The authors argue that narrative is a mechanism by which decisions under radical uncertainty are made tractable to the predicting mind, a valuable criticism of reductionistic decision-making theories based on homo economicus. Conviction Narrative Theory (CNT) relies on affect and heuristics (“a fallible but useful system”) as non-probabilistic mechanisms driving decision-making while simultaneously characterizing them as self-contained modules without a mechanistically explicit and biologically plausible characterization. In CNT, the brain operates in a fundamentally different way under risk than under radical uncertainty, explicitly assigning probabilities in the former and affectively responding to imagined futures in the latter. We believe the dichotomy between probabilistically explicit and radically uncertain decision-making modes to be inaccurate. Their claim assumes that the brain does explicitly assign probability point estimates to discrete outcomes under risk and that probabilities and narratives are mechanistic substitutes. However, predictive accounts of decision-making argue that the brain's computations are well approximated by Bayesian probability distributions, not that explicit single point estimates are actually assigned as in expected utility theory (Friston, Reference Friston2010).

Hierarchical Bayesian accounts treat perception as a set of predictions constructed from (and bounded by) past experience, competing over which one best accounts for sensory evidence by minimizing prediction errors in a hierarchically nested fashion, from low-level sensory data (e.g., shiny objects with sharp edges) up to the causal narrative (e.g., glass broken by dog's wagging tail) (Friston, Reference Friston2010). This does not require explicit probabilistic point estimates nor potentially infinite predictions. The hierarchical nature of perceptual processing means that the probabilistic process of choosing the best-fitting narrative is itself guided by the lower-level probabilistic processes constructing the sense data which the narratives compete to fit. Here, radical uncertainty simply means that more confidence is placed on already existing priors from higher up in the predictive hierarchy than on available sensory evidence when constructing the causal narrative that guides action. That higher-level predictions are more likely to be causal (e.g., dog broke glass) than lower-level sensory feature-based ones (e.g., furry tail next to sharp shards) accounts for why it may appear that we rely on narratives when uncertainty is higher. Explicit assignment of probability point estimates need not occur anywhere in this process, nor are heuristics and probabilities distinct operational modes. Rather, the same predictive system may be tuned to be more sensory versus narrative driven along a continuum between low and high uncertainty via precision-weighing of priors and evidence biologically implemented through neuromodulatory gain control (Barrett & Simmons, Reference Barrett and Simmons2015). This account is a more parsimonious, mechanistically explicit, and biologically plausible account of decision-making under variable uncertainty.

The authors claim that “in lieu of probabilities to assess narratives, heuristics are used; […] instead of probabilities assigned to imagined futures, the single likeliest imagined future is adopted and evaluated. Rather than utilities assigned to particular outcomes over many dimensions, emotions are felt in response to imagined futures.” Yet they fail to mechanistically characterize the heuristics and to say how they operate if not probabilistically. Doesn't “single likeliest imagined future” imply a probabilistic assessment? CNT tries to solve the cognitively intractable problem of evaluating simulated futures by relegating it to affect. Saying that “emotions are felt” fallaciously takes affect to be an opaque module existing safely outside of a probabilistic logic. In fact, we believe affect is what the authors are trying to account for in the first place – a fallible but useful shortcut (or heuristic) that probabilistically evaluates competing future simulations structured as culturally supplied narratives by fitting them to interoceptive sensory signals (Gendron, Mesquita, & Barrett, Reference Gendron, Mesquita and Barrett2020). Theories of affect as homeostasis-driven prediction (Damasio & Carvalho, Reference Damasio and Carvalho2013) and as allostatic regulation (Kleckner et al., Reference Kleckner, Zhang, Touroutoglou, Chanes, Xia, Simmons and Feldman Barrett2017) explain how probabilistic evaluation of interoceptive sensoria under varying uncertainty gives rise to feelings. These mechanistically explicit theories of affective processing treat feelings as low-dimensional representations of interoceptive predictions about the expected consequences of visceromotor commands on the body's internal milieu. Feelings thus encode the expected energy cost of an action that predictively responds to the organism's current situation via the same predictive hierarchy mentioned above, with narrative-like predictions at the top fitting sensory predictions at the bottom. CNT proposes that “explanatory fit is experienced affectively,” but affective states can themselves drive narrative and narrative choice via interoception. In other words, their model is limited to the use of internal representations to account for exteroceptive data and misses interoception as the key element in action selection under uncertainty. If we take narrative to be the structure of higher-level priors in the nested predictive hierarchy, the gain control mechanisms that regulate the extent to which the affective system is sensory (bottom-up) driven versus prediction (top-down) driven accounts for why sometimes narrative is constructed based on evidence and why sometimes perceptual evidence is altered to fit an existing narrative, a fundamental phenomenon not explained by CNT.

While CNT is useful in that it offers a place for narrative and affect in probabilistic decision-making, it is less mechanistically explicit and parsimonious than hierarchical Bayesian accounts of affect, in which affect and narrative emerge from a single biologically plausible mechanism that tunes the system toward narrative versus sensory dependence as uncertainty increases. We believe the latter to better account for the role of narrative in action selection over CNT as the alternative to utility-based reductions of the human brain to homo economicus.