The doxastic shear pin: delusions as errors of learning and memory

Abstract

We reconsider delusions in terms of a “doxastic shear pin”, a (cognitive) mechanism that errs so as to prevent the destruction of the (brain) machine and to permit continued function at an albeit attenuated capacity. Delusions may disable flexible (but energetically expensive) inference, leaving the increasingly habitual delusion to flourish. With each recall, delusions are reinforced further, strengthened, and made more resistant to contradiction. Psychodynamic theories posit benefit from delusions in the form of defensive function. We aim to respond to deficit accounts of delusions – that delusions would be only a problem without any benefit – by considering delusion formation and maintenance in terms of predictive coding.

We posit that brains conform to a simple computational principle: to minimize prediction error (the mismatch between prior top-down expectation and current bottom-up input) across hierarchies of brain regions and psychological representation. Recent data suggests that delusions may form in the absence of constraining top-down expectations. Then, once formed, they become new prior expectations (priors) that motivate other beliefs, perceptions, and actions by providing strong (sometimes overriding) top-down expectation.

We argue that delusions form when the shear-pin functions correctly – by breaking. This permits continued engagement with an overwhelming world, and ongoing function in the face of what was a paralyzing difficulty. This crucial role should not be ignored when we treat delusions: we need to consider how a person will function in the world without them.

I. Introduction

Delusions are fixed false beliefs that are considered cardinal symptoms of schizophrenia (Garety, 1992). They can be very distressing for the people who espouse them. Therefore, it may seem mischievous to consider potential adaptive functions of delusions (Hingley, 1992; Roberts, 1991). Dennett and McKay have explored adaptive misbeliefs – incorrect beliefs that, though wrong, may be helpful (R. T. McKay & Dennett, 2009). They argued that only positive illusions are adaptive misbeliefs. Psychodynamic theory holds that some psychotic beliefs serve a defensive function – for example, grandiose beliefs protect against low self-esteem (Neale, 1988). Evolutionary theorists have made similar claims: if a person convinces himself that he has important insight, when he shares that insight, he may gain social status (Hagen, 2008).

Dennett and McKay introduced a helpful metaphor – the doxastic shear-pin – a mechanism by which a misbelief serves to protect its author (R. T. McKay & Dennett, 2009).

We extend the doxastic shear-pin to delusion formation and maintenance, describing a mechanism in terms of aberrant predictive coding. These processes are themselves adaptive, allowing exploitation of environmental contingencies and flexible responding when those contingencies change.

We discuss the biological underpinnings of a shear-pin model, conceiving of beliefs as akin to stimulus-response habits that become resistant to contradictory evidence through overtraining.

We relate our approach to the philosophical concept of epistemic innocence – delusions provide an explanation for ineffable experiences, and are the best conclusion given the available data. Finally, an adaptive function of delusions would imply that with treatment, the adaptation the delusions provide could be lost. We will discuss the relevance of these ideas for how we treat people with delusions.

II. An example: the Capgras delusion

People with the Capgras delusion report that previously familiar people have been replaced by imposters (Capgras, 1923). Although these imposters look familiar, they do not evoke the feelings of recognition that characterize familiar people (M. Coltheart, Langdon, & McKay, 2007). We hold in mind this example as we consider the potential for adaptive function of delusions.

III. The doxastic shear pin

For McKay and Dennett, delusions might be adaptive, but in a psychological (i.e. wishful thinking) not biological sense. So for McKay and Dennett, delusions are not adaptive misbeliefs (R. T. McKay & Dennett, 2009). In response, we suggested that delusions might be biologically adaptive (A. L. Mishara, Corlett, P.R., 2009).

McKay and Dennett introduced the doxastic shear-pin (R. T. McKay & Dennett, 2009). In engineering, shear pins are built into systems to disable machines in trouble so that continued functioning does not destroy them. A broken shear-pin allows continued function, albeit at an attenuated level. We argue that delusions form when the doxastic shear-pin breaks. This approach is similar to psychological description of defenses and biases (R. T. McKay & Dennett, 2009). Sometimes wrong beliefs (like overconfidence in one’s abilities) can confer an adaptive advantage, e.g. when the benefits of winning contested resources outweigh the costs of competition (D. D. Johnson & Fowler, 2011). We argue that a broken doxastic shear pin (i.e. delusion formation) allows some continued engagement with the world, rather than no action at all (A. L. Mishara, Corlett, P.R., 2009).

IV. Learning and delusions

We have previously explained delusions as erroneous causal inferences (P. R. Corlett, Honey, & Fletcher, 2007; Hemsley, 1994; Miller, 1976). We focus on prediction errors (PEs): teaching signals that mark discrepancies between expected and actual events. In the 1960s, engineers working in neural networks (Widrow, 1960) and experimentalists working in animal models (R. A. Rescorla, Wagner, A.R., 1972) similarly accounted for new learning with PEs. As contingencies change, PEs update associations between cause and effect. Updating can occur directly by changing association strength, or indirectly by allocation of attention (Pearce & Hall, 1980). Indeed, stimuli that engender PEs garner more subsequent attention (Hogarth, Dickinson, Austin, Brown, & Duka, 2008).

The brain works to minimize uncertainty. It maintains a set of predictive associations (based on past experience) that is flexible enough to adapt, yet robust enough to avoid superstitions and instabilities (Friston, 2005b, 2009). PE minimization occurs at all levels from the single neuron (Fiorillo, 2008) up through the hierarchical neuroanatomy (Friston, 2005b, 2009). Expectations based on established associations are communicated from areas with more abstract representations downwards through the hierarchy (Mesulam, 2008). PEs are either cancelled by top-down expectancy (i.e. something unexpected is ignored), or propagated and used to update associations (i.e. new learning) (Friston, 2005b, 2009). Whether a PE is discarded or incorporated depends on precision – precise (consistent) errors drive new learning, and imprecise errors are less likely to garner updates. Precision is signaled by specific slow neuromodulators dedicated to each inference (e.g., acetylcholine for perceptual inference, dopamine for motor inference). These neuromodulators are implicated in the pathophysiology of psychosis (Adams, Stephan, Brown, Frith, & Friston, 2013; Friston, 2005a).

This model of mind/brain function and dysfunction, which is committed to veracity, may seem at odds with the generation of psychotic symptoms like hallucinations and delusions (P. R. Corlett, Taylor, Wang, Fletcher, & Krystal, 2010a). How can the complex and strongly held misbeliefs that characterize psychotic illness arise from a truth-seeking system? We know from behavioral economists that humans can depart from responses that minimize punishment and maximize reward (Kahneman, 1982). Can predictive coding depart likewise? We think so. For example, Bayesian models of message-passing in crowds can recapitulate the rumors and panic that arise after a salient event, such as a major disaster (Butts, 1998). We (and others) posit that delusions may, similarly, be explained within the predictive learning framework (P. R. Corlett et al., 2010a). Aberrant PE brain signals and attention to irrelevant stimuli have been associated with delusions in patients (Corlett et al, 2007) and delusion-like beliefs in controls (R. Morris, Griffiths, Le Pelley, & Weickert, 2013) (Le Pelley, Schmidt-Hansen, Harris, Lunter, & Morris, 2010).

Learning about environmental reward contingencies allows organisms to explore and exploit their environments (A. Dickinson, Shanks, D.R., 1995). The neurobiological changes in early schizophrenia, however, disrupt this learning (Murray et al., 2008). How, then, might delusions contribute to survival?

The effects of motivation have been challenging for simple rules to minimize prediction error (Dayan & Balleine, 2002). To reconcile the effects of satiation on subsequent responding, researchers have posited multiple instrumental controllers that compete to guide action choices. These are thought to work through a balance between goal-directed and habitual learning (Daw, Niv, & Dayan, 2005) (Hitchcott, Quinn, & Taylor, 2007). Goal-directed learning involves flexible action-outcome relationships based on computations in prefrontal cortex (Daw et al., 2005). This learning is sensitive to rapid changes in outcome value (Dayan & Balleine, 2002). Habits, more rigid representations of stimulus-response relations, are encoded in striatum (Daw et al., 2005). A recent account of reinforcement learning actually suggests that the goal-directed system is associated with processing higher up the Bayesian hierarchy. This depends on bootstrapping from simpler habitual reflexes lower down (Pezzulo, Rigoli, & Friston, 2015), and as mentioned above, depends on precision of expectations (Daw et al., 2005) – control is ceded to the most precise level (Pezzulo et al., 2015).

Delusions may arise because goal-directed learning (at the highest levels of the hierarchy) is impaired (Gold et al., 2013; Gold et al., 2012; Gold, Waltz, Prentice, Morris, & Heerey, 2008). Simpler associative mechanisms (lower in the hierarchy) might drive ongoing but less flexible instrumental engagement (A. L. Mishara, Corlett, P.R., 2009).

We propose a single impairment in PE (across the hierarchy) in three stages:

1. Delusional mood. In the prodrome, attention is drawn to irrelevant stimuli: people report feeling uncomfortable and confused (Kapur, 2003a; McGhie & Chapman, 1961). This may reflect inappropriate PEs. Functional neuroimaging studies of drug-induced and endogenous early psychosis reveal PEs in frontal cortex in response to unsurprising events, and PE intensity correlates with delusion severity (P. R. Corlett et al., 2006; P. R. Corlett, Murray, et al., 2007).

   During the prodrome, the stress-mediator cortisol increases by up to 500% (Sachar, Mason, Kolmer, & Artiss, 1963). Heightened stress impairs goal-directed learning and promotes habit formation (Schwabe & Wolf, 2009).

2. Delusion formation. In response to prodromal confusion and stress, the “doxastic shear-pin” breaks. Delusions appear in an aha-moment, when explanatory insight occurs and flexible processing is disabled. Habitual responses are preserved and possibly even enhanced (P. R. Corlett, Krystal, et al., 2009; P. R. Corlett et al., 2010b). Cortisol falls as delusions crystalize (Sachar et al., 1963). Forming the delusion is associated with ‘insight relief’ that helps consolidate it in memory (Miller 2008; Tsuang et al. 1988). Cortisol rises once more as delusions conflict with reality (Sachar et al., 1963). As people recover and relinquish their delusions, cortisol responses normalize (Sachar et al., 1963).

3. Explaining things with the delusion. As in overtraining during instrumental learning, the delusion becomes increasingly habitual as it is used (P. R. Corlett et al., 2010b)). In the shear-pin metaphor, delusions are an adaptive product of the shear-pin break. They enable patients to stay engaged with the environment and to exploit its regularities, though the patient may be inflexible and unresponsive to corrective feedback (P. R. Corlett et al., 2010b)).

Delusion formation, associations, and psychodynamic motivation

In a psychodynamic frame, some symptoms arise from conflict between conscious and unconscious motivation. For example, someone with hidden feelings of social failure may develop paranoia, sensing that many people are so interested in him that they constantly intrude and observe him (Lyon, Kaney, & Bentall, 1994). A man with conflicted feelings about his wife may suffer the Capgras delusion (R. McKay, Langdon, & Coltheart, 2005).

Some have considered psychodynamic processes in information processing terms (Bowlby, 1980; Pally, 2005, 2007). Cognitive 2-factor explanations posit that both a perceptual dysfunction (Factor 1) and a belief evaluation deficit (Factor 2) are necessary for delusions to form. McKay and colleagues suggested that motivational processes could influence Factor 2 (R. McKay, Langdon, & Coltheart, 2007), i.e. that wishful thinking changes belief evaluation. On the other hand, factor 1 is a further possibility, as people may actually sense things differently due to motivated biases (R. T. McKay & Dennett, 2009).

In our model (and those before us (Helmholtz, 1878/1971)), the two factors are strongly inter-related (P. R. Corlett & Fletcher, 2015). Differentiating top-down (belief) from bottom up (sensation) effects is a challenge, since, in a generative system, top-down and bottom-up processes sculpt one another. Learned biases can alter perception. We see illusory stimuli that conform to our expectations rather than the sensory data incident on the retina (Pearson & Westbrook, 2015). Regina Pally also implicates top-down effects in her analysis of the relationship between PE and psychodynamics (Pally, 2005, 2007). She posits that predictions are responsible for the recapitulation of harmful early-life relationships during adulthood (Pally, 2005, 2007).

Computational psychiatry and predictive coding have engaged directly with psychoanalysis (Carhart-Harris & Friston, 2010). In reviewing data about confabulation (memory errors that fill in gaps in memory), Aikaterini Fotopoulou notes that, consistent with ideas of motivated self-deception (Hagen, 2008), many confabulations are positively valenced, i.e. aligned with the patient’s wants, and involving limbic reward processing regions (Fotopoulou, 2010).

Self-deception is relevant to delusions. It entails simultaneously believing some proposition (p) and its antithesis (not-p). Subjects may be psychologically motivated to state one belief but act according to another (Harold A Sackeim & Ruben C Gur, 1979). This is relevant to the double-bookkeeping in which some delusional patients engage (see below (L.A. Sass, 1994)).

In one laboratory based self-deception task, subjects first predict an uncertain outcome, then describe the outcome when they see it. Some subjects (self-deceivers) stick with their initial prediction even when presented a contrary outcome (as if they can’t see what’s right in front of them). They are more likely to engage in this deception when incentivized for correct prediction (Mijovic-Prelec & Prelec, 2010). To explain this finding, the Prelecs call on an actor-critic model, such as those proposed to explain instrumental learning with PE (Sutton, 1998). By this account, the mind is organized into multiple interacting agents, each operating on different information and each revealing different outputs to its co-agents. The actor chooses an action and the critic gives that action a score. The critic tries to learn the actor’s policy and the actor tries to get the best possible score (perhaps even better than it deserves). This architecture portends self-deception – the actor tries to fool the critic (Mijovic-Prelec & Prelec, 2010). In reinforcement learning applications of actor-critic, the critic learns the environmental states and the actor learns an action policy given those states. Prediction errors update the actor and the critic. Actor and critic have been localized to different striatal sub-regions (actor – dorsal striatum, critic – ventral (O’Doherty et al., 2004)). If actor and critic are disconnected in computational models, the critic no longer trains the actor, sensitized reward responses ensue, and behavior becomes inflexible (Takahashi, Schoenbaum, & Niv, 2008). Though previously focused on cocaine addiction, this model might explain the genesis and maintenance of self-deception and delusions. However, we advise caution. The actor-critic architecture of reinforcement learning may not map neatly onto the Prelecs’ model. Whether it does is an empirical matter.

Gur and Sackeim examined self-deception using galvanic skin response (GSR) to mark the progress of conditioning. GSR is a metric of salience. The skin sweats more in response to, or in anticipation of, salient events. In their examination of self-deception, Gur and Sackeim played recordings to their subjects of the subjects’ own voice and others’ voices. Subjects were asked if the voice was their own or another person’s. People show increased GSR to their own voice. Here, self-deception is defined as saying the voice belongs to someone else, despite increased GSR signaling which would suggest that the subject actually finds it familiar. People are more likely to self-deceive in the lab if they also endorse self-deceptive statements like “I have never lied” or “I have never stolen,” which are unlikely to be true (H. A. Sackeim & R. C. Gur, 1979).

Learning theorists often use GSR to assay predictive learning in human subjects. As conditioning progresses, GSR tracks conditioned cues rather than the outcomes they predict. However, GSR changes in just one trial with reassuring instructions (e.g. “That cue is now safe”); what was learned over multiple trials is immediately extinguished (Lovibond, 2003, 2004). Perhaps for GSR conditioning, conscious expectation of the outcome must develop to the cue (Dawson & Furedy, 1976) although there are also contradictory data (Schell, Dawson, & Marinkovic, 1991).

Consider also the Perruchet effect (Perruchet, 1985). This is the observation that while conscious predictions demonstrate a gambler’s fallacy (treating independent pairings as non-independent – “If I haven’t had a salient event following the cue in a while, then one must be coming”), GSR responses do not (McAndrew, Jones, McLaren, & McLaren, 2012). There is a dissociation between expressed belief and skin conductance, which is perhaps a metric of unconscious belief. This implies that there are multiple learning systems representing conflicting beliefs. Those systems may be responsible for the self-deception examined by Gur, Sackeim, and the Prelecs.

Clearly, the number and representational nature of learning systems in people and other animals is still a topic of active inquiry (Mitchell, De Houwer, & Lovibond, 2009). Furthermore, although we point out the methodological and inferential overlap between self-deception studies and studies of conditioning and expectation, this overlap may only be superficial. Psychodynamic notions of self-deception may not align with associative learning theory (although they may (Bowlby, 1980)). Whether they do will be best settled with new data rather than argument from the extant literature.

A. Working around delusions

Our analysis centers on the role of delusions in sustaining action. However, not all delusions motivate action. Some people double-bookkeep, claiming a delusional belief and yet acting otherwise (Bleuler, 1908; L. A. Sass, 1994). Someone might claim their food is being poisoned yet continue to eat (L.A. Sass, 1994). Bortolotti and Broome appeal to biological motivation here. Patients more afflicted by negative symptoms, who lack motivation and whose prospective cognition is impaired, may be less likely to act on their delusions (L. Bortolotti, Broome, M., 2012).

On the other hand, acting in spite of one’s delusions could be biologically adaptive (maintaining nutrition, housing, etc.). Dickinson and Balliene (1993) accounted for the impact of incentive motivation on instrumental responding in animal learning in their associative cybernetic model. These authors sought to explain how motivational states invigorate action in some situations but not others. According to their account, instrumental behavior is invigorated by a motivational system that codes the current value of an outcome (e.g. a reward) – thus allowing for the effects of changes in outcome value by satiation. This system may be instantiated in the orbitofrontal cortex (OFC) and ventral striatum (Daw et al., 2005), both structures associated with amotivation (Lebreton et al., 2009). The interaction of this motivational system with learning systems may dictate the degree to which delusions are acted upon (L. Bortolotti, Broome, M., 2012). For example, delusion severity is associated with disrupted OFC responses during Pavlovian-to-Instrumental transfer (R. W. Morris, Quail, Griffiths, Green, & Balleine, 2015). Normally, stimuli with learned Pavlovian incentive value invigorate instrumental responding (R. A. Rescorla & Solomon, 1967). However, in patients with delusions, who attribute salience inappropriately to irrelevant cues (P. R. Corlett, Murray, et al., 2007; Kapur, 2003a), this transfer seemed to occur even with irrelevant events, such that they too drove instrumental action (R. W. Morris et al., 2015). Future computational modeling work may well discern the causal models being inferred during Pavlovian to Instrumental Transfer (Cartoni, Puglisi-Allegra, & Baldassarre, 2013), and how those models may be perturbed in patients with delusional beliefs.

B. Delusion maintenance

While many delusions have upsetting content, they may nonetheless ease the overwhelming chaos of the prodrome (Kapur, 2003b). They serve to infer the best explanation for that chaos (M. Coltheart, Menzies, & Sutton, 2010). Delusions are also remarkably elastic: they expand and morph around contradictory data (Garety, Hemsley, & Wessely, 1991; Milton, Patwa, & Hafner, 1978; Simpson & Done, 2002). Of note, patients can learn about other new things (they don’t have an all-encompassing learning deficit) and can even critically evaluate others’ delusions (Rokeach, 1964). However, once a delusion is formed, subsequent PEs are explicable in the context of the delusion and serve to reinforce it (P. R. Corlett, Krystal, et al., 2009; P. R. Corlett et al., 2010b). Hence the seemingly-paradoxical observation that challenging subjects’ delusions can actually strengthen their conviction (Milton et al., 1978; Simpson & Done, 2002).

The illusory truth effect is relevant here – merely having considered a proposition enhances judgment of its veracity in the future (Begg, Anas, & Farinacci, 1992). Patients with delusions are particularly susceptible to this effect (Moritz, 2012). Similar effects have been observed with conditioned memory reactivation in rodents. Merely reactivating a fear-conditioned context (by reminding animals of prior fear memories) can strengthen future responding (Lee, 2008). We recently showed that reactivating fear memories in rodents and humans on ketamine enhanced subsequent memory strength (Corlett et al, 2013; Honsberger et al, 2015). Similarly, backfire effects are observed in politics (Bullock, 2009) and science (McRaney, 2013), whereby data that clearly contradict a cherished belief strengthen rather than weaken it. When a belief does crucial explanatory work, contradictory data may strengthen the belief by engaging it and reconsolidating it more strongly, much like the habitization associated with instrumental overtraining (P. R. Corlett, Krystal, et al., 2009; P. R. Corlett et al., 2010b).

V. Capgras and PE

Let’s return to the Capgras delusion. Cognitive neuropsychiatric (Halligan & David, 2001) explanations of Capgras (and other delusions) range from single factor (B. A. Maher, 1974), to two-factor (Max Coltheart & Davies, 2000) to interactionist (Young, 2008). The single factor account appeals to a deficit in perception of familiarity; the delusion formation process being a reasonable or sensible consequence of such an unsettling experience (B. A. Maher, 1974; B.A. Maher, 1988b). Two-factor theorists appeal to deficits in both familiarity processing and belief evaluation such that the unlikely explanation (“My loved one has been replaced”) is favored over the actual explanation (“Something is wrong with my brain”). We attempt some rapprochement between prediction error, 2-factor theory, and psychodynamic motivational accounts of delusions. Under motivational explanations, people concocted imposter beliefs like Capgras (for example) to cover up new (and guilty) lack of affection while maintaining a sense of (good) self (R. McKay et al., 2007). We suggest that PE theory can also incorporate such functionality.

Recall that several features are part of the PE account: 1) through experience we learn to expect a certain set of circumstances, 2) the consequences when these expectancies are violated, which are either to discard the dissonant experience or to incorporate it by updating the set of expectancies. Absent but expected events are crucial in the phenomenology of Capgras (M. Coltheart et al., 2007) and perhaps delusions more broadly (L. Sass & Byrom, 2015). When confronted with someone who resembles a loved-one, we 1) expect to feel familiarity, and 2) generate a PE when that feeling is absent (P. R. Corlett & Fletcher, 2015; P. R. Corlett et al., 2010b). We suggest that the continued aberrant PE leads to Capgras (P. R. Corlett et al., 2010b), since sustained PEs call for a new explanatory belief (P. R. Corlett, 2015).

PE guides the Capgras sufferer toward a particular conclusion: that a replica has replaced their loved-one. Two-factor theorists argue that this explanation is so irrational that it must require a deficit in belief evaluation. They underline reports of patients who lack familiarity responses for their loved-ones, but who never endorse delusional explanations (Tranel & Damasio, 1985). It may be that the imposter belief is not particularly unlikely (R. McKay, 2012). Furthermore, PEs have been invoked to explain selection between beliefs (Waldmann, 1998) (FitzGerald, Dolan, & Friston, 2014), so we suggest that PE dysfunction could lead to deficits in both factors 1 and 2 (P. R. Corlett, Fletcher, P.C., 2014). Predictive coding theory (and its application to psychosis) does not draw as strong a distinction between perception and belief as do 2-factor theorists (P. R. Corlett, Fletcher, P.C., 2014). Top-down expectations may sculpt perception to conform with priors (as is observed with perceptual illusions – (Gregory, 1996)). Physiological motivation can also alter top-down perception – hungry subjects perceive food images in noise (Sanford, 1937). Psychological motivation can do likewise – poorer children perceive coins as physically larger than do their wealthier counterparts (Bruner & Goodman, 1947) – although these particular studies did not replicate and the New Look approach to perception fell out of favor (Erdelyi, 1974). There are top-down attentional effects from prefrontal and parietal regions onto sensory cortices that may mediate some of these effects (Firestone & Scholl, 2015) and have relevance to delusions (Dima et al., 2009; Schmidt et al., 2012). Psychodynamic defenses might likewise alter the influence of top-down priors (R. McKay et al., 2007) so that believing changes seeing through effects on attentional allocation. Furthermore, knowledge (based on prior experience) may penetrate perception more so in people with psychosis than those without (Teufel et al., 2015), adding credence to our proposal that top-down priors may be overly engaged and sculpting perception inappropriately in psychosis. But, again, the penetrability of perception by belief is still a subject of ongoing debate that has yet to be resolved empirically (Firestone & Scholl, 2015).

VI. Why this particular belief?

Delusions are often socially relevant: they are ideas about oneself in relation to others. Their content is crucially related to an individual’s specific concerns (Reed, 1972). Recently, social learning has been analyzed in terms of predictive coding (Behrens, Hunt, Woolrich, & Rushworth, 2008). We form beliefs about others and make predictions about their future actions using PE (Behrens et al., 2008; King-Casas et al., 2008). Phenomena from simple associative learning, like Kamin blocking (Kamin, 1969), have also been demonstrated in social learning –learning that one worker is productive blocks the attribution of productivity to a new worker (Cramer et al., 2002). Key questions for future research include whether there is neural circuitry dedicated to social learning or whether social learning draws more extensively upon canonical predictive circuitry because social inference is difficult and computationally intensive (Heyes & Pearce, 2015). We note with interest that voltage gated calcium channels have been implicated in the genetic risk for psychotic illnesses in genome wide association studies (Jiang et al., 2015) and may contribute to prediction error signaling (Liu et al., 2014) and to social learning in rodent models (Jeon et al., 2010).

Also important are the particulars of past and present experience, including personal and cultural context. Cold war era persecutory delusions commonly involved KGB agents (Kihlstrom, 1988). Delusions of reference have evolved from worries about radio or satellite monitoring to concerns about the internet, from The Matrix to Transcendence (Stompe, Ortwein-Swoboda, Ritter, & Schanda, 2003) (Gold et al., 2012).

VII. Delusions and Epistemic Innocence

Epistemically innocent (L. Bortolotti, 2015) beliefs are defined as those that confer adaptive advantage by increasing knowledge, and that are based on the available data. Delusions decrease uncertainty, providing a new explanatory framework for knowing. Importantly, in doing so, they allow the person with a delusion to re-engage on some level with the otherwise too chaotic world. Also, others might have access to contradictory evidence, but the authors of delusions often do not.

What of epistemic innocence and associationism? Unexplained PE or uncertainty is stress inducing. We do not like being surprised. Explaining surprise so that events can be better predicted in the future drives belief formation. If the belief is wrong, or even delusional, it still explains away the uncertainty.

Computational modeling of learning and perception allows us to test the consequences of specific changes in a model learner (Stephan & Mathys, 2014). These experiments may shed further light on the requirements for epistemic innocence in a model where fine manipulations are possible. For example, some models produce biases, e.g. spreading of erroneous rumors in a social network (Butts, 1998), tendency to ignore base rates when making probabilistic decisions (Soltani & Wang, 2010), and habit formation (FitzGerald et al., 2014).

One relevant example is the confirmation bias (Lord et al, 1979; Nickerson, 1998), in which prior beliefs bias current decision making, specifically, contradictory data are ignored if they violate a cherished hypothesis. The confirmation bias has been tied to striatal PE learning through theoretical (Grossberg, 2000) and quantitative computational models (Doll, Jacobs, Sanfey, & Frank, 2009) as well as genetics (Frank, Moustafa, Haughey, Curran, & Hutchison, 2007; Heyser, Fienberg, Greengard, & Gold, 2000)) (Doll et al 2009). Of interest to this discussion, confirmation bias is increased in individuals with delusions (Balzan, Delfabbro, Galletly, & Woodward, 2013). However, patients with chronic schizophrenia do not show an enhanced fronto-striatal confirmation bias (Doll et al., 2014) – the relationship with delusions in particular was not examined. It is possible that confirmation biases are specific to delusion contents (encapsulated) rather than a general deficit. Woodward and colleagues showed delusion-related confirmation biases (Balzan et al., 2013). At first, it is hard to think that maintaining a belief in the face of contradiction could be adaptive. However, Boorstin (1958) has argued that confirmation bias permitted the 17th-century New England Puritans to prosper: they had no doubts and allowed no dissent, so were freed from religious argument, and more able to focus on practical matters. Their doctrine was so clear and strongly held that they had an all-encompassing explanation. As in this example, confirmation bias may save energy and allow work on more pressing tasks. Confirmation bias also protects ones’ sense of self as a person with a consistent and coherent web of beliefs living in a predictable world.

VIII. Implications for treatment

In the normative approach to defining delusions that dominates most cognitive neuroscience approaches, delusions are conceived of as symptoms to be eradicated. The recovery movement has campaigned for people who experience mental health symptoms to be able to advocate for themselves, to improve their care, and to have more influence on clinicians and researchers alike. In describing her own experience of delusions, Amy Johnson has invoked WB Yeats – suggesting that clinicians and scientists ought to tread lightly in their work, as they tread on patients’ delusions (A. Johnson, Davidson, L., 2013). This is perhaps the most resounding endorsement of the epistemic innocence of delusions – Amy’s delusions matter to her.

The non-clinical situations in which people with radically different beliefs clash may be instructive. For example, confronting anti-vaccination believers with contrary evidence can backfire, strengthening their conviction that vaccines are harmful (Nyhan & Reifler, 2015; Nyhan, Reifler, & Ubel, 2013). We, and others, have argued that delusions are often grounded in personal experiences; “I know it sounds crazy, but I saw it with my own eyes, Doctor”). Likelihood of relinquishing beliefs on the basis of others’ testimony is strongly correlated to the credibility of the source (Nyhan et al., 2013). Do the people trying to change one’s mind have a vested reason to disagree? Perhaps large-scale anti-stigma educational activities in mental health would benefit from including more people with lived experience to spread the word about mental illness (Corrigan, 2012).

In summary, delusions can engender significant suffering and distress. However, in addition to the problems they can bring, delusions form through neurobiological and psychological efforts to adapt (to learn from prediction errors, to use defenses to maintain a functioning self). Delusions are epistemically innocent: new knowledge that may be flawed, but nonetheless useful, and based on the data available to the author. We have emphasized that delusions may not be so different from non-delusional beliefs, including religious ideas, political affiliations, and scientific theories. Delusions formed by adaptive reinforcement learning processes can become cherished and difficult to replace (B. A. Maher, 1974). Clinicians would do well to consider the utility of delusions both in terms of effective approaches to change and compassionate work with the people who have made these beliefs.