominate both insight and non-insight solution finding. Insight is assumed to arise via the problem solver's ability to capitalize on repeated failures to find the solution non-insightfully. The repeated failures bring forth new information to work with, guiding the solution search to new, promising directions that unveil ill-held assumptions about the problem. Once these ill-held assumptions are rectified (i.e., restructuring), the complete solution path is revealed all at once, and the solution is found with insight. Although Weisberg (2015) argued that insight might also be achieved via unconscious processes, he perceived this as a rare phenomenon (Weisberg, 2018).

As the above-theoretical views indicate, the debate concerning insight's (un)conscious nature is not settled yet. One way of debunking whether insight relies on cognitive resources is by assessing how WM relates to insight and non-insight problem solving (e.g., DeCaro, Van Stockum, & Wieth, 2016). WM is considered a central processing hub where information is shortly stored and updated to cope with ongoing task demands (Baddeley, 1986; Shipstead et al., 2016); executive functions associated with problem solving (e.g., Cooper & Marsh, 2015; De Neys, 2006). The capacity of WM is limited (Cowan, 2010). Some have defined this limited capacity in terms of the number of informational chunks that can be hold in the scope of attention (for a review, see Cowan, 2010; Oberauer, Farrell, Jarrold, & Lewandowsky, 2016), whereas others have conceived it as being more closely related to the limits of attentional control processes (e.g., filtering efficiency; see Oberauer, 2019, for an overview).

Thus, WM capacity imposes limits to our ability to maintain and update the mental representations of a problem (Shipstead et al., 2016; Wiley & Jarosz, 2012). Such limits should then clearly influence one's ability to solve complex, multistep problems through non-insight. But what about Insight? If insight is indeed associated with unconscious, Type 1 processes, it should not be bound by the limits of WM capacity (Fleck, 2008; Gilhooly & Fioratou, 2009). However, if it depends on conscious, effortful Type 2 processes, it should also be constrained by WM capacity limitations (Chein & Weisberg, 2014; Chuderski & Jastrzebski, 2018). One way of testing these assumptions consists of using a dual-task paradigm where participants are asked to perform a primary problem-solving task while concurrently executing a secondary WM task (e.g., retaining a series of digits; De Dreu, Nijstad, Baas, Wolsink, & Roskes, 2012). Loading WM in such a way should therefore decrease its capacity and hence its availability to executive functions. Thus, if the primary task also involves WM, then concurrently executing the WM task should hamper performance on the primary task (e.g., Camarda et al., 2018). In contrast, if the primary task does not tax WM, then concurrently executing a WM task should not impact performance (e.g., Abadie, Waroquier, & Terrier, 2013).

A few studies have already explored the role of WM in problem solving. Lavric, Forstmeier, and Rippon (2000) presented participants with classical insight and non-insight problems. Single- and dual-task conditions were created by adding a concurrent tone-counting task for some participants but not for others. The insight problems were correctly solved by a comparable number of participants in both conditions. However, the non-insight problems were correctly solved by a larger number of participants in the single-task conditions, as compared to the dual-task condition. This finding is consistent with the assumption that insight, unlike non-insight, depends on Type 1 processes. While other studies have found similar results (e.g., Korovkin & Savinova, 2014; Korovkin, Vladimirov, Chistopolskaya, & Savinova, 2018), De Dreu et al. (2012), however, found that the number of correctly solved insight problems decreased as the concurrent WM load increased. The authors concluded that insight problem solving depends on WM and hence competes for limited WM capacity. This study and other studies (see also Lin & Lien, 2013; Wieth & Burns, 2014 for similar results) suggests that insight, like non-insight, relies on Type 2 processes.

The inconclusive nature of these findings might originate from their dependence on classical insight and non-insight problems and on the assumption that these problems reliably index insight and non-insight, respectively. However, these two types of problems are very different, making it difficult to draw strong conclusions based on their comparison (Bowden & Jung-Beeman, 2007; Webb et al., 2016). Moreover, it has been shown that insight problems are sometimes solved with non-insight and non-insight problems with insight (Danek, Wiley, & Öllinger, 2016; Webb et al., 2016). Therefore, studies using classical insight problems without a non-insight comparison group might have yielded confounding results (e.g., De Dreu et al., 2012). These findings indicate that insight and non-insight problem solving are difficult to pin down in a specific problem type. Therefore, some have suggested that the problem solver is the most designated person to decide how a solution was found because what sets insight apart from non-insight resides in the problem solver's phenomenology rather than the nature of the problem type (see Bowden & Grunewald, 2018).

To address these issues, researchers have developed problem types that have an almost equal likelihood of being solved with insight and non-insight, keeping the type of problems constant across both solution types. For example, in the compound remote associates test (CRA; Bowden & Jung-Beeman, 2003), participants receive three cue words (break/bean/cake) on each trial and are requested to search for a fourth compound solution word (coffee break/coffee bean/coffee cake). After each solved CRA trial, participants indicate whether they solved the problem with insight or non-insight based on their subjective, solution-finding experience (e.g., Salvi, Simoncini, Grafman, & Jung-Beeman, 2020).

However, this procedure to classify insight and non-insight solutions based on the participant's subjective report has not been without critique (Danek & Salvi, 2018; Laukkonen et al., 2021). The finding that insight solutions are more often correct, receive higher solution confidence, and are solved faster than non-insightful ones (e.g., Cranford & Moss, 2012; Hedne et al., 2016) raises the question if those behavioral or phenomenological indices do not bias the retrospective insight classification. However, strong correspondence has been observed between participants' self-reported insights and physiological indices of insight (i.e., participants' squeeze strength on a dynamometer upon solving a problem; Laukkonen et al., 2021). Moreover, subjective self-reports of insight have been associated with distinct brain-pattern activation (i.e., activation burst across right temporal lobe; see Kounios & Jung-Beeman, 2014), physiological responses (i.e., increased skin conductance, heart rate, and pupil dilation; Salvi et al., 2020; Shen et al., 2017), and phenomenological qualia (feelings of happiness and relief; Stuyck, Aben, Cleeremans, & Van den Bussche, 2021). This body of research provides additional support for the validity of self-reports to study insight and non-insight problem solving.

In the current study, we, therefore, used a problem type (the CRA) that can be solved with insight and non-insight, which was determined based on participants' self-reports. We manipulated WM load by creating no-load, low-load, and high-load conditions. As non-insight problem solving is expected to be an effortful, conscious, Type 2 process, we predicted a detrimental influence of WM load on the performance of the problems solved with non-insight. If insight problem solving relies on an automatic, unconscious, Type 1 process, it should not be influenced by WM load. However, if insight relies on WM in the same way as non-insight, it should be impacted similarly by the WM load.