If what you experience is built by the brain rather than received from the world, then perception, meaning and orientation are all models — and models can be examined, tested and refined. This is the question beneath the entire project.
Across neuroscience, cognitive science and philosophy of mind, a dominant framework has formed: the brain is fundamentally a prediction machine. Rather than building experience bottom-up from raw sensory input, it runs a generative model — its best guess about the causes of its sensory stream — and uses incoming signals mainly to correct prediction errors. Perception, on this view, is a "controlled hallucination" reined in by the world (Seth; Clark; Hohwy). Karl Friston's Free Energy Principle formalises it: organisms persist by minimising prediction error (free energy), either by updating the model (perception) or by acting to make the world fit the model (active inference).
Two independent research programmes — one from neuroscience, one from evolutionary game theory — arrive at the same unsettling place: what you perceive is not a faithful readout of the world. They agree on that, and disagree sharply on what it means. That disagreement is where the frontier is.
Active inference, developed by Karl Friston, treats the brain as a prediction machine running a generative model of the world. It does not assemble experience from raw sensory data flowing inward; instead it continuously generates its best inference about the hidden causes of its sensations, and uses the senses mainly to correct that model where it errs. Formally, the system acts to minimise prediction error — "variational free energy," a measure of the gap between what it expects and what it senses. Crucially, it can close that gap two ways: by updating the model (this is perception) or by acting on the world to make the sensations match the prediction (this is action). Perceiving and acting become two faces of the same loop. So what you experience is a controlled construction, tuned to support useful inference and action — not to mirror reality faithfully.
Donald Hoffman pushes the point further, and from a different direction. With Chetan Prakash he proved what they call the Fitness-Beats-Truth (FBT) theorem: in evolutionary game-theoretic models, organisms whose perceptions are tuned to fitness reliably outcompete and drive to extinction those tuned to perceive objective reality as it is. In their simulations the probability that a truth-perceiving strategy survives natural selection rounds, strikingly, to roughly zero. Hoffman's conclusion — his Interface Theory of Perception — is that our senses evolved like a desktop interface: the icons are useful precisely because they hide the underlying machinery rather than reveal it. On this view space, time and physical objects are species-specific data structures, not the furniture of reality.
Read this precisely: the "≈0%" is a result within a formal evolutionary model under stated assumptions, not an unconditional empirical fact. Its strength is the theorem; its limit is the modelling assumptions — which is exactly what a careful reader should interrogate.
Both theories agree perception is non-veridical — built for usefulness, not truth. But they part on the metaphysics. Friston's active inference is broadly naturalist and physicalist: the brain is a physical organ modelling a physical world it cannot see perfectly. Hoffman runs the same non-veridicality against physicalism, arguing for "conscious realism" — that consciousness, not matter, is fundamental. Same premise, opposite conclusions. Resisting the urge to collapse that into one tidy story — and instead asking which assumptions drive the divergence — is the actual frontier work here.
The interesting move is not to summarise the science but to ask what it implies:
These are interpretive implications drawn from the framework, not established experimental findings — flagged as such.
Look at the same picture through my own framework, Recursive Field Theory (RFT). RFT sees the mind as a system always trying to "close the loop" — to settle into a stable, coherent read of the world. Your existing beliefs are the memory it starts from; surprise is what unsettles it; and you resolve that surprise the same two ways active inference describes — by updating your picture, or by acting to change the world. When the loop settles, you feel oriented and clear; when it can't, you feel fragmented. That's why better models of reality genuinely bring more clarity and agency — it's the same settling process, seen from the inside.
Recursive Field Theory is my own framework, offered as a way of seeing — not established neuroscience. It describes the structure of the process, not why it feels like anything.
Primary sources linked where available; books cited by title. Always consult the originals — this synthesis describes emphasis and findings, not verbatim claims.
Learning outcomes. After this investigation you should be able to: (1) explain the predictive-processing / active-inference account of perception; (2) state the Fitness-Beats-Truth result and its modelling caveats; (3) articulate why active inference and interface theory agree perception is non-veridical yet diverge metaphysically.
To complete the unit (this is what makes it ~45 minutes of reflective learning):
Self-certified, CPD-eligible structured learning — not an endorsement by any statutory regulator. Hours shown reflect estimated reflective-learning time; log only genuine time spent. Certificate available to members.
Our standards. Every investigation is built to be tested, not believed: claims are sourced, strong evidence is kept separate from contested evidence, competing theories and contradictions are shown rather than smoothed over, confidence is stated explicitly, and each piece names what would change its mind — and a human reviews every word before it is published. Reality is the arbiter.
Join the Frontier — free every week →
© 2026 David Fleming · Models of Reality / The Frontier. All rights reserved. Recursive Field Theory (RFT) is the original work of David Fleming. No part of this publication or the system that produces it may be reproduced, reverse-engineered or used to train or build a competing service without permission.