Feedback is information returned to an actor or system about the relationship between its current state, recent action or output and a target, expectation or environment.
A marked paper, error message, thermostat reading, customer review, coach’s comment, sensor signal, audit finding and performance dashboard can all provide feedback. But feedback differs in purpose, timing, evidence, specificity, authority and whether it actually changes future behaviour.
Good classification therefore separates feedback from signal, reward, instruction and outcome. A signal carries information; feedback is information returned into the system in a way that can influence subsequent action. A reward changes payoff; feedback can be purely informational. An instruction tells what to do; feedback tells something about what happened or how current performance relates to a target.
Quick answer: how should feedback be categorised?
- Source: self, teacher, peer, customer, expert, sensor, system, audit or environment?
- Purpose: inform, evaluate, correct, reinforce, warn, calibrate or diagnose?
- Object: action, process, output, strategy, behaviour, state or outcome?
- Timing: immediate, delayed, scheduled, continuous or event-triggered?
- Specificity: broad judgement, criterion-level, error-level or action-level?
- Evidence: observation, measurement, comparison, rule, standard or interpretation?
- Actionability: can the receiver identify a useful next action?
- Loop structure: open-loop observation or closed-loop adaptation?
- Delay: how long between action, consequence and returned information?
- Noise: how reliable and stable is the signal?
- Consequence: what changes if the feedback is acted on?
- Review: how is feedback quality itself tested?
This page complements How to Categorise Signals, How to Categorise Systems, How to Categorise Controls, How to Categorise Outcomes and How to Categorise Incentives.
Feedback is not signal
A signal is any information-bearing input. Feedback is specifically returned information connected to prior action, state or output and capable of influencing what happens next.
Feedback is not reward
A high mark can be both feedback and reward, but the two functions differ. As feedback, the mark indicates performance relative to a criterion. As reward, it may alter motivation or future behaviour through payoff.
Feedback is not instruction
“Your evidence does not support the conclusion” is feedback. “Add a second independent source and explain the inference” is an instruction. Strong feedback can include or generate instructions, but diagnosis and next action remain distinct.
Feedback is not outcome
An outcome is what happened. Feedback is information about that outcome or the path toward it. The same outcome can generate different feedback depending on target, interpretation and user need.
Informational feedback reduces uncertainty
“The file uploaded successfully”, “your answer is correct” and “the temperature is now 72°C” primarily tell the receiver what state exists.
Evaluative feedback judges quality
Grades, ratings, performance reviews and audit conclusions compare work with a standard, expectation or peer group. Evaluation can summarise performance but may be too coarse to support improvement.
Corrective feedback identifies a gap
Corrective feedback explains what is wrong, missing or misaligned and points toward repair. It is most useful when the receiver can connect the correction to a controllable action.
Reinforcing feedback confirms effective behaviour
Feedback is not only about errors. Identifying what worked can help preserve strong habits, methods and decision patterns.
Diagnostic feedback searches for the first weak link
A student’s wrong answer may result from vocabulary, prerequisite knowledge, careless arithmetic, method selection or time pressure. Diagnostic feedback seeks the underlying source rather than merely marking the surface error.
Calibrating feedback compares judgement with reality
Prediction accuracy, confidence calibration and estimate error help actors improve not only outcomes but their own sense of uncertainty and competence.
Warning feedback signals approach to a boundary
Alerts, amber indicators and early warnings tell an actor that current behaviour may soon cross a threshold even if failure has not yet occurred.
Immediate feedback shortens the learning loop
When feedback arrives quickly, the action and consequence remain easy to connect. This can support correction in skill learning, control systems and interactive software.
Delayed feedback can support deeper reflection
Immediate correction is not always best. Some learning tasks benefit from productive struggle, delayed review or self-checking before external feedback arrives. Timing should match the cognitive or operational purpose.
Continuous feedback streams update constantly
Dashboards, sensors and live monitoring can return information continuously. Continuous feedback is useful for fast-moving systems but can create noise, distraction or overreaction if every fluctuation triggers action.
Periodic feedback supports slower cycles
Weekly reviews, monthly reports and termly assessments suit systems where underlying change occurs more slowly and constant intervention would add little value.
Event-triggered feedback activates on defined conditions
An alert after threshold crossing, failed validation or unusual behaviour provides feedback only when a meaningful state occurs.
Broad feedback summarises overall performance
“Good work” or “needs improvement” may influence motivation but offers little diagnostic value. Broad feedback is often best paired with more specific evidence.
Criterion-level feedback identifies the dimension that matters
Accuracy is high but explanation is weak; timing is good but evidence is incomplete. Separating dimensions helps avoid one overall score hiding the first weak link.
Error-level feedback points to a specific failure
“You changed the sign when moving the term” is more actionable than “algebra weak”. Error-level feedback should still connect the local mistake to the underlying concept where possible.
Action-level feedback suggests the next controllable move
“Before simplifying, write one line showing the inverse operation on both sides” translates diagnosis into behaviour the learner can execute.
Process feedback focuses on method
An outcome may be correct through luck or fragile strategy. Process feedback examines whether the reasoning, sequence, checking or control path is reliable.
Outcome feedback focuses on result
Result feedback answers whether the target was met. It is necessary but often insufficient for learning because it may not reveal why the result occurred.
Self-feedback comes from observing one’s own performance
Checking, reflection and metacognition help actors detect error without waiting for another person. Self-feedback improves with clear criteria and accurate self-observation.
Peer feedback adds another perspective
Peers can notice clarity, usability and misunderstanding that the producer misses. Peer feedback quality depends on competence, criteria and independence.
Expert feedback adds domain judgement
Experts can diagnose patterns and hidden assumptions, but expertise should remain domain-bounded. Strong feedback identifies the evidence and criterion rather than relying only on authority.
System feedback comes from instrumentation
Error codes, test results, metrics and sensors can provide consistent feedback, but they are limited to what the system measures and how those measures are interpreted.
Customer feedback reveals user experience
Reviews, complaints and support tickets reveal real friction, but respondents are self-selected and may not represent all users. Qualitative feedback should be linked with behavioural and operational evidence when possible.
Feedback quality depends on a target or standard
“Too slow” is difficult to interpret without a target latency or user need. Feedback becomes stronger when the comparison point is explicit.
Feedback should distinguish evidence from interpretation
“Three of five users abandoned the form at step four” is observation. “Step four is confusing” is an interpretation. Both can be useful if the boundary remains visible.
Feedback specificity should match user capability
Experts can use compressed feedback; novices may need explicit examples and next steps. Overly detailed feedback can overwhelm, while vague feedback can leave the receiver unable to act.
Feedback should be actionable when improvement is the goal
A useful receiver should know what can be changed, what must remain fixed and what evidence will show improvement. Feedback about uncontrollable traits is less useful than feedback about controllable strategies and behaviours.
Feedback can be accurate but mistimed
Detailed correction during a safety-critical action may distract from immediate control. Feedback timing should consider task phase, cognitive load and consequence.
Feedback delay changes system stability
In control systems, delayed feedback can cause overshoot, oscillation or repeated correction based on stale state. Human organisations experience similar problems when metrics arrive long after decisions were made.
Overcorrection is a feedback failure
If every small fluctuation triggers a large response, the system may oscillate. Thresholds, smoothing and review cadence can prevent noise from driving unstable action.
Underreaction is also a feedback failure
Feedback can be received and ignored because authority, incentives, resources or ownership are missing. A feedback channel without an action path is informational but not operationally closed.
Closed-loop feedback changes future action
A closed loop senses current state, compares it with target, adjusts action and observes the new state. Learning, thermostats, quality control and adaptive planning all use versions of this structure.
Open-loop reporting does not guarantee adaptation
A dashboard can display poor performance while no owner changes anything. Reporting becomes feedback only operationally when the returned information enters a decision or control loop.
Positive feedback in systems amplifies change
In systems language, positive feedback means reinforcing the direction of change, not necessarily praise. More users attract more content, which attracts more users; rising panic causes selling, which creates more price decline.
Negative feedback in systems resists deviation
Negative feedback pushes the system toward a target or stable range: a thermostat corrects temperature deviation; a teacher adjusts task difficulty when performance is too high or too low.
Praise and criticism are different from positive and negative feedback loops
Everyday language uses positive feedback to mean praise and negative feedback to mean criticism. Systems language uses the terms for reinforcing or balancing dynamics. Classification should preserve which meaning is intended.
Feedback noise can mislead action
Random variation, biased samples, inconsistent markers and instrument error can create noisy feedback. Repeated evidence, calibration and uncertainty estimates reduce overreaction.
Feedback bias can be systematic
Customers who complain, students who ask questions and employees who complete surveys may differ from silent groups. Feedback collection should examine who is missing as well as what respondents say.
Conflicting feedback should be localised
One source may value speed while another values depth. Disagreement can reflect different goals, contexts or standards rather than one source being wrong.
Feedback should identify source authority
A regulator, customer, teacher, peer and automated checker do not carry identical authority. Source type helps the receiver decide what must change and what should be interpreted as one perspective.
Feedback should not become personal attack
“This argument lacks independent evidence” targets work. “You are bad at reasoning” targets identity. Work-focused feedback is usually more actionable and less likely to create defensiveness.
Too much feedback can reduce learning
If every error is corrected instantly, learners may stop monitoring themselves. Feedback frequency should support increasing independence rather than permanent dependence.
Feedback can fade as competence grows
Novices often need explicit guidance; experts benefit from less frequent, higher-level feedback. Fading support shifts responsibility toward self-monitoring.
Feedback literacy matters
Receivers need to interpret, prioritise and use feedback rather than merely receive it. Teaching learners how to compare feedback with criteria and choose the next action improves the loop.
Feedback providers also need calibration
Markers, reviewers and evaluators should compare judgements across cases, standards and other competent reviewers. Inconsistent feedback weakens trust and learning.
Feedback effectiveness should be measured
The test is not whether feedback was delivered but whether it improved understanding, behaviour, output or system stability. Measure uptake, correction, transfer and later recurrence of the same error.
Repeated identical feedback signals a deeper problem
If the same correction appears repeatedly, the issue may be weak prerequisite knowledge, unclear instruction, poor incentives, inaccessible tools or a feedback design that does not produce action.
Educational feedback should separate mark from diagnosis
A score tells a learner where performance landed. Diagnosis explains why. Corrective feedback identifies what to repair. A next task verifies whether the repair transferred. These are separate stages of a strong learning loop.
Good feedback can reduce unnecessary practice
If one discriminating question reveals the weak prerequisite, practice can target that point instead of repeating entire chapters. High-quality feedback narrows the search space.
Organisational feedback requires psychological and operational safety
People may hide bad news when delivering it is punished. Feedback systems should make truthful reporting safer than concealment, especially for early warnings and near misses.
Metrics are feedback only when interpreted well
A dashboard metric can encourage useful adaptation or produce metric chasing. Feedback design should preserve the underlying goal, uncertainty and known proxy limitations.
AI can generate feedback at scale
Models can identify patterns, explain errors and propose revisions quickly. But automated feedback can be confidently wrong, inconsistent with the rubric or too generic to be useful. Source, evidence and review status should remain visible where consequence matters.
AI feedback should distinguish observation from inference
“The response contains no quoted evidence” is observable. “The student does not understand the text” is an inference. Good machine feedback preserves that distinction and asks for discriminating evidence before escalating diagnosis.
Agent feedback loops need stop conditions
Automated systems that repeatedly adjust themselves can oscillate or optimise the wrong proxy. Hard bounds, monitoring, review and rollback prevent uncontrolled feedback cycles.
A practical feedback record
- feedback ID and title;
- source and authority;
- recipient or target system;
- object: action, process, output, state or outcome;
- purpose;
- reference standard or target;
- observation and evidence;
- interpretation;
- specificity level;
- timing and delay;
- reliability and noise;
- recommended or available action;
- loop state: reported, received, acted on, verified;
- effect on later performance;
- conflicts with other feedback;
- reviewer or calibration status;
- version and date.
Worked example: feedback on a mathematics error
A student writes 2x + 3 = 7, then x = 5. “Wrong” is evaluative but not diagnostic. “You subtracted 3 incorrectly” is more specific. If the student repeats the error across several questions, the deeper feedback may be that inverse operations or signed-number fluency is unstable.
The strongest loop then gives a discriminating task, observes the response, repairs the first weak link and retests transfer. Feedback becomes part of diagnosis rather than a comment attached after the work is finished.
Worked example: operational feedback
A service dashboard shows latency rising. That signal becomes feedback when an owner compares it with the target, identifies the cause, adjusts capacity or configuration and verifies whether latency returns to the acceptable range.
Questions to ask before accepting feedback
- What action, state or output is the feedback about?
- What target or criterion is being used?
- What was directly observed?
- What is interpretation?
- How reliable is the evidence?
- How quickly did the feedback arrive?
- Is it specific enough to support action?
- Can the receiver control the recommended change?
- What other feedback conflicts with it?
- Who has authority to act?
- Was the feedback acted on?
- Did later performance improve?
- What would show the feedback itself was wrong or unhelpful?
Common classification mistakes
- Confusing feedback with reward or punishment.
- Giving evaluation without diagnosis.
- Giving diagnosis without an actionable next step.
- Using broad identity labels instead of work-focused evidence.
- Ignoring delay and stale information.
- Overreacting to noisy feedback.
- Treating every metric movement as meaningful.
- Delivering feedback without an owner or action path.
- Correcting so frequently that self-monitoring never develops.
- Assuming customer or survey feedback represents silent users.
- Failing to test whether feedback actually improved performance.
- Accepting AI feedback without checking its evidence or rubric.
The deeper idea
Feedback is information with a return path. Its purpose is not merely to describe the past but to improve the next state of the learner, process or system.
That is why the best feedback systems preserve target, observation, interpretation, timing, action and verification as separate steps. When one of these is missing, feedback can become noise, judgement or reporting rather than a genuine adaptive loop.
To categorise feedback well is to preserve what it is responding to, which target or standard it uses, what evidence supports it, how quickly and specifically it arrives, what action it enables, and whether the resulting change is later verified.
Final answer
Categorise feedback by source, purpose, object, timing, specificity, evidence, actionability, loop structure, delay, noise and consequence. Keep feedback separate from signals, rewards, instructions and outcomes, and judge its quality by whether it helps the receiver adapt accurately rather than merely by whether a comment or metric was delivered.