Skip to content

Efference Copy

Prime #
819
Origin domain
Neuroscience
Subdomain
sensorimotor integration → Neuroscience
Aliases
Corollary Discharge

Core Idea

Efference copy names the recurring structural pattern in which a system that issues a control command also routes an internal copy of that command to its perception, monitoring, or audit subsystem, which uses the copy to predict the self-caused consequences and subtract them from the incoming signal, leaving only the world-caused remainder for further processing. The structural commitment is the separation of self-caused effects from world-caused effects via a forward signal broadcast at the moment of action — not via retrospective inference.

The pattern requires five jointly necessary elements: a controller that issues commands to effectors; an effector that executes the command and changes the world or the controller's own state in ways the perceiver will register; a perceiver or monitor that senses both world-caused and self-caused changes through the same channel, with no intrinsic way to tell them apart from the raw signal; an internal copy of the command, broadcast concurrently from controller to perceiver and bypassing the effector path; and a predictor and subtractor that uses the copy to anticipate the self-caused component of the upcoming sensation and removes, attenuates, or flags it before downstream processing. The diagnostic signature is self-attenuation: predicted self-caused signals are systematically smaller than identical world-caused signals, because the predictor cancels them, and when the prediction is wrong — effector drift, unexpected load, interference — the residual is unattenuated and downstream systems see the discrepancy. This is distinct from generic feedback, which corrects after the world responds, and from broad predictive comparison, because efference copy is specifically the self-versus-world attribution mechanism built from broadcasting a command copy to the perceiver.

How would you explain it like I'm…

Can't Tickle Yourself

You can't tickle yourself, because your brain knows your own hand is coming and gets ready for it, so it doesn't feel surprising. When you decide to move, your brain sends a little secret note to itself saying 'I'm about to do this,' so it can tell apart what YOU did from what the world did. That note is why your own touch feels different from someone else's.

The Secret Heads-Up Note

When your brain tells your body to do something, it also sends an INTERNAL COPY of that command to the part that senses things, as a heads-up. That sensing part can't normally tell apart what it caused itself from what the world caused, but the copy lets it PREDICT the self-caused part and subtract it out, leaving just the world's part to notice. That is why you can't tickle yourself: your brain predicted your own touch and turned it down. If the prediction is wrong, like someone bumps your arm unexpectedly, the surprise isn't subtracted and you really feel it.

Self-Versus-World Subtraction

Efference copy is a pattern where a system that issues a control command also routes an INTERNAL COPY of that command to its own perception or monitoring part, which uses the copy to predict the self-caused consequences and SUBTRACT them from the incoming signal, leaving only the world-caused remainder for further processing. The key structural commitment is separating self-caused effects from world-caused effects using a forward signal broadcast AT THE MOMENT OF ACTION, not by figuring it out afterward. The diagnostic signature is self-attenuation: a predicted self-caused signal comes out systematically smaller than an identical world-caused one, because the predictor cancels it, which is why you cannot tickle yourself. When the prediction is wrong, from unexpected load or interference, the leftover is un-attenuated and downstream systems see the discrepancy. This is different from ordinary feedback, which corrects AFTER the world responds, and narrower than general prediction, because efference copy is specifically the self-versus-world attribution mechanism.

 

Efference copy names the recurring pattern in which a system that issues a control command also routes an INTERNAL COPY of that command to its perception, monitoring, or audit subsystem, which uses the copy to predict the self-caused consequences and subtract them from the incoming signal, leaving only the world-caused remainder for further processing. The structural commitment is the SEPARATION OF SELF-CAUSED FROM WORLD-CAUSED EFFECTS via a forward signal broadcast at the moment of action, not via retrospective inference. The pattern requires five jointly necessary elements: a CONTROLLER that issues commands to effectors; an EFFECTOR that executes the command and changes the world or the controller's own state in ways the perceiver will register; a PERCEIVER or monitor that senses both world-caused and self-caused changes through the same channel, with no intrinsic way to tell them apart from the raw signal; an INTERNAL COPY of the command, broadcast concurrently from controller to perceiver and bypassing the effector path; and a PREDICTOR-AND-SUBTRACTOR that uses the copy to anticipate the self-caused component of the upcoming sensation and removes, attenuates, or flags it before downstream processing. The diagnostic signature is SELF-ATTENUATION: predicted self-caused signals are systematically smaller than identical world-caused signals, because the predictor cancels them, and when the prediction is wrong, effector drift, unexpected load, interference, the residual is unattenuated and downstream systems see the discrepancy. This is distinct from generic feedback, which corrects after the world responds, and from broad predictive comparison, because efference copy is specifically the self-versus-world attribution mechanism built from broadcasting a command copy to the perceiver.

Structural Signature

the controller issuing commands — the effector executing them and changing the sensed world — the perceiver sensing self- and world-caused changes through one channel — the internal command copy broadcast concurrently, bypassing the effector path — the predictor-and-subtractor that cancels the anticipated self-effect — the self-attenuation signature with unattenuated residual on prediction error

A configuration exhibits efference copy when each of the following holds:

  • A controller. Some component issues control commands to effectors.
  • An effector. It executes the command and changes the world, or the controller's own state, in ways the perceiver will register.
  • A perceiver with a single channel. A monitor senses both world-caused and self-caused changes through the same channel, with no intrinsic way to tell them apart from the raw signal.
  • An internal command copy. A copy of the command is broadcast concurrently from controller to perceiver, at the moment of action and bypassing the effector path — a forward signal, not a retrospective inference.
  • A predictor and subtractor. The perceiver uses the copy to predict the self-caused component of the upcoming sensation and removes, attenuates, or flags it before downstream processing.
  • A self-attenuation signature. Predicted self-caused signals are systematically smaller than identical world-caused signals; when the prediction is wrong (effector drift, unexpected load), the residual is unattenuated and downstream systems see the discrepancy.

Composed, these separate self-caused from world-caused effects via a forward broadcast rather than after-the-fact correction — distinguishing the pattern from generic feedback (which corrects after the world responds), from plain self-monitoring (no concurrent copy), and from broad signal extraction (the separated component need not be self-caused). Attribution bugs live at the forward channel, not at the sensor or controller alone.

What It Is Not

  • Not feedback. Feedback corrects after the output deviates and the world responds; efference copy broadcasts a command copy concurrently with action, predicting the self-effect before it arrives — forward, not retrospective.
  • Not signal_extraction. Signal extraction separates any component of interest from noise; efference copy specifically separates the self-caused component using a command copy — the distinguished component must be self-generated.
  • Not observer_effect. The observer effect is measurement disturbing the observed; efference copy is the cancellation of self-caused disturbance via a forward prediction, the opposite move.
  • Not plain self-monitoring. Self-monitoring observes one's own state but need carry no concurrent command copy; efference copy's defining element is the broadcast-at-the-moment-of-action forward channel.
  • Not mere_exposure_effect. That is a familiarity-driven preference shift; efference copy is a sensorimotor attribution mechanism — the embedding proximity is incidental.
  • Not predictive coding in general. Broad predictive comparison anticipates any input; efference copy is the specific self-versus-world attribution built from broadcasting a command copy to the perceiver.
  • Not latency management. Latency concerns delay per se; efference copy addresses the attribution problem that delay creates (delayed-self versus new-world), not the delay itself.
  • Common misclassification. Responding to "the system keeps reacting to its own actions" by upgrading the sensor. That treats an attribution failure as a signal-to-noise one; the fix is a forward channel, and sensor quality is irrelevant.

Broad Use

The canonical instance is neuroscience and motor control: corollary discharge in oculomotor circuits, saccadic suppression during eye movements, the cerebellar forward model for limb dynamics, and the attenuation of self-produced touch all instantiate the broadcast-and-subtract architecture. It recurs in speech production, where speakers attenuate perception of their own voice via command-driven predictions in auditory cortex and compensate when feedback is artificially altered. It recurs in control engineering as the Smith predictor and internal model control: to control a plant with significant delay, the controller maintains an internal model and operates on the prediction error rather than the raw output. It recurs in robotics, where a forward model predicts the expected reaction force of a commanded torque and subtracts it, leaving only the unexpected component for collision detection. It recurs in security operations, where a scheduled maintenance command emits a concurrent notification so the monitoring system anticipates and cancels the self-caused changes. It recurs in software change management, where a deploy emits an expected-effect manifest in parallel with the change, so the observability system attenuates the expected signature and alerts only on deviations. And it recurs in organizational accountability, where an announced policy change is in effect an efference copy to the audit function, so downstream metric movements are read as expected rather than anomalous.

Clarity

Efference copy separates two phenomena that the raw signal cannot: world-caused changes that require action and self-caused changes that do not. Without the pattern, every signal looks like something happening to the agent; with it, the agent perceives itself as a distinct cause and reserves its alarm response for genuinely external events. It also separates the prediction from the correction: the forward copy predicts the expected sensation, the actual sensation arrives, and the comparison yields a residual — a three-part structure crisper than "self-monitoring" or "internal model" alone, both of which collapse the prediction-versus-residual distinction. And it separates two scalars that are routinely confused: signal-to-noise, a ratio property of the sensor improved by better sensors, and self-versus-world attribution, a predictive property of the perceiver improved by better efference copy. A sensor swap will not fix an attribution failure, and naming the pattern is what makes that non-obvious fact visible.

Manages Complexity

The pattern compresses a substrate-spanning design problem into a single named architecture: broadcast the command sideways at the moment of issue, and subtract the predicted self-effect from the incoming signal before downstream processing. Engineers meeting the distinguish-self-from-world problem in a new domain need not re-derive the solution; they can run the five-part diagnostic — where is the controller, the effector, the monitor, the forward channel, the predictor — and locate the intervention point. By factoring attribution into these five components, the pattern lets an analyst reason about where an attribution failure must lie without modelling the whole sensorimotor or monitoring loop, and it cleanly separates the question of sensor quality from the question of predictive cancellation, which are otherwise entangled under a vague heading of "the system keeps reacting to its own actions."

Abstract Reasoning

Recognising the pattern supports inference about systems that lack it. The failure mode of absent efference copy is that a system cannot distinguish its own contributions from external ones, so it either over-alerts on every self-action or learns to silence whole sensor channels, losing world-attributable signal as well — both failure modes carrying distinct signatures. The Smith-predictor argument holds that a system acting under perceptual delay performs better with a forward model than with raw feedback alone, because feedback cannot disentangle delayed-self from new-world. The saccadic-suppression argument holds that any system moving its own sensor must solve the self-motion-versus-world-motion problem, and the cleanest solution is a command copy to the processor. The pattern even illuminates pathology: a failure of efference copy in inner-speech monitoring has been proposed as a mechanism for misattributing self-generated signal to an external source, and the structural inference — self-signal not cancelled, therefore experienced as world-signal — recurs in any system whose attribution layer fails. The unifying structural insight is that attribution bugs live at the forward channel, not at the sensor or the controller individually.

Knowledge Transfer

Because the self-versus-world attribution problem is substrate-independent and the forward-copy-plus-predictor-plus-subtractor solution is its substrate-independent answer, the architecture transfers cleanly across domains. The Smith predictor is the control-engineering instantiation of the cerebellar forward model: both compute what the sensor should be reading if the command is executing as intended, and both act on the residual. The saccadic-suppression shape transfers to change-window suppression in monitoring systems, where a maintenance window's announcement is the efference copy and the reduced sensitivity during the window is the attenuation, carrying the intervention vocabulary — precise timing, expected-effect model, unattenuated residual — across the gap. Forward-model torque cancellation in robotics is the same structure as vestibulo-ocular compensation. The practice of emitting an expected-effect manifest before a deploy, so observability tools attribute consequent metric movement correctly, is structurally the efference-copy move in software. In every port the diagnostic is the five-part decomposition and the prescription is the same: install a concurrent forward channel from the actor to the monitor, predict the self-effect, and subtract it before alarms fire. The transfer carries its boundary as well: a receiving domain must distinguish efference copy from reactive feedback (which corrects after the response arrives), from generic self-monitoring (which need not use a concurrent command copy), and from broad signal extraction (where the separated component is not specifically self-caused). A practitioner who has built this architecture in one substrate arrives at the next already asking where the controller, the monitor, and above all the forward channel sit — and already knowing that when attribution fails, the forward channel is where to look.

Examples

Formal/abstract

Saccadic suppression and the oculomotor corollary discharge is the prime's canonical neuroscience case, and it isolates every role. The controller is the oculomotor circuitry issuing a command to move the eyes; the effector is the extraocular muscles, which sweep the eye and so sweep the retinal image. The perceiver is the visual system, which receives a single channel — the moving image on the retina — in which a self-caused sweep (the eye moved) and a world-caused sweep (the scene moved) are intrinsically indistinguishable from the raw signal. The internal command copy is the corollary discharge, broadcast concurrently from the oculomotor command to visual areas, bypassing the muscle-and-retina path. The predictor and subtractor uses that copy to anticipate the retinal smear the saccade will cause and suppress it, which is why you do not perceive the world lurching every time your eyes dart. The diagnostic self-attenuation signature is exact and testable — the same retinal motion is perceived when externally imposed (push your eyeball gently and the world does appear to jump) but cancelled when self-generated, because only the self-generated case carries a matching command copy. And when prediction fails, the unattenuated residual is seen: this is why a system whose eye muscles are paralysed and tries to move them perceives the world as shifting — the command copy predicts a sweep that the failed effector never produced.

Mapped back: The oculomotor command is the controller, the eye muscles are the effector, vision is the single-channel perceiver, corollary discharge is the concurrent command copy, saccadic suppression is the predict-and-subtract, and the pushed-eyeball jump is the unattenuated residual on prediction error.

Applied/industry

The Smith predictor in control engineering instantiates the same prime in a process-control substrate, solving the self-versus-world problem under transport delay. The controller is the feedback regulator; the effector is the plant — say a chemical reactor whose temperature responds to a heating command only after a long pipe-and-mixing delay. The perceiver is the temperature sensor, whose single reading cannot tell, at any instant, how much of the current temperature is the delayed result of the controller's own earlier command versus a new world disturbance (a feed change). The internal command copy is the controller's internal plant model, run forward concurrently with the real command; the predictor and subtractor computes what the sensor should read if the command is executing as intended and operates on the residual between predicted and actual output. Self-caused, delayed effects are thereby cancelled, and only the unexpected component — the genuine disturbance — reaches the control law, exactly the prime's broadcast-and-subtract architecture. The clarifying payoff matches the prime's: a better sensor will not fix instability caused by delay (that is an attribution problem, not a signal-to-noise one); a forward model will. A structurally identical applied instance is software deploy management, where a release emits an expected-effect manifest in parallel with the change, so the observability system attenuates the predicted metric movement and alerts only on deviations — the deploy notification playing the role of the command copy.

Mapped back: The regulator is the controller, the delayed plant is the effector, the sensor is the single-channel perceiver, the internal plant model is the concurrent command copy, operating on the prediction residual is the predict-and-subtract, and an unexplained residual is the world-caused disturbance the architecture isolates.

Structural Tensions

T1 — Prediction Accuracy versus Cancellation Aggressiveness (measurement/sign). The predictor subtracts the anticipated self-effect, but the subtraction is only as good as the forward model. An over-confident predictor cancels too much, masking real world-caused signal that resembled the predicted self-effect; an under-confident one cancels too little, leaving self-noise. The prime's "subtract the self-effect" presumes a well-calibrated model. Failure mode: a drifted forward model that confidently cancels a genuine disturbance because it looked like the expected self-effect, so a real event is silently suppressed. Diagnostic: monitor the residual distribution; systematically near-zero residuals can mean over-cancellation hiding real signal, not perfect prediction.

T2 — Efference Copy versus Reactive Feedback (boundary with a competing prime). The prime is specifically forward — a concurrent command copy — distinct from feedback that corrects after the world responds. The boundary blurs when delays are short: fast feedback can mimic prediction, tempting the designer to skip the forward channel. But under significant delay, feedback cannot disentangle delayed-self from new-world. Failure mode: relying on reactive feedback for a self-versus-world attribution problem that only a forward model can solve, so the system oscillates or mis-attributes under delay. Diagnostic: ask whether the self-effect arrives after a delay long enough that feedback cannot separate it from new disturbances; if so, feedback is the wrong prime.

T3 — Self-Attenuation versus Lost World-Signal (scopal/over-suppression). The self-attenuation signature is the prime's diagnostic virtue — predicted self-signals are smaller than identical world-signals. But attenuation applied through a shared channel risks suppressing world-caused signal that coincides with the self-action window. The mechanism that cancels self-noise can blind the system during its own actions. Failure mode: a monitoring system in a maintenance window suppressing not just the expected self-caused changes but a real incident that happened to occur during the window. Diagnostic: ask whether attenuation is scoped to the predicted signature or to the whole channel during the action; blanket suppression during self-action loses concurrent world events.

T4 — Concurrent Timing versus Effector Delay (temporal/synchronisation). The command copy is broadcast at the moment of action, but its predicted effect must be aligned in time with when the self-effect actually arrives at the perceiver — and that lag varies (effector drift, variable transport delay). A mistimed cancellation subtracts the right magnitude at the wrong moment, creating two artefacts instead of zero. Failure mode: subtracting the predicted self-effect on a fixed schedule when the true effector delay has shifted, so the cancellation misaligns and produces a spurious residual where there was none. Diagnostic: check whether the predicted-effect timing tracks the actual effector latency; a forward model with a stale delay estimate cancels into the wrong time bin.

T5 — Self-versus-World Attribution versus Signal-to-Noise (measurement/category). The prime sharply separates an attribution property (improved by better efference copy) from a sensor property (improved by better sensors) — and warns these are routinely conflated. The tension is that the symptom ("the system keeps reacting to its own actions") invites the wrong fix. Failure mode: responding to an attribution failure by upgrading the sensor, spending on signal-to-noise when the problem is that no forward channel cancels the self-effect — the sensor swap changes nothing. Diagnostic: ask whether the spurious reactions are to self-caused effects specifically; if so, the fix is a forward channel, and sensor quality is irrelevant.

T6 — Forward Channel versus Single Point of Attribution Failure (coupling/fragility). The prime locates attribution at the forward channel — "when attribution fails, the forward channel is where to look." But this concentrates the entire self-versus-world distinction in one pathway, making it a single point of failure: corrupt or sever the command copy and self-signal is experienced as world-signal. Failure mode: a broken forward channel causing systematic misattribution of self-generated signal to an external source (the inner-speech pathology the prime cites), with no redundancy to catch it. Diagnostic: ask what cross-checks the forward channel's predictions against ground truth; a forward model trusted without any independent attribution check fails silently and wholesale when the channel corrupts.

Structural–Framed Character

Efference copy sits well onto the structural side of the structural–framed spectrum: the pattern — route an internal copy of a command to the perceiver so self-caused effects can be predicted and subtracted, leaving the world-caused remainder — is a bare relational control architecture, with only a mild residual frame from its neuroscientific birthplace.

Three diagnostics read fully structural. Evaluative weight is zero: broadcasting a command-copy to cancel self-caused signal is neither good nor bad in itself — the same architecture distinguishes a useful sensorimotor cancellation from a deploy-notification that suppresses a self-triggered alert, value-neutral until specified. Human-practice-bound is zero: the five-part structure runs in purely physical and engineered substrates — a Smith predictor in a control loop, a robot subtracting its own actuation, a SIEM tagging self-generated events — needing no human practice; biological nervous systems are one substrate among several, not a prerequisite. Import-vs-recognise leans recognition (0.5): to diagnose efference copy is to notice that a forward command-copy is being used to separate self from world at the moment of action, a structure present in the system, though the "corollary discharge" framing is sharpened by the neuroscience lens. The two diagnostics at the half-mark are vocabulary and origin: "efference copy," "corollary discharge," "the perceiver," "reafference" carry a sensorimotor-neuroscience home lexicon that control engineering, robotics, and monitoring must translate, and the origin is a specific discipline rather than a pure formal relation.

The honest reading is that nothing here imports approval or human ceremony, and the architecture runs in engineered and physical substrates indifferently — which holds it firmly on the structural side — while the neuroscientific vocabulary and disciplinary origin keep it off the pole. Neutral, substrate-indifferent, recognised structure against a half-translated lexicon and domain-specific origin yields an aggregate of 0.3, matching the assigned mixed-structural grade.

Substrate Independence

Efference copy is a strongly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its domain breadth is wide (4 / 5): the concurrent command-copy that enables self-versus-world separation recurs across neuroscience (a copy of the motor command predicting reafference), motor control, speech (predicting one's own auditory feedback), control engineering (the Smith predictor), robotics, security operations (a SIEM correlating an expected internal action against observed events), and deployment notifications (a deploy event tagged so downstream alerts know it was self-caused). Its structural abstraction is high (4 / 5): the architecture — fork the command, predict its sensory consequence, subtract it to isolate the external signal — is stated in medium-neutral terms, imports no approval or human ceremony, and runs in engineered and physical substrates indifferently, holding it firmly on the structural side. What holds it to a 4 is the neuroscientific vocabulary and disciplinary origin (transfer evidence 4 / 5): the structural transfer is strong and documented, but each domain adopts the efference-copy lexicon rather than already owning it.

  • Composite substrate independence — 4 / 5
  • Domain breadth — 4 / 5
  • Structural abstraction — 4 / 5
  • Transfer evidence — 4 / 5

Relationships to Other Primes

One-hop neighborhood: parents above, mutual partners to the right, children below.Efference Copysubsumption: Predictive CodingPredictiveCoding

Parents (1) — more general patterns this builds on

  • Efference Copy is a kind of, typical Predictive Coding

    The file: 'not predictive coding in general; efference copy is the specific self-versus-world attribution built from broadcasting a command copy to the perceiver.' predictive_coding (predict input, propagate error) is the genus; efference copy is the self-attribution specialization.

Path to root: Efference Copy → Predictive Coding → Feedback

Neighborhood in Abstraction Space

Efference Copy sits in a sparse region of abstraction space (97th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely rather than landing on a neighbor.

Family — Anticipation & Forward Models (15 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-06-14

Not to Be Confused With

The most important contrast is with feedback, because efference copy is defined against it. Feedback is retrospective: the system acts, the world responds, the output deviation is sensed, and a correction is sent back — the loop closes only after the consequences arrive. Efference copy is prospective: at the moment of issuing a command it broadcasts an internal copy sideways to the perceiver, which predicts the self-caused consequence and subtracts it before the sensation is processed. The boundary blurs when delays are short — fast feedback can mimic prediction — but it becomes decisive under significant delay, where feedback cannot disentangle a delayed self-effect from a new world disturbance, and only a forward model can. The practical consequence is exactly the Smith-predictor argument: a system acting under perceptual delay needs a forward model, not faster feedback, because no amount of after-the-fact correction can solve a self-versus-world attribution problem. Reaching for feedback where efference copy is required produces oscillation or persistent mis-attribution under delay.

A second confusion is with signal_extraction (and broad predictive comparison). Both separate a wanted component from a mixed signal, and both rely on a model of what to expect. But signal extraction is agnostic about the source of the separated component — it pulls signal from noise by whatever statistical structure distinguishes them. Efference copy is specifically the separation of the self-caused component, and it does so by a command copy — a forward broadcast of the system's own action — not by general statistical modelling. The distinguishing requirement is that the cancelled component be self-generated and the prediction driven by the system's own command. A receiving domain that treats efference copy as generic signal extraction will look for any separable component and any predictor, missing the load-bearing specifics: the concurrent command broadcast and the self-versus-world attribution it uniquely enables.

Finally, efference copy is distinct from the observer_effect, with which it shares the theme of an agent's own activity entangling with what it perceives. But the two run in opposite directions. The observer effect is the problem — the act of measuring disturbs the measured quantity, contaminating the reading. Efference copy is a solution to a related problem — it cancels the self-caused disturbance by predicting it from a command copy and subtracting it, leaving the world-caused remainder clean. Where the observer effect names an unavoidable contamination of measurement by action, efference copy names the architecture that removes the self-caused contamination so genuine external events stand out. Confusing them would lead one to treat a solvable attribution problem (cancellable via a forward channel) as an irreducible measurement disturbance, or to overlook that the very self-effect one might lament is exactly what an efference-copy mechanism is built to subtract.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.