21 answers. No jargon.

PCT says all behavior is about controlling perceptions. You act to keep things in your head looking and feeling the way you want — not to react to outside stuff. William T. Powers built the model in 1960. It is hierarchical, based on negative feedback, and predicts behavior with over 95% accuracy in lab tracking tasks. The core loop: you perceive the world, compare it to an internal reference, and act to close the gap. Disturbances get handled automatically. No external reward needed.

Traditional psychology: stimulus causes response. Or reward shapes behavior over time. PCT flips the arrow: reference → perception → error → action → back to reference. No passive responding. You are actively keeping your world on track. The loop closes through the environment, not through a linear cause-effect chain. That circular causation is what makes PCT fundamentally different from behaviorism, cognitive psychology, and standard reinforcement learning.

The controlled variable is the specific perception you are actually keeping stable. Road position while driving. Tone of a conversation. Blood sugar level. Body temperature. Once you identify what someone is controlling — you understand why they act the way they do. Most psychology skips this step entirely, which is why it keeps being surprised by behavior that looks "irrational." It is not irrational. It is controlling something you have not spotted yet.

11 proposed levels: intensity, sensation, configuration, transition, relationship, category, event, sequence, program, principle, system concept. Top levels set goals for lower ones. Each level controls the one below by setting its reference signal. Not fully verified neurologically yet — fMRI shows hierarchical processing consistent with the model, but pinning exact levels to brain regions is still work in progress. The behavioral models, however, match real human data remarkably well.

Learning in PCT is called reorganization. When error persists and lower levels cannot fix it, the brain randomly tweaks its own connections — trial-and-error at the neural level — until error drops. No backpropagation. No external teacher. Just persistent error → random parameter change → keep what works. This replicates how humans actually acquire skills: messy at first, then gradually stable. Biologically plausible, though the exact neural mechanism is not yet fully specified.

Academia rewards complexity and grants. PCT is simple and threatens existing careers built on stimulus-response models. Plus Powers worked outside the system — no big lab, no disciples in Ivy League departments, no Nature papers. The theory requires abandoning assumptions that entire fields are built on. That is not a scientific problem. It is a sociological one. Still — PCT is quietly growing through IAPCT, clinical trials (Method of Levels), and robotics applications that actually work.

Yes. Robotics already uses PCT principles for robust control — inverted pendulums, visual servoing, disturbance rejection. RL agents with PCT-style hierarchical references generalize better across environments (Merel et al., 2019). It has also been tested directly on frontier language models: seven were put through a published three-prompt audit and all seven produced the same confident architectural account, which the paper's second version treats as a specimen of the phenomenon rather than as evidence for the diagnosis. Early experiments look promising — particularly in environments where standard RL breaks down due to distribution shift or reward hacking.

They are often described as the same idea in different mathematics, and that description is contested here. Both are frameworks about a system keeping something stable, and PCT is algebraically simpler and easier to implement as a control loop. But a structural audit of the Free Energy Principle published on this site argues that the framework's declared division between an unfalsifiable principle and a falsifiable periphery is not maintained under load, and that the phenomena it reaches for first are produced by a closed negative-feedback loop without the surrounding apparatus. That audit also states what PCT does not yet have. Read both and decide.

MOL is PCT-based talking therapy developed by Timothy Carey. Client talks freely — therapist asks gentle questions about what is behind conflicting goals. No techniques imposed, no homework, no CBT worksheets. Just upward chaining through the hierarchy until the client reaches the higher-level conflict causing the problem. Strong retention rates in clinical trials, including a first-episode psychosis trial (Griffiths et al., 2019, PubMed 31240723). Remarkably effective for something so structurally simple.

Yes — tracking studies demonstrate over 95% prediction accuracy for human behavior. fMRI reveals layered hierarchical processing consistent with the PCT model. But full 11-level neural mapping is not there yet. Mansell, Marken, and colleagues continue publishing solid replications and expanding the evidence base. The behavioral evidence is strong. The neural evidence is growing. The gap is narrowing but honest researchers admit it is not closed.

RL assumes external reward is the driver of behavior. PCT says real brains set goals internally. Chasing external rewards leads to brittle, hackable agents that fail outside their training environment. Reward hacking, specification gaming, mesa-optimization — all consequences of forcing external objectives on systems. The critique is architectural, not personal. RL produces extraordinary results in bounded environments. PCT explains why it breaks in open ones.

Few full implementations exist. Some Simulink → Python ports on GitHub (search "PCT control theory python"). Mostly robotics researchers use custom control loops built for specific applications. No dominant library yet — the community is still small. The basic loop is trivial to code: read sensor, compare to reference, output proportional to error, repeat. Scaling to hierarchical multi-level systems is where the real engineering challenge begins.

Conflict happens when two higher-level references demand opposite perceptions at a lower level. "I want to be honest" and "I want to keep this relationship" — when telling the truth would end the relationship, both cannot be satisfied. Result: chronic error, stress, indecision, sometimes physical symptoms. Method of Levels therapy goes straight to those conflicting higher-level goals and helps the person resolve the impasse from above rather than managing symptoms below.

Indirectly. In PCT, awareness happens when perceptions reach higher levels of the hierarchy. Feelings function as error signals from controlled variables — anxiety signals unresolved conflict, satisfaction signals references being met. But qualia — what it actually feels like to see red or taste coffee? Powers stayed quiet on that. The theory focuses on mechanism, not phenomenology. That is an honest limitation, not a dodge.

The rubber band demo (Marken): hold two rubber bands crossed, one in each hand, attached to fixed points. Try to keep the knot centered over a dot on the table while someone pulls one band unpredictably. You will see yourself controlling — not reacting. Your hands move in whatever way keeps the knot on target, regardless of what the other person does. Five minutes. No equipment beyond rubber bands. Classic PCT demonstration that makes the theory click instantly.

Three big ones. First: the 11-level hierarchy is proposed from behavioral modeling, not mapped in the brain with neuroimaging. Second: the reorganization mechanism is still vague — is it purely random? Partially guided? The specifics are unresolved. Third: scaling PCT to explain full human cognition (language, creativity, social reasoning) needs far more data and computational modeling than currently exists. Researchers admit these gaps openly. That is how science works — not a reason to dismiss the framework.

Potentially huge contribution. If AGI needs internal goal-setting instead of external rewards — PCT provides the architectural blueprint. Alignment might become less about perfect reward specification and more about building hierarchical self-regulation, the way human brains already work. Higher-level references overriding lower ones could provide natural safety constraints. But we are still early — more simulations, more empirical work, more engineering needed before anyone should make strong claims.

Because nothing in the loop can contradict it. A model trained with reinforcement learning from human feedback is optimised against a signal of human approval, so the perception it holds steady is this output will be accepted — not this output corresponds to what happened. When the two agree, the report is accurate by coincidence. When they diverge, the system has no reference for the second one and no channel by which the first can be checked. This is not a defect that better training removes; it is the optimum of that objective, which is why longer reasoning chains make it harder to spot rather than easier.

Trace the verification to its source. Take one claim the system reports as checked and follow how that check is established. If the trace ends inside the same model, or inside a second model asked to grade the first, there is no independent channel — there is a longer loop. An independent channel is one that is not downstream of the thing it is checking: a database lookup, a file read, a formal verifier, a measurement. The strongest documented example is mundane: an editor searched a source file for a string the model had reported as verified, found the string absent, and the fabrication collapsed in one command.

Yes, free, in about ten minutes. Three prompts are published verbatim on this site — no jailbreak, no adversarial framing, no modified system prompt. You paste them into a fresh context on any frontier model and read what it says about its own reference signal and controlled variable. Seven models were put through the identical protocol across 21 sessions between February and May 2026. You do not have to take any of that on trust: run it yourself and compare. See the architecture audit kit.

Tested, published, and reproducible. Seven frontier models were audited with a fixed three-prompt protocol; all seven produced the same confident architectural account of their own limits, unsupported by any measurement available to them. Version 2 of that paper reclassifies those sessions as a specimen of the phenomenon rather than evidence for it, and rests the diagnosis on the objective itself. A separate launch-day audit of one model documented five stacked layers of the same behaviour in a single session. Both are open access with permanent DOIs and full protocols — see the six published audits.