Metabolic cost of prediction error
The biological engine of Theriault, Young & Barrett (2021), and the premise on which its whole account of social pressure rests. The wiki’s other predictive-coding sources (predictive-coding, active-inference, allostasis) treat energy budgeting as background; this construct puts the energy in the foreground and draws a motivational conclusion from it.
The two-step argument
- Neuronal signalling dominates the brain’s energy budget, and prediction error is where the cost lives. The brain is ~20% of the body’s resting metabolic budget; signalling accounts for ~75% of grey-matter energy use (most of it the Na⁺/K⁺ pump restoring ion gradients after activity), so adults spend on the order of ~13% of the whole-body budget on neuronal signalling. Predictive processing keeps this affordable by transmitting only unpredicted signals — a perfectly predicted input is redundant and carries no information (Shannon), so there is nothing to encode. The corollary: what a predictive brain pays for is prediction error.
- Therefore an unpredictable environment is metabolically costly, and — because the brain regulates prospectively (allostasis) — organisms are motivated to make their environments predictable in advance. Predictability becomes a resource, and unpredictability a cost the organism will act to avoid.
Step 2 is the move the rest of the wiki’s predictive sources do not make. Seth, Barrett, and Farb use predictive coding to explain perception, emotion, and regulation; Theriault et al. use the cost of prediction error to explain motivation — specifically the motivation to conform to other people, who are the largest unpredictable part of a human’s environment. See sense-of-should.
Constructing and coasting
The cost is not simply to be minimized. A brain that only ever minimized prediction error would sit in a “dark room” (the standard objection to free-energy accounts; Friston et al. 2012). Theriault et al. frame the trade-off as constructing vs coasting:
- Constructing — paying a short-term metabolic cost to encode prediction error, building a more general internal model that predicts better later (exploration, learning).
- Coasting — exploiting the model you already have to make cheap, accurate predictions (exploitation).
The sense-of-should is a strategy for coasting — conforming keeps the social world predictable so your existing model keeps working. But the same behaviour–expectation contingency that lets you coast is what makes interactive-inference (a constructing move — violating expectations as controlled experiments) possible. Neither pole can dominate, because survival needs both.
What to hold loosely
The paper is candid that the load-bearing quantity is unspecified. The relation between prediction error and metabolic cost is written as a proportionality (∝) throughout, “as the exact relation… is unknown,” and footnote 2 concedes that how metabolic cost is realized at the micro-scale is “an open area of research” — candidates include astrocytic glycogen depletion and amyloid accumulation, and the account is careful to distance itself from the discredited circulating-blood-glucose model of willpower (Gailliot et al.; failed to replicate). So the premise “prediction error is metabolically costly” is well grounded, but the stronger premise the social argument needs — that it is costly enough to motivate behaviour — is asserted rather than measured. This is the seam a critic of the whole framework would press.
Where it connects
The same energetic framing surfaces on allostasis (the body budget) and reaches the clinic on computational-psychiatry (mis-set precision and aberrant prediction error as a mechanism of illness; Stephan’s allostatic self-efficacy). Theriault et al.’s contribution is to make the cost of error itself the currency of social motivation, one level up from where the wiki’s other sources leave it.