The elegance of Semantic Cognition: A Parallel Distributed Approach (Rogers & McClelland Reference Rogers and McClelland2004; henceforth Semantic Cognition) is in the explanatory power of several very simple principles. Rogers & McClelland (R&M) show the principles of distributed units with sigmoidal activation functions; varying connection strengths; and predictive, error-driven learning account for a whole slew of nontrivial phenomena from cognitive and developmental psychology and linguistics. This is not entirely a news flash – as the authors point out, they have built upon a strong foundation of past connectionist models in this area. R&M's novel contribution is the sweeping scale of the cognitive phenomena they simulate with a uniquely robust architecture. Indeed, their simulations suggest this is how we know what we know, and the basis for how we reason about the world.

The more I read, the more I wanted R&M to expand beyond their trees, flowers, birds, fish, and mammals, and predict when my children will stop believing in Santa Claus, or why Anselm of Canterbury believed existence was a virtue, thereby concluding God existed. These seemed well within the explanatory power of the model (the first deriving from competing error-driven learning, and the second from coherent covariation). In this commentary, I will limit my comments to another domain that did not immediately appear to be within the explanatory power of the model. This is the formation and persistence of delusions.

Delusions are beliefs that arise in the absence of evidence (they are self-evident to the believer and generally very personally poignant), and are resistant (though not immune) to invalidating evidence. Upon occasion, delusions can be quite fanciful and bizarre. A number of dementias lead to delusional beliefs, including some kinds of brain trauma, late-stage Alzheimer's disease and AIDS, Parkinson's disease, affective disorders, and my particular domain of expertise, schizophrenia. Although experimental evidence is sparse as to the nature of cognitive changes during the development of a delusional psychosis, clinicians' reports are generally consistent about the typical progression. A delusion begins as a vague sense of something, for example, that someone has malicious intentions toward me. Over the span of weeks or months, a delusion begins to crystallize; for example, the government is spying on me, then the FBI is following me, then the FBI is tracking my thoughts using an implanted device. Perhaps the FBI's reasons for this elaborate artifice also become clearer to me – “I'm a threat to the government” slowly morphs into “I've been chosen by God to bring down this heretical regime.” Although delusions take on many forms, they are not random. They are characterized by personal significance and are emotionally charged. No one develops the delusion that popcorn comes from barley or that the pavement just happens to be made of worn-out carpet. Delusions experienced by depressed people are nearly always depressing; delusions experienced by manic people are nearly always expansive. Delusions appear to emerge not out of random associations, or even a random breakdown of associations. There is a landscape that appears to constrain the kinds of delusions that people will report (and presumably experience).

The landscape of one type of delusion is particularly constrained. Capgras delusions occur across several forms of dementia, including schizophrenia. In a typical Capgras case, the patient believes a loved one has been replaced by an identical imposter. One prominent hypothesis used to account for this phenomenon observes that visual form is processed by paths ascending to occipital cortex, whereas its emotional salience can be processed by subcortical paths that project directly to the amygdala (Morris et al. Reference Morris, Ohman and Dolan1999). A Capgras delusion, therefore, may occur when the occipital pathway is intact (thereby allowing facial recognition), but the subcortical pathway is lesioned (denying the perceiver the usual emotional salience of the face) (Ellis & Young Reference Ellis and Young1990; Frith Reference Frith2004). There is some support for this conjecture: Five psychiatric patients with Capgras delusions, in contrast to controls, showed no modulation of their autonomic responses to familiar and unfamiliar faces (Ellis et al. Reference Ellis, Young, Quayle and De Pauw1997). Indeed, across the different classes of delusions, there is reason to believe that an abnormal feeling drives an a posteriori cognitive explanation that is manifest as a delusion (Frith Reference Frith2004).

It is not clear how the landscape of delusional susceptibility is manifest by the model described by R&M. In accounting for the possibility of “preparedness” to learn, they readily acknowledge the possibility that the strength of some pathways are not random and may have evolved stronger links (Semantic Cognition, p. 368). This is one way to incorporate emotional experiences. However, this brief consideration of the development of delusions, and evidence for the relationship between delusions and changes in emotional salience, is not clearly reconcilable within the current model. Given the boost to psychopathology research derived from “breaking” other connectionist models (e.g., Cohen & Servan-Schreiber Reference Cohen and Servan-Schreiber1992), many readers will welcome any and all efforts R&M might make to account for these distressing phenomena. How can this model be “broken” to make it delusional?