Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Formative Assessment and Instructional Feedback

Abstract. Formative assessment and instructional feedback constitute a recurring system in which evidence about learners' current understanding is interpreted and used to adjust teaching, study, and subsequent performance. This fundamental research monograph investigates the historical development, scientific definition, psychological mechanisms, measurement requirements, classroom practices, and evidence base of assessment for learning. Its critical narrative synthesis examines foundational experiments on evaluation and motivation, theoretical accounts of standards and self-regulation, investigations of classroom questioning and dialogue, and quantitative reviews of feedback interventions. It differentiates informational content from instructional change, immediate performance from durable learning, and associative findings from causal contrasts. Major evidence syntheses show variable positive mean effects alongside substantial heterogeneity, negative effects in some conditions, and important uncertainties about implementation and transfer. The analysis develops a testable description of the diagnostic and instructional sequence linking elicitation, interpretation, action, and independent learning. It addresses feedback quality and fidelity, construct validity, peer and self-assessment, linguistic and developmental differences, equity, and the governance of student information. Finally, it evaluates recent research on artificial and generative-AI feedback in the Artificial Era, identifying needs for reliable evidence, human oversight, independent outcome measurement, and longitudinal trials. This work is a critical narrative monograph of existing research; it reports neither a new experiment nor an independently conducted systematic review or meta-analysis.

Keywords: formative assessment; instructional feedback; assessment for learning; classroom questioning; feedback literacy; evaluative judgment; self-regulated learning; feedback intervention theory; metacognition; generative AI feedback; educational measurement

Introduction: The Scientific Problem of Learning-Guided Assessment

Formative assessment is the use of evidence about learning to inform decisions that can improve subsequent learning. Instructional feedback is information generated through an encounter with performance, understanding, or a learning process that may help teachers or learners adjust their next action. Together they form a recurring educational process: articulate a learning aim; elicit evidence; interpret it against a defensible model of knowledge or skill; communicate or otherwise act on the diagnosis; and examine whether a subsequent attempt shows change. A quiz, a teacher comment, a rubric, or a digital dashboard becomes formative through its consequences within instruction. The scientific unit of analysis is therefore a sequence of interactions and decisions, rather than the mere presence of an assessment instrument (Black & Wiliam (2009); Sadler (1989)).

This monograph examines six connected research problems. How can brief observations reveal learners' evolving conceptual states? What information constitutes usable feedback? Which cognitive, motivational, social, and self-regulatory mechanisms explain differences in response? How strong is the experimental evidence for classroom formative assessment as a coherent intervention? How should researchers distinguish feedback information from instructional change, and short-term performance from durable learning? Finally, what changes when automated systems, including generative artificial intelligence, mediate the exchange of evaluative information? These questions require analysis across educational psychology, classroom discourse, psychometrics, motivation, cognitive science, and human–artificial interaction. Their interdependence is precisely why the field cannot be captured by a single average effect size (Hattie & Timperley (2007); Wisniewski et al. (2020)).

The method is a critical narrative synthesis of historically influential theoretical texts, original experiments and observational studies, peer-reviewed meta-analyses and methodological evaluations, and recent research on technology-assisted feedback. Studies were selected for their direct bearing on the research object, historical explanatory significance, methodological distinctiveness, and evidentiary relevance; the sources do not comprise a protocol-defined exhaustive search. Numerical findings are attributed to the particular population, design, and outcome of the cited research. The analysis is a research monograph and does not present a newly conducted experiment, registered systematic review, or independent pooled estimate.

Definition, Scope, and Scientific Identity

An assessment event yields evidence about performance or understanding, whereas formative assessment describes the use of that evidence in an instructional process. Summative assessment ordinarily supports retrospective certification, reporting, selection, or accountability decisions. The same task may serve formative and summative functions depending on who receives the information, when it is available, and how it enters decision making. A test used to revise the next lesson can have a formative function; an ostensibly low-stakes diagnostic survey that disappears into a database does not necessarily complete a formative cycle. This functional distinction appears in Black & Wiliam (2009) and clarifies why classifying activities by product names, assessment frequency, or score type is scientifically inadequate.

Instructional feedback may take the form of knowledge of response accuracy, identification of a faulty step, a worked explanation, comparison with success criteria, targeted prompting, a peer question, an opportunity to revise, or information that prompts a teacher to replace an ineffective explanation. Feedback thus crosses a boundary between information and action. An objectively accurate comment can remain educationally inert when the learner cannot interpret it, lacks time for revision, or does not trust the source. Conversely, a brief observation can be consequential when the teacher uses it to uncover an error shared by many students and reorganizes instruction. Shute (2008) emphasizes the variability of feedback types and the interaction among timing, task demands, and learner characteristics.

The distinctive object of this dossier is the assessment-for-learning cycle and its informational and instructional feedback mechanisms. It intersects with learning theory, cognitive load, self-regulation, motivation, classroom interaction, instructional design, and educational measurement, but it is conceptually narrower than the total psychology of learning. It also differs from psychoeducational evaluation for disability eligibility, clinical assessment, standardized placement, and school accountability. Those activities have their own validity arguments, professional obligations, and consequences. They enter this monograph only when evidence from them is repurposed for actual teaching and learner decision making.

Within Psychology by Bogdanova, the principal scientific home is Educational and School Psychology. Learning and Behavioral Psychology addresses general acquisition and reinforcement; Cognitive Psychology addresses representation, retrieval, and metacognition; Research Methods and Psychometrics addresses measurement and inference. The present dossier synthesizes their intersection at the point where evidence is interpreted and turned into a change in instruction. Its scope is historical, theoretical, empirical, operational, and technological; every explanatory claim must remain distinguishable from a specific intervention outcome.

Historical Formation: From Measurement to Assessment for Learning

Educational evaluation predates contemporary formative-assessment terminology, and feedback has a separate history in behavioral psychology, cybernetics, and programmed instruction. Twentieth-century test research often emphasized how outcomes could be scored and compared, whereas teaching-oriented feedback research investigated whether learners could use information from a response to correct later performance. The later classroom literature connected these strands to a social and institutional question: when evidence exposes a discrepancy between current and intended competence, what changes in classroom interaction, lesson design, and the learner's own judgment? The early review by Bangert-Drowns et al. (1991) retained an error-correction perspective while documenting variation across testing conditions, instructional materials, and modes of response.

Butler (1988) already demonstrated why informational feedback could not be reduced to a message's evaluative valence. In a school-based manipulation of numerical grades, individual comments, and their combination, learners' interest and performance responded differently to task-oriented and ego-oriented evaluation. The research became influential because it situated feedback inside a motivational context. A comment that guides improvement has a different social meaning from a judgment that positions the learner in a status hierarchy, even when both concern the same piece of work. Experimental conditions cannot be generalized unchanged to every age, task, or grading system, but they made the role of interpretation empirically visible.

Sadler (1989) placed standards, judgment, and the capacity to take corrective action at the center of formative assessment. Black & Wiliam (1998) subsequently assembled a large body of classroom and feedback literature into a widely influential argument that learning gains were possible when assessment practices changed. Their review catalyzed research and policy attention, but subsequent methodological criticism established that heterogeneous interventions and weak designs could not sustain a single portable effect estimate. A historical account must therefore recognize both the field-building importance of the 1998 synthesis and the need to re-evaluate its quantitative claims against later, more stringent analyses (Kingston & Nash (2011); Bennett (2011)).

In the 2000s, the field integrated learning intentions, success criteria, classroom questioning, feedback, peer involvement, and self-regulation. Nicol & Macfarlane-Dick (2006) argued for the learner's active production and use of feedback, while Black & Wiliam (2009) presented a more explicit theoretical rationale for formative practice within pedagogy and classroom discourse. Research then broadened to computer-mediated assessment, feedback literacy, process tracing, and automated explanations. These developments form a history of progressively richer models of the learner, rather than a simple chronology of new assessment devices.

Standards, Quality Judgment, and the Closing of Learning Gaps

Sadler's account treats the learner's grasp of quality as an instructional achievement in its own right. To improve complex work, a learner must recognize what counts as good performance, compare current production against a relevant standard, and command strategies for reducing the discrepancy. This sequence is especially important for essays, scientific explanations, mathematical arguments, creative projects, and professional judgments that admit multiple dimensions of quality. Correctness messages alone may reveal whether a response matches an expected answer while leaving the quality of reasoning unexplained. Sadler (1989) regarded the development of evaluative expertise as essential for progressively independent performance.

Learning goals and success criteria help make this interpretive process explicit. A goal such as “reason quantitatively about proportional relationships” provides a competence target, while examples, counterexamples, rubrics, and guided comparison show what counts as stronger reasoning. Standards must also be responsive to disciplinary knowledge: evidence of understanding in history differs from evidence in experimental science or literary interpretation. The psychometric problem is to establish that the visible response measures the intended construct; the pedagogical problem is to make the learner's response informative enough to guide action. A good score representation is insufficient if its cognitive meaning cannot be interpreted.

The word “gap” can imply a simple distance along a single scale, but many learning difficulties are structural. A student may use a coherent yet mistaken causal model, know a procedure while misunderstanding its purpose, or perform accurately on familiar examples without being able to transfer. The next instruction must therefore be matched to the kind of discrepancy. Providing the answer may address a missing fact; contrasting two worked cases may address discrimination; eliciting an explanation may uncover a misconception; a later transfer task may test whether conceptual reconstruction occurred. This interpretation links formative assessment to the broader distinction between performance and learning developed by Soderstrom & Bjork (2015).

Effective closure of a gap requires an opportunity to attempt, interpret, revise, and test again. The feedback loop fails when information arrives after the curriculum has moved on, when success criteria remain opaque, or when the student is required to read comments but receives no meaningful opportunity to act. A rigorous investigation should observe the intermediary decisions: Did the teacher change an example or pacing? Did the student revise a strategy? Did an error recur on a delayed task? These observations transform an appealing cyclical metaphor into potentially testable causal and process hypotheses.

Classroom Interaction and the Social Construction of Feedback

Classrooms are collective environments in which assessment evidence emerges through talk, gesture, written work, shared artifacts, and participation norms. Ruiz-Primo and Furtak's 2007 classroom study analyzed informal formative assessment through elicitation, student response, recognition, and instructional use. Their ESRU model illustrated the difference between asking a question and exploiting its evidentiary value. A teacher may elicit an answer and evaluate it without learning why it was produced; a more informative exchange probes reasoning, recognizes the underlying conceptual model, and uses the response to direct the next question or explanation.

The evidence is shaped by the circumstances of elicitation. Public answering can amplify reputational pressure; short-answer tasks may conceal reasoning; teacher follow-up may privilege a narrow set of acceptable formulations; linguistic demands may be mistaken for disciplinary ignorance. Furtak and Ruiz-Primo (2008) examined how different science prompts elicited different varieties of student thinking in written and oral settings. Their findings caution against treating a single response modality as a transparent view of cognition. What appears measurable is partly determined by task openness, familiarity, and the structure of classroom discussion.

The recipient also matters. A private written comment, public whole-class explanation, anonymous voting pattern, and peer dialogue create different opportunities for uptake. Learners may interpret a teacher's statement in light of prior relationships, perceived competence, cultural norms, and classroom expectations. The teacher's access to student reasoning is likewise partial and institutionally constrained. Formative practice must coordinate diagnostic accuracy with communicative clarity, emotional safety, and a credible path to improvement. Studies of children's interpretations of feedback demonstrate that the meaning attributed by recipients deserves measurement alongside the teacher's intentions (Tunstall & Gipps (1996)).

A scientifically complete account therefore treats formative assessment as a distributed activity with informational, cognitive, motivational, and relational components. Teachers provide opportunities, interpret evidence, and alter instruction; students disclose partial understanding, judge feedback, and attempt change; peers act as alternative audiences and evaluators; tools store, transform, and present traces of performance. No participant has perfect access to the state of learning. Each observed response is evidence under an explicit or implicit model, and each subsequent decision becomes another part of the learning environment.

Feedback Intervention Theory and the Allocation of Attention

Feedback intervention theory proposes that a response to feedback depends partly on where attention moves after an evaluative event. Kluger & DeNisi (1996) synthesized a broad performance-feedback literature and distinguished task-learning processes, task motivation, and self-related or meta-task processes. Their argument links outcome variability to the mechanisms that a message activates: information that helps solve a problem may support learning, while information that invites global comparison of personal worth may divert attention from the task. The theory provides a mechanism for the paradox that feedback can be accurate, salient, and well intentioned yet produce a negative effect.

This distinction should not be converted into a universal prohibition on praise or evaluation. The original meta-analysis included many settings beyond school classrooms; the meaning of a message depends on learners' goals, task complexity, and the credibility of the source. A teacher who says “You have selected the right equation; now explain why these quantities are proportional” directs attention to structure and action. A teacher who says “You are excellent at mathematics” communicates a different kind of information. Whether the first statement changes durable achievement depends on subsequent problem solving, not merely on the analyst's preferred classification. Attention is a hypothesized mediator that requires direct observation or manipulation.

The central methodological implication is that a feedback treatment should be coded by its operational content, not assigned a binary present/absent label. Such coding can distinguish diagnostic explanations, affective reassurance, social comparisons, hints, procedural corrections, and opportunities for response. The same written sentence can activate different processes across learners, so experimental manipulation of message type should be supplemented with evidence about interpretation and uptake. Subsequent reviews of feedback models reaffirm the importance of examining source, message, learner, and context together (Lipnevich & Panadero (2021)).

Task, Process, Self-Regulation, and Personal Evaluation

Hattie & Timperley (2007) differentiates feedback directed at the task, the processes used to accomplish it, the student's self-regulation, and the person. Task-level information can identify whether a claim is supported by evidence; process-level information can explain how to choose a more appropriate strategy; self-regulatory information can help the student monitor and decide independently. Self-directed praise or criticism is distinguished because it may communicate comparatively little about how to improve the work. The model also organizes feedback around questions about the intended destination, current position, and subsequent action.

These categories are useful analytical lenses rather than mechanically separable boxes. A single comment may ask a learner to check an inference, explain why a strategy worked, and plan a different approach for the next problem. The question is which kind of information becomes actionable at a particular stage of learning. A novice may need an explicit correction or demonstration; a more advanced learner may gain more from identifying a principle, comparing alternative arguments, or testing an independently chosen method. Feedback design must therefore consider prior knowledge, representational difficulty, and the learner's ability to evaluate the response.

A persistent error in educational discussions is the presumption that “more detailed feedback” is always superior. Details impose reading, interpretation, and working-memory costs. An exhaustive annotation on an essay may obscure the one structural revision that would produce the greatest improvement. A hint that triggers productive search may be preferable to a full worked solution for a sufficiently knowledgeable learner, while the same hint can leave a novice stranded. Shute (2008) reviewed the interactions among feedback complexity, learner characteristics, and task attributes. The design problem is to match information to the learner's state and the next realistic action.

Cognitive Mechanisms: Error Detection, Representation, and Transfer

Feedback can support learning by detecting an incorrect response, replacing a mistaken rule, discriminating between similar cases, prompting retrieval, elaborating a causal explanation, or reorganizing a representation. These mechanisms have different evidentiary signatures. A correction of factual recall may produce a higher score on the next recognition test without changing conceptual reasoning. A process-oriented explanation may lead to a delayed transfer advantage even if immediate performance is slower. Researchers therefore need outcome measures that reveal the intended mechanism rather than a single convenient test aggregate.

A wrong response is not necessarily noise. Errors can expose the explanatory commitments or partial patterns that a learner uses. However, only a carefully selected elicitation task may distinguish a stable misconception from a careless slip or a language misunderstanding. The teacher's interpretation should be treated as a testable inference: the proposed diagnosis predicts how the learner will respond to a discriminating item or an alternative example. Ruiz-Primo and Furtak (2007) showed how recognition and use of student thinking could be analyzed in authentic classroom interaction. A response to an informative probe is often more educationally useful than a simple aggregate score.

Retrieval practice complicates this picture because answering a question can itself change memory. Dunlosky et al. (2013) identified practice testing and distributed practice as broadly useful learning techniques across a substantial literature. When a formative assessment embeds retrieval, the observed gain may arise partly from retrieval practice and partly from the feedback that follows. An experiment seeking the incremental causal effect of feedback should compare otherwise equivalent retrieval opportunities and control for exposure to correct answers. This principle protects theory building from attributing every learning benefit of a quiz to the explanatory quality of its feedback.

Motivation, Evaluation, and the Meaning of a Learning Error

Assessment information is also a social signal about competence, belonging, and future opportunities. Learners can perceive errors as invitations to revise a strategy, evidence of inadequacy, or threats to standing within a group. Butler (1988) contrasted task-involving comments with ego-involving numerical evaluation in a controlled school study. The result was influential because it showed that feedback format and the meaning of evaluation can alter both performance and interest. The interpretation extends more broadly to the design of safe learning opportunities but does not prove that every numerical grade harms motivation.

The emotional response to corrective information may influence attention, effort, avoidance, and the willingness to seek help. Learning environments that publicly rank students, attach high stakes to early errors, or make revisions impossible can suppress the very diagnostic responses formative systems need. However, a pleasant emotional reaction does not establish learning. A reassuring statement may protect engagement while failing to correct the underlying misconception; a demanding but comprehensible critique may initially lower self-reported confidence while producing later improvement. Both affective outcomes and independent achievement should be measured.

Motivation is neither an exogenous background characteristic nor an automatic consequence of message positivity. Goals, task value, confidence, peer norms, and the anticipated consequences of mistakes influence whether learners invest effort in using feedback. Researchers examining motivational mediators should track feedback seeking, time spent revising, the quality of the revision, and subsequent independent performance. Such measurements help distinguish a short-lived emotional reaction from an intervention that changes the regulation of future learning.

Self-Regulated Learning and Feedback Literacy

Self-regulated learning involves planning, monitoring, selecting strategies, evaluating outcomes, and adapting future actions. From this perspective, feedback is not exhausted by information delivered by a teacher; learners generate information through comparison, retrieval attempts, self-explanation, and inspection of their own work. Nicol & Macfarlane-Dick (2006) presented seven principles intended to connect formative assessment with learner control. Their model locates feedback inside a system of internal and external judgments rather than restricting it to comments after grading.

The concept of feedback literacy gives this account a more explicit social and dispositional form. Carless & Boud (2018) describe the understandings and capacities that allow learners to appreciate feedback, make evaluative judgments, manage affect, and take action. An accessible comment may still fail when the learner cannot translate its general principle into a concrete revision. Feedback literacy develops through structured opportunities to examine exemplars, compare judgments, discuss criteria, and show how an earlier critique affected a later attempt.

This theory creates an important distinction between two outcomes: the immediate improvement of a particular product and the development of the capacity to judge and improve future work independently. A teacher may make a paragraph better by supplying its missing sentence, while the student learns little about how to identify the defect next time. A learner who uses a criterion to diagnose an analogous problem in a later independent task demonstrates a stronger form of transfer. Longitudinal research should assess this capability with novel tasks and reduced scaffolding. The strongest account of formative assessment is one in which dependence on external information can progressively diminish as evaluative judgment develops.

Feedback as Dialogue, Interpretation, and Joint Activity

Feedback exchanges are commonly studied as information moving from a knowledgeable sender to a less knowledgeable recipient. The dialogic perspective treats the meanings of feedback as jointly interpreted and subject to revision. A student can misunderstand a comment, ask for an example, resist a judgment, or supply information that changes the teacher's diagnosis. The teacher may be correct about an error but wrong about its cause. When feedback becomes a conversation, questions and revised explanations are observable features of the process rather than informal additions.

The educational relevance of dialogue is especially clear in extended assignments and authentic performances. A peer or instructor may challenge a claim and require its defense, revealing the student's reasoning and changing what the next feedback should address. Dawson et al. (2019) investigated perceptions of effective feedback among students and staff, illustrating the value of examining how feedback is experienced and interpreted. Such perceptions are necessary evidence about implementation and acceptability but should be distinguished from independently assessed competence gains.

Dialogic models also expose the institutional resources required for productive use. Class size, workload, contact hours, opportunities to revise, language accessibility, and norms for disagreement all shape what forms of exchange are feasible. Digital tools can make comments rapidly available without ensuring meaningful discussion; conversely, a brief whole-class intervention based on a diagnostic pattern can change many learners' next opportunities at once. A complete psychological theory must account for the organization of interaction as well as the informational content of messages.

Eliciting Evidence and the Problem of Construct Validity

An assessment response is evidence only relative to a proposed construct and a model connecting observable performance to knowledge or competence. If the instructional target is the ability to explain a scientific mechanism, a multiple-choice response may detect a likely misconception but may not reveal the explanation that produced it. If the target is fluent retrieval of factual relationships, a lengthy essay may add language and writing demands irrelevant to the intended measurement. Designing formative tasks therefore begins with a specification of the inference to be supported and the alternative interpretations that must be considered. The teacher's need for rapid evidence does not suspend ordinary principles of validity.

The quality of an item is partly determined by how strongly it separates plausible mental models. A distractor that represents a documented misconception can be diagnostically meaningful; an ambiguous answer that is wrong for several different reasons has limited instructional resolution. Multiple representations, follow-up questions, explanations, and short demonstrations can triangulate the meaning of a response. This does not require psychometric precision identical to a high-stakes standardized examination in every classroom exchange. It requires matching the confidence of a diagnosis to the strength of the observed evidence and the cost of the resulting instructional decision. Bennett (2011) highlights conceptual and measurement challenges in claims about formative assessment.

Teachers and students also contribute to the evidence-generating process. An open invitation to explain may allow a student to express a partially formed idea, whereas a strongly leading question may produce the expected words without establishing understanding. Furtak and Ruiz-Primo (2008) found that different forms of formative prompt elicited different patterns of student thinking in the context of middle-school science. Such variation is intrinsic to assessment design. A classroom procedure should be assessed not simply by how many answers it collects but by the range, relevance, and interpretability of the information it exposes.

Learning Intentions, Success Criteria, and Exemplars

Learning intentions specify the developing competence that instruction seeks to change; success criteria translate that competence into observable qualities or performances. In quantitative problem solving, relevant criteria might include choosing a warranted model, explaining assumptions, showing a coherent inference, and checking the result. In argumentative writing, the criteria may include a defensible claim, evidence selection, counterargument handling, and integration. These representations help students understand what is being judged and what a revision should improve.

A rubric can make criteria visible, but a rubric is not intrinsically formative. Used only to assign a final grade, it may simply formalize evaluation. Used prospectively with exemplars, guided comparison, self-assessment, and revision, it can support the development of judgment. Panadero & Jonsson (2013) evaluated research on formative uses of scoring rubrics and identified promising pathways alongside important methodological and contextual limitations. A rubric's potential mechanism depends on how learners interpret and use it, not solely on whether a scoring table is distributed.

Exemplars provide concrete referents for abstract standards. Asking students to compare two solutions with similar scores but different reasoning can reveal evaluative criteria more clearly than reading criteria alone. However, examples can invite copying when differences in task structure are not explained. An effective sequence includes comparison, articulation of the criteria, independent application, and feedback on the quality of judgment. This design allows investigators to test whether criteria have become productive representations rather than memorized vocabulary.

Content, Specificity, and Actionability of Feedback

Feedback content can be classified according to whether it verifies correctness, supplies an answer, explains why an answer is right or wrong, offers a hint, identifies a missing step, or prompts a strategy choice. Shute (2008) showed that a range of such forms had been studied under the broad label of formative feedback. Their educational significance differs because they support different next actions. “Incorrect” identifies a mismatch; “your conclusion follows from an untested assumption” points to a reasoning operation; “compare the boundary conditions in these two examples” proposes a learning activity.

Actionability is a relational property. A technically specific instruction may be unusable to a novice who lacks necessary vocabulary or intermediate knowledge. A concise question can be highly actionable to an expert who recognizes the underlying distinction. Successful feedback design therefore requires diagnostic alignment with the learner's current state, an interpretable representation, and a feasible opportunity for implementation. Assessing actionability only through the teacher's judgment risks overlooking learner interpretation; collecting student ratings of perceived clarity without observing revision risks the opposite error.

Research coding should identify the location and target of each comment, its explanation level, reference to criteria, temporal placement, personal or comparative language, and the intended learner response. These features can then be linked to revision traces and subsequent independent performance. Panadero & Lipnevich (2022) organizes feedback around message, implementation, student, context, and agents, emphasizing that content and delivery cannot be analyzed in isolation.

Timing, Frequency, and the Architecture of Opportunity

The timing of feedback can influence whether an error is corrected, a problem is independently completed, and learning is retained. Immediate verification may prevent repeated practice of an incorrect response, but delaying an explanation can sometimes preserve opportunities for retrieval, reflection, or independent problem solving. Bangert-Drowns et al. (1991) reviewed feedback in test-like events and demonstrated that outcomes vary with feedback conditions and opportunities for mindful processing. van der Kleij et al. (2015) examined timing among several dimensions of computer-based feedback, finding substantial heterogeneity and limits to simple rules about when feedback should arrive.

Time must be understood relative to the learning activity. Feedback “immediate” at the scale of a classroom lesson may be delayed relative to a single response but early enough to affect the next problem. Feedback “delayed” by a week may arrive after the student's attention has moved to another topic, making enactment difficult. A meaningful experiment must specify the interval, task demands, planned activity during the delay, and available follow-up opportunities. Otherwise timing conditions confound information delivery with opportunities to work.

Frequency introduces another tradeoff. Very frequent correction can support accurate early acquisition but can also remove opportunities for self-diagnosis or independent monitoring. Infrequent feedback may allow maladaptive strategies to persist. The optimal schedule is not a single fixed interval; it is a policy conditioned on error severity, learner expertise, task complexity, stakes, and available time. Researchers should measure feedback exposure and actual response rather than assuming that an issued comment was processed.

Feedback Across Disciplinary and Task Contexts

Formative assessment acquires its psychological specificity through the kind of knowledge being learned. A mathematics response may reveal a misconception about proportionality, a procedural substitution error, or a weak grasp of the conditions under which an algorithm applies. A science explanation can display causal structure and relations among evidence, models, and observation. An essay may require evaluative judgments across several dimensions at once. Feedback that treats every task as a correctness check misses these different forms of knowledge.

Disciplinary differences can also alter the relationship between short-term and long-term outcomes. A learner may immediately reproduce a corrected formula but fail when a later task changes representation. In writing, revisions within one assignment may improve the product without transferring to a new genre. In laboratory science, students can learn vocabulary while retaining mistaken causal explanations. The corresponding assessment designs must measure the intended form of learning: explanation, transfer, reasoning, communication, or procedural competence. This is consistent with the learning–performance distinction articulated by Soderstrom & Bjork (2015).

Kingston & Nash (2011) reported different pooled estimates by subject area in their restricted K–12 meta-analytic sample, with larger estimates for English language arts than for mathematics or science. These comparisons were drawn from small, uneven collections of eligible studies and should be treated as exploratory moderators. They identify a design problem for future research: whether differences reflect the intrinsic affordances of feedback, the nature of the tasks and assessments, or implementation and study selection.

Formative Cycles and Instructional Adaptation

Formative assessment is often represented as a continuous loop, but classroom practice consists of nested cycles operating at different timescales. An interaction lasting thirty seconds may reveal confusion that the teacher addresses through a follow-up question. A lesson-level exit task may guide regrouping and examples for the next day. A multiweek sequence of work and revision can reveal which conceptual distinctions require renewed attention. Each cycle has a target, evidence source, interpretation rule, decision, and subsequent opportunity to observe learning.

An instructional change is the critical step that converts assessment evidence into a formative intervention. A teacher might give a clearer explanation, offer contrasting examples, revisit prerequisites, change the order of problems, form temporary groups, or invite student-generated questions. A learner might allocate more practice to a weak concept, test an alternative solution strategy, or request clarification. An evaluation should specify which decisions were possible and documented rather than counting how many diagnostic tasks were administered. Black & Wiliam (2009) emphasize that formative activity is embedded within a broader pedagogical system.

Fidelity should be understood functionally. Two teachers can implement different surface activities yet both elicit, interpret, and use evidence effectively. Conversely, two teachers can distribute the same exit ticket while only one modifies instruction on its basis. A research design can code the completeness of cycles, the diagnostic alignment of response, and the enactment of an adjustment. This is more informative than treating the number of quizzes or comments as a proxy for intervention quality. It also makes it possible to distinguish teacher professional learning from the mere adoption of classroom technology.

Original Experimental Evidence: Evaluation Format and Learning Response

Original experiments supply causal leverage that conceptual frameworks alone cannot provide. Butler (1988) compared numerical grades, individual comments, and a combination of the two after tasks completed by fifth- and sixth-grade pupils. Twelve classes were randomly assigned to feedback conditions, and the reported analysis included 132 selected students. Interest and performance were highest under task-oriented comments in the examined conditions; combining grades with comments did not reproduce the comments-only pattern. The study therefore supports a context-specific causal claim about feedback format and motivational interpretation, with clear limits related to setting, age, tasks, and the social meaning of grading.

The experiment is particularly valuable because it shows that the presence of more evaluative material cannot be assumed to create a stronger intervention. It also illustrates the need to distinguish the form of feedback from the process it evokes. The manipulated conditions changed attention, comparison, and expectations as well as information. Later studies should measure those mediators independently and include a delayed transfer criterion. A short-term advantage on the experimental tasks cannot by itself establish changes in a durable general learning capacity.

Classroom observational work offers complementary process resolution while imposing weaker causal control. Ruiz-Primo and Furtak (2007) traced elicitation, recognition, and use of responses across three middle-school science classrooms. Their study is informative about the practical organization of feedback cycles, not a randomized estimate of their effect on achievement. Furtak and Ruiz-Primo (2008) further analyzed how prompt design affected elicitation of scientific ideas. Combining these kinds of study helps identify which active components should later be randomized or experimentally contrasted.

Feedback Interventions: Early Meta-Analytic Evidence

Kluger & DeNisi (1996) assembled 607 effect sizes representing 23,663 observations across a broad feedback-intervention literature. They reported an average positive standardized effect around 0.41, while more than one-third of the interventions reduced performance. Their review is a fundamental challenge to explanations that assign an automatically positive effect to feedback. The analysis includes workplace and other performance settings as well as learning, so its pooled estimate must not be presented as the effect of classroom formative assessment.

The authors proposed that feedback can redirect attention toward task learning, task motivation, or self-related processes. In this respect, the negative effects are theoretically important observations rather than statistical anomalies to be averaged away. Subsequent evidence would need to establish which feedback features shifted attention, whether recipients adopted the proposed strategies, and what outcomes were sustained. The comparison of positive and negative effects within one synthesis underlines the necessity of preregistered moderators and context-sensitive causal explanations.

Bangert-Drowns et al. (1991) reviewed 58 effects from 40 reports on feedback in test-like events. Their analysis identified differences associated with feedback content, test and pretest arrangements, opportunities for correction, and instructional context. Such studies clarify the operation of feedback on specific questions, but the estimand is usually different from the effect of a teacher's entire classroom assessment routine. An instructor may generate a benefit through changed teaching even when the student never receives an individualized correctness statement; a computer may provide correctness immediately without affecting the next lesson. Pooling these interventions without conceptual differentiation obscures mechanism.

Formative Assessment Programs: What Can Be Concluded About Average Effects?

Black & Wiliam (1998) assembled evidence for promising gains when classroom assessment was strengthened and became a landmark in assessment-for-learning theory. Its importance lies partly in its broad argument that changing the use of evidence can change pedagogy. Yet the historically circulated claims of very large general effects require qualification. Different studies operationalized formative assessment as feedback, mastery learning, computer systems, professional development, or other instructional packages. Their outcome measures and causal designs were far from equivalent.

Kingston & Nash (2011) critically evaluated more than 300 candidate K–12 studies. Only 13 supplied sufficient interpretable information for their analysis, yielding 42 independent effect sizes. The authors reported a median observed effect of 0.25 and a random-effects weighted mean of 0.20, rather than an often-cited estimate near 0.70. Their work provides a narrower, more cautiously derived quantitative anchor, accompanied by an explicit warning that the field needed substantially better studies. Their mean is not a timeless universal constant: eligibility criteria, subject populations, and intervention definition determine the estimand.

The evidence problem is therefore two-sided. Positive classroom effects should not be dismissed merely because early studies were heterogeneous; the scientific conclusion is that the strongest claims must be grounded in credible individual designs and appropriate syntheses. Equally, broad endorsement of “formative assessment” cannot substitute for showing which instructional operations produced the effect. Bennett (2011) emphasized weaknesses in conceptualization, inference, and implementation. A modern cumulative research program must document both treatment components and how those components alter instruction.

Quantitative Synthesis of Feedback Content and Educational Outcomes

Wisniewski et al. (2020) synthesized 435 studies, 994 effect sizes, and more than 61,000 participants across educational feedback research. The article's abstract reports a medium overall effect around d = 0.48; the main results report an initial random-effects weighted average of d = 0.55 with a 95% confidence interval of 0.48–0.62. These numbers refer to different reported calculations or reporting contexts within the paper and should not be silently treated as identical estimates. The study documented substantial heterogeneity, with I-squared above 86% in its general analysis, and emphasized that informational content moderated outcomes. Seventeen percent of the extracted effects in one reported analysis were negative.

A large meta-analysis is stronger than a small series of convenience studies for detecting broad patterns, but its scale does not eliminate definitional heterogeneity, outcome dependencies, publication bias, or differences in historical schooling. The median study year in this synthesis was 1985, limiting its direct representativeness for contemporary digital classrooms. Its reported stronger effects on cognitive and motor outcomes than on motivational and behavioral outcomes also cautions against a claim that feedback works equally well across outcome domains. A scientifically informative account uses the synthesis to establish heterogeneity as a central finding and to specify better moderator hypotheses.

In the more narrowly defined domain of computer-based item feedback, van der Kleij et al. (2015) analyzed 40 studies and 70 effects. Their reported estimates were approximately 0.49 for elaborated feedback, 0.32 when the correct response was supplied, and 0.05 for knowledge of response accuracy alone. The article concluded that explanatory feedback was especially useful for higher-order learning outcomes. These contrasts are compelling as evidence about feedback content in computer-based tasks, but the categories encompass multiple interventions and do not prove that more explanation is always better for every learner or objective.

Moderator Interpretation and the Problem of Heterogeneity

Moderator analyses are often read as instructional prescriptions before their inferential strength has been established. A reported difference between language arts and mathematics may reflect different study populations, outcome reliability, observation periods, or package components. A difference between delayed and immediate feedback can be confounded by whether learners were allowed to retrieve, reflect, or revise during the interval. Subgroup estimates based on few studies may be unstable. The analysis should report uncertainty intervals, the number of studies per subgroup, and whether comparisons were preregistered or exploratory.

Outcomes also require separation. Correctness on the next item, performance on a course examination, quality of a revised essay, delayed recall, far transfer, interest, confidence, and strategic independence represent different constructs. A positive effect on one does not entail the others. Soderstrom & Bjork (2015) articulate why immediate performance can be an unreliable indicator of long-term learning. In assessment research, this distinction is especially consequential because the intervention itself can increase familiarity with the criterion test.

Study-level dependence further complicates synthesis. A trial may report several outcomes, repeated time points, or correlated achievement domains. Treating all estimates as independent can distort uncertainty. More recent three-level models attempt to address some dependence structures, but the resulting estimate remains conditioned on included studies and their risk of bias. The basic principle is to treat a meta-analytic effect as a description of an identified evidence set and causal contrast, accompanied by moderator and quality information, not as an inherent numerical property of the term “feedback.”

Contemporary Evidence Beyond the Classical Classroom

Recent research extends the feedback problem to digital environments and artificial systems. Ba et al. (2025) synthesized 129 studies of AI-assisted feedback, describing a rapidly expanding field with multiple feedback foci and mixed implementation practices. The review identified promising outcomes while also documenting incomplete transparency about the underlying algorithms. Such descriptive and outcome evidence supports research on adaptive feedback design, but the field still needs trials that separate personalization, immediate availability, explanatory content, and instructor oversight as distinct causal components.

A more focused Huang et al. (2026) included 36 experimental or quasi-experimental studies of generative-AI feedback published between 2023 and 2025. It reported an academic-achievement estimate of g = 0.61, 95% CI [0.42, 0.80], across 4,538 participants, alongside high heterogeneity (I-squared 86.3%). The estimated positive effect is relevant to instructional feedback but concerns a selected and comparatively recent collection of interventions. Its authors noted that estimated metacognitive benefits were based on very few effects and that noncognitive outcomes did not yield a robust pooled advantage. The study should not be generalized to all generative-AI systems, learners, or educational institutions.

Feedback technologies may change the scale and speed of information delivery while preserving the fundamental explanatory questions: What evidence is collected? What cognitive diagnosis is justified? Who acts on the feedback, and how? What independent learning follows? The greater capacity to generate comments heightens the importance of distinguishing the amount of feedback from the knowledge or skill that students actually acquire.

Questioning, Peer Explanation, and Whole-Class Evidence

Questions become formative when they make particular features of thinking visible and provide grounds for a subsequent instructional decision. An open question about why a model works may expose a causal explanation, while a carefully designed multiple-choice item can efficiently reveal whether a misconception is widespread. Both approaches may be appropriate, provided that the teacher understands the inference each response supports. The relevant comparison concerns the diagnostic value of the evidence, the resources needed to obtain it, and the instruction made possible by it. In Ruiz-Primo and Furtak's study of informal classroom assessment, the connection among eliciting, recognizing, and using students' ideas distinguished richer cycles from simple question-and-answer exchanges.

Peer discussion can alter the evidentiary landscape because a student may articulate a reason to a classmate that remains hidden in an individual answer. Asking students to commit to a response, compare reasoning, and then answer a new question allows the teacher to observe the relation between explanation and performance. Yet peer conversation can also propagate misconceptions or favor the most confident speaker. Its contribution should be assessed through independent post-discussion tasks and the accuracy of explanations, not by participation counts alone. A rigorous classroom experiment would distinguish the effects of retrieval, peer explanation, teacher explanation, and feedback content, with equal instructional time where feasible.

Whole-class patterns can support efficient instructional adjustment. If many students make the same error, a targeted contrasting example may be more useful than individual comments that repeat identical information. If responses divide among several explanatory patterns, the teacher may compare alternatives publicly while allowing individual reasoning to remain anonymous. The teacher should consider nonresponse: students who do not submit an answer can differ systematically from those who do. The legitimacy of an inference depends on who was observed, what the task elicited, and whether the resulting adjustment addresses the construct of interest. These principles link classroom formative practice with evidence-centered measurement.

Peer Assessment and the Development of Evaluative Judgment

Peer assessment can perform two functions: it can supply information to a recipient, and it can train the reviewer to recognize quality in another person's work. The second pathway is theoretically significant because the act of comparing evidence against criteria may generate more learning than reading a comment received from someone else. Studies should therefore separate gains among reviewers and recipients, distinguish anonymous from identifiable peer exchanges, and observe whether the judgments remain accurate on later independent tasks. A peer score that agrees with a teacher's grade does not by itself establish that the peer's explanation was useful.

Training may involve analyzing exemplars, using rubrics to justify a criterion, revising comments, and discussing disagreements about quality. The quality of peer feedback depends on disciplinary expertise, communicative skill, trust, and the organization of opportunities to respond. Panadero & Jonsson (2013) suggests that shared criteria can support self-regulation, but implementation and study designs vary. Carless & Boud (2018) place evaluative judgment and taking action at the center of feedback literacy, offering a framework for studying the developmental benefits of judging others' work.

Peer assessment also creates fairness and belonging concerns. Friendship, status, language fluency, group dynamics, and disciplinary norms can influence how a comment is generated and accepted. Investigators should observe not only average achievement but also whose work receives actionable guidance, who experiences reputational costs, and who uses the feedback. Transparent criteria and teacher supervision may mitigate some problems, yet evidence of equitable delivery requires direct assessment across learners and contexts.

Self-Assessment, Reflection, and Calibration

Self-assessment can be productive when learners compare their work against external standards and check the accuracy of their judgments. It can be misleading when interpreted merely as asking students how confident they feel. Confidence may reflect familiarity, social experience, or recent feedback rather than accurate knowledge. A useful operationalization requires a measurable target, a judgment before or after an attempt, and an independent criterion of performance. Calibration can then be studied as the correspondence between perceived and demonstrated competence.

The strongest instructional use of self-assessment couples judgment with evidence and next steps. A student may predict how well an explanation will transfer, test it on an unfamiliar question, inspect the discrepancy, and choose a study action. This procedure integrates formative assessment with metacognition and retrieval. Nicol & Macfarlane-Dick (2006) emphasize that external feedback should help develop internal regulation; Soderstrom & Bjork (2015) explain why a feeling of fluent performance can be an inadequate indicator of durable learning.

An effective experiment would measure self-assessment accuracy, revision behavior, delayed independent learning, and the amount of external scaffolding still needed. Merely asking students to reflect in a journal can introduce writing practice and time-on-task effects. The assessment component becomes identifiable when judgments are tested against subsequent performance and used to choose actions. Self-assessment can thus be understood as a learned skill with developmental variation, rather than an automatic benefit of a reflective prompt.

Mastery Learning, Differentiation, and Feedback-Guided Instruction

Mastery-oriented systems use evidence about competence to determine whether instruction should progress, repeat, or change form. Formative feedback is one informational mechanism within such systems, but their causal packages may also include additional time, tutoring, alternative explanations, and criterion-based reassessment. Attributing all observed benefit to feedback would misidentify the treatment. To establish the incremental contribution of information, researchers must compare conditions that hold the amount of instructional time and opportunity for practice as constant as possible.

Differentiation presents a related challenge. Teachers may identify groups that need examples, vocabulary support, advanced tasks, or alternative representations. The meaningful research question is whether the diagnosis was accurate and the consequent adaptation improved learning compared with a realistic alternative policy. If a group receives more time, more instructor attention, or different content, the observed effect cannot be uniquely attributed to “personalization.” The local scientific map of Psychology by Bogdanova rightly treats Mastery Learning and Differentiated Instruction as a neighboring dossier with its own intervention object, while the present inquiry concerns assessment information and feedback decisions.

Continuous adjustment should not become perpetual testing. Collecting evidence imposes time, attention, and emotional costs, and teachers face finite opportunities to interpret it. A useful diagnostic should have sufficient expected instructional value to justify those costs. This can be examined through decision analyses: how often does an assessment change instruction, which changes are effective, and whether a simpler signal would have yielded an equally good decision. The scientific goal is better instructional action, not maximum surveillance of performance.

Computer-Based Feedback and Learning Analytics

Computer-based environments can standardize the timing and format of responses, log sequences of attempts, and deliver explanations contingent on selected answers. These features support controlled research because feedback types can be assigned reproducibly and learners' subsequent actions can be recorded. van der Kleij et al. (2015) showed that elaborate feedback, correct-answer information, and verification-only feedback were associated with different average outcomes in the studies they synthesized. The important design distinction is informational content and its interaction with task demands, rather than a presumed benefit of placing the intervention on a screen.

Learning analytics may display predictive indicators, patterns of error, or inferred risk states. Such outputs are only as defensible as the observations, populations, and models from which they were derived. A prediction of lower future achievement does not identify which explanation or assignment would help a learner; a cluster of incorrect answers does not necessarily identify a single misconception. To become formative, analytics must be linked to an interpretable diagnosis, a feasible teaching response, and evidence that the response works. The transformation from prediction to intervention is an empirical step requiring separate validation.

A digital dashboard may therefore improve the speed of access to evidence while contributing little to instructional change if teachers lack time or confidence to interpret it. Researchers should collect implementation logs, teacher interviews, decisions, and learning outcomes, and should avoid treating dashboard clicks as equivalent to pedagogical use. Systems can support a teacher's distributed attention, but the teacher's professional judgment remains part of the causal chain being studied. Automatically generated visual summaries must be evaluated as representations of fallible data, not authoritative portraits of individual minds.

Generative Artificial Feedback and the Problem of Epistemic Reliability

Generative artificial intelligence can produce fluent explanations, suggested revisions, worked solutions, and conversational guidance at a scale that changes the economics of feedback. The educational research problem is whether these outputs diagnose the learner's actual difficulty, guide an appropriate response, and improve independent understanding. Fluency of expression is not a validity test. A system may provide an elegant but incorrect mathematical derivation, invent a textual reference, misread a student's intent, or overcorrect a creative answer. Consequently, factual accuracy, disciplinary alignment, interpretability, and demonstrable learning must be evaluated separately.

The Ba et al. (2025) describes growing research on AI-assisted feedback alongside incomplete reporting of algorithmic procedures. The Huang et al. (2026) reports an average positive achievement effect within its eligible experimental and quasi-experimental studies but substantial heterogeneity. Neither review establishes that any individual language model can reliably infer an individual learner's cognitive state, that every style of personalized prompt is beneficial, or that AI feedback can substitute for professional oversight across disciplines. The evidence supports testable comparisons of feedback policies, not unconditional deployment claims.

Experimental designs should separate generated feedback from established explanatory feedback, distinguish the language model from its retrieval and verification tools, and measure the incidence and impact of erroneous advice. Students may treat an automated explanation as authoritative because it is available, conversational, and apparently individualized. Error detection and contestability are therefore educational outcomes in their own right. A mature system should expose uncertainty when warranted, allow students to ask for evidence or alternative reasoning, and support escalation to a qualified teacher when the instructional consequence is significant.

Teacher Learning, Organizational Conditions, and Professional Judgment

The implementation of formative assessment demands professional capacities that extend beyond writing good questions. Teachers must recognize plausible misconceptions, infer what a partial response may indicate, decide when to change pacing, choose discriminating examples, and evaluate the success of the adjustment. Professional development should therefore rehearse diagnostic interpretation and response selection using authentic student work, alongside collaborative examination of later learning. A training program that merely teaches the vocabulary of “assessment for learning” may leave these substantive practices unchanged.

Institutional constraints influence what can be enacted. Large classes, tight examination schedules, limited planning time, fragmented digital systems, and inflexible curriculum sequences can reduce the opportunity to use evidence. A classroom intervention may appear ineffective because the feedback was uninformative or because a high-quality diagnosis could not be acted upon. Evaluation designs should document these distinctions. Bennett (2011) warned against conceptual and methodological shortcuts that equate a named practice with demonstrated effectiveness; the warning applies equally to program adoption and educational policy.

A school-wide evidence culture can facilitate learning when teachers exchange examples of student reasoning, analyze common problems, and revise instructional sequences. The same infrastructure can produce distortions if formative data are used primarily to rank teachers or penalize students, thereby increasing pressure to conceal errors. Governance must specify the purpose of evidence collection and the legitimate users of the information. Formative assessment requires trust in the diagnostic process because candid demonstrations of partial understanding are part of the evidence the system depends upon.

Research Design: Identifying the Causal Effect of Feedback

The experimental evaluation of formative assessment requires a precisely specified treatment contrast. A study comparing a teacher who gives diagnostic comments and opportunities for revision with a teacher who assigns a grade estimates the effect of a package of practices, not the isolated effect of comment content. A study comparing elaborated feedback with correct-answer feedback while holding tasks, timing, and practice constant more directly estimates an informational difference. Cluster-randomized classroom trials can preserve the social organization of instruction, while item-level randomized experiments can isolate message features. Each addresses a different causal question, with its own threats to validity.

Randomization should account for assignment units, spillover, and student nesting. Pupils within a classroom share teachers, materials, peer conversations, and time, so treating their scores as independent observations can substantially understate uncertainty. A randomized trial that assigns entire schools faces fewer contamination problems but may have few independent clusters. Baseline differences, attrition, differential time-on-task, and missing assessments can further distort inference. Investigators should report participant flow, intended treatment, delivered treatment, analytic assumptions, and estimates with uncertainty intervals rather than relying exclusively on significance tests.

The central estimands can be distinguished: the effect of an offer of formative support, the effect among students who actually receive or use it, and the effect of a particular feedback component under comparable instructional conditions. These are different quantities. A per-protocol analysis selecting only engaged students may confound engagement with preexisting motivation, whereas an intention-to-treat estimate preserves the initial assignment at the expense of describing heterogeneous uptake. Mechanism claims require evidence that the hypothesized process changed, temporally preceded the learning outcome, and contributed causally beyond competing pathways.

Implementation Fidelity and the Observation of Actual Feedback Use

A protocol may specify that a teacher gives individualized formative comments, yet its actual implementation may vary in diagnostic accuracy, specificity, tone, timing, and whether students have a chance to revise. Fidelity should therefore be operationalized at several levels: whether assessment evidence was elicited, whether it was interpreted accurately, whether an instructional adjustment followed, and whether a subsequent response was observed. Counting materials distributed or comments posted records exposure, while measuring feedback use records enactment. Black & Wiliam (2009) supports treating the process as an activity embedded in pedagogy.

Observer coding can identify the content of teacher questions and responses, but coding instruments need explicit definitions, rater training, agreement estimates, and validation. A category such as “high-quality feedback” is not an adequate observable variable until it is linked to behavioral indicators and examples. Independent coders might disagree whether a teacher's statement supplies an explanatory hint or merely signals correctness. Such disagreements should be studied and resolved transparently because they affect estimated treatment dose and the interpretation of mechanisms.

Students' responses provide the complementary record. Data can include revisions, choice of subsequent problems, help-seeking, explanations generated after feedback, and later independent task solutions. Digital logs can help capture the sequence of actions, but timestamps and clicks are imperfect proxies for attention and comprehension. An intensive process study might combine trace data with targeted interviews and samples of student thinking, while making clear that introspective reports also have limitations. Together these observations distinguish a message that is issued, seen, understood, believed, and acted upon.

Psychometrics, Outcome Selection, and Inference About Learning

Formative and summative instruments often share measurement problems. A measure intended to diagnose conceptual reasoning must distinguish reasoning from reading difficulty, language proficiency, task familiarity, and test-taking strategies. Reliability can be reduced when the diagnostic task has only a few items; at the same time, repeating many items imposes an opportunity cost in instruction. Validity should be evaluated for the particular inference, not granted globally to an assessment brand. A short classroom probe can be sufficient for a low-cost reversible adjustment while being entirely inadequate for placement or high-stakes certification.

Outcome measures should be independent enough to avoid overestimating the effect of instruction that closely matches test items. Improvement on the same questions immediately after feedback may reflect item memory, correction, or new understanding. Delayed, structurally different, and independently scored assessments are important for testing durability and transfer. The distinction articulated by Soderstrom & Bjork (2015) is foundational: behaviors that facilitate performance during training can diverge from those that support long-term learning. Researchers should specify which form of achievement their instrument measures and avoid treating all positive scores as evidence of conceptual reconstruction.

Assessing complex products introduces additional challenges. Essay quality, mathematical explanation, or scientific argument can require human judgment and involve multiple dimensions. Rubrics may increase transparency but introduce rater effects and local criterion dependence. Blinded scoring, anchor examples, calibration exercises, and analyses of disagreement can improve interpretability. In studies using automatically evaluated writing or reasoning, validation against qualified independent scorers and diverse response styles is especially important. Machine-generated scores are another assessment claim requiring evidence, not an external ground truth by definition.

Mediation, Competing Mechanisms, and the Limits of Causal Stories

The fact that an intervention increases achievement does not establish the psychological process through which it works. Elaborated comments might improve comprehension, but they may also increase attention, time on task, retrieval opportunities, or perceived instructor support. An adaptive digital system may outperform a conventional condition because it grants more practice rather than because its personalization algorithm makes superior diagnoses. Theoretically meaningful mechanism tests require manipulation or credible measurement of candidate pathways, adequate temporal ordering, and consideration of unmeasured confounding.

Studies can strengthen mechanism inference through factorial contrasts. For example, explanation versus verification could be crossed with opportunities for immediate revision, or learner-generated versus instructor-generated criteria could be crossed with rubric availability. Independent measures can track conceptual understanding, confidence calibration, and revision quality at different times. A mediation analysis is only as credible as the assumptions that connect manipulated treatment, observed mediator, and outcome. The field benefits when researchers present proposed psychological mechanisms as falsifiable explanations rather than treating a program's label as proof of its active ingredient.

A useful distinction separates process measures from surrogate outcomes. A student who opens a feedback panel or writes more revision comments has engaged in an observable behavior, but this behavior may or may not improve later independent work. Similarly, improved confidence may indicate actual competence or misplaced certainty. Robust trials should report both process and achievement, examine discrepant results, and test whether the presumed mediator predicts learning under a design capable of supporting the causal interpretation. The path from information to understanding must be demonstrated rather than inferred from activity counts.

Reproducibility, Reporting Quality, and Publication Bias

The formative-assessment literature contains a persistent problem of broad conceptual labels attached to noncomparable interventions. Kingston & Nash (2011) found that a small fraction of candidate K–12 studies could support interpretable effect calculations, and Bennett (2011) questioned whether some high-profile interpretations exceeded the available evidence. These findings call for explicit treatment protocols, preregistered hypotheses where appropriate, accessible materials, and outcome reporting that separates primary measures from exploratory comparisons.

Replication should preserve the functional elements of an intervention while documenting differences in setting. A feedback approach may be replicated with new surface examples but equivalent diagnostic questions and comparable opportunities for revision. If the instructional context changes, the replication should document teacher expertise, curricular content, student characteristics, and institutional constraints so that variation can inform theory. A failed replication can reveal that an originally reported result depended on a particular treatment package or that the effect is sensitive to conditions. These possibilities are scientifically informative when investigated transparently.

Meta-analyses should assess study selection, selective reporting, multiple dependent effects, and between-study heterogeneity. Wisniewski et al. (2020) demonstrate that substantial dispersion remains even in a large evidence base. Average effects should be reported alongside the number of studies, contrast definitions, confidence intervals, and plausible moderators. Researchers should attend to gray literature and protocols where available and consider the possibility that the published literature overrepresents favorable findings. Sensitivity analyses can quantify how estimates vary under alternative assumptions but cannot automatically eliminate unmeasured bias.

Developmental, Linguistic, and Cultural Variability

The interpretation of feedback develops with learners' language, executive functions, prior experience, and familiarity with disciplinary quality criteria. Young children may require concrete demonstrations and opportunities to act, whereas advanced students may benefit from abstract criticism or self-generated comparisons. The same learner can require different information as expertise grows. A developmental theory therefore predicts interactions between feedback form, prior knowledge, and the capacity to identify the next action, rather than assuming age is a simple linear moderator.

Language accessibility presents an especially consequential measurement problem. A teacher may mistake difficulty interpreting task instructions or a feedback comment for a deficit in disciplinary knowledge. In multilingual classrooms, conceptual reasoning can be expressed through multiple languages, drawings, mathematical notation, or oral explanation. A formative procedure that relies on an unnecessary linguistic barrier may produce systematically misleading evidence. Researchers should distinguish linguistic demands intrinsic to the learning target from those introduced by the assessment format and document accessibility accommodations when interpreting outcomes.

Culture and institutional experience can also affect how critique, praise, public disagreement, and peer feedback are understood. A message that appears supportive in one classroom may carry a different social meaning elsewhere. Broad claims require studies across settings and populations rather than simplistic cultural stereotypes. Researchers should measure these differences directly through interpretation, uptake, and learning, and should test whether the proposed mechanism transports to new contexts. In this sense, diversity is a condition for explaining the phenomenon, not merely an item on a reporting checklist.

An Integrative Causal Architecture of Formative Learning

The evidence supports modeling formative assessment as a chain of conditional transformations: an instructional goal defines what should be learned; an elicitation task produces an observable response; an inferential process maps the response to a provisional judgment about competence; a feedback event or teaching decision changes the learning environment; a learner processes the information and acts; and an independent later task tests the resulting competence. The chain is a synthesis of established research problems, not a newly validated psychometric scale. Its value lies in locating potential failure points that are otherwise concealed by broad intervention labels. A feedback program can fail because evidence was weak, the diagnosis mistaken, the message incomprehensible, the opportunity for revision absent, or the eventual outcome inadequately measured.

The components are dynamically related. Assessment itself can induce retrieval practice; a teacher's diagnosis can be revised through dialogue; students' confidence and identity beliefs can alter the interpretation of corrective information; peer interaction can reshape the evidence presented to the teacher; and new performance can feed into a further learning cycle. Black & Wiliam (2009), Sadler (1989), and Nicol & Macfarlane-Dick (2006) provide distinct theoretical foundations for these links. Their differences are productive: classroom formative theory emphasizes contingent instruction, Sadler emphasizes quality judgment, and self-regulation theory emphasizes the learner's capacity to monitor and act.

A robust synthesis should resist the inference that the strength of any individual link guarantees an outcome. Precisely diagnosed misconceptions can persist when the subsequent instruction is ineffective; useful explanations can fail to transfer beyond practiced examples; motivationally agreeable feedback can leave understanding unchanged. Evidence accumulated across the chain is more informative than a single global satisfaction measure. For experimental purposes, one can isolate particular links while retaining the sequence as a model of what successful practice must accomplish.

Evidence-Based Design Principles and Their Conditions

Several practical principles follow from the literature when treated as conditional propositions. First, articulate the target competence and select a task that can elicit interpretable evidence about it. Second, use the resulting response to form a tentative instructional diagnosis that can be refined. Third, provide information or redesign instruction at a level the learner can interpret and act upon. Fourth, create a genuine opportunity for subsequent work, and assess the outcome with a task capable of distinguishing learning from repetition. These principles describe an evidentiary sequence derived from the research, not a universal scripted procedure.

The information content of feedback matters. van der Kleij et al. (2015) found substantial differences between verification-only, correct-answer, and elaborated feedback across their computer-based evidence set. Hattie & Timperley (2007) distinguish task, process, and self-regulation information, while Kluger & DeNisi (1996) identify attention to task and self-related processes as a moderator of feedback effects. The resulting design recommendation is to ask which information will support the next useful action for a particular learner and task, and then test that proposition rather than accumulating comments as an end in itself.

Implementability is part of scientific validity. Diagnostic tasks must fit the time available, instructor knowledge, student access, and the curriculum's opportunities for revision. Digital systems may alter costs and provide individual feedback, but their outputs should be verified and integrated with legitimate instructional decisions. Teacher professional learning, rubric calibration, peer-feedback training, and assessment design remain researchable components. A well-specified intervention states which of these components are present, how they operate, and which outcomes they are expected to influence.

Persistent Contradictions and Unresolved Research Questions

The central evidence gap concerns mechanisms. Many studies compare a named feedback treatment with a comparison condition but cannot establish whether benefits arise from explanation, motivation, extra practice, altered teacher attention, or subsequent revision. Future studies should manipulate components independently, measure theoretically relevant mediators, and report delayed independent outcomes. The field also needs transparent estimates of harmful or null feedback effects, especially for vulnerable learners, complex tasks, and poorly calibrated automated systems. Wisniewski et al. (2020) demonstrates that positive mean effects coexist with substantial variability.

A second unresolved issue is the scale of generalization. Research conducted on specific questions, courses, or disciplinary tasks cannot automatically justify claims about lifelong metacognition or broad academic success. Follow-up studies should test whether a capacity for judgment developed in one setting travels to new problems and remains evident when support is withdrawn. This is especially important for peer assessment, automated tutoring, and interventions designed to increase self-regulation. Outcome designs should separate product improvement, strategy acquisition, conceptual understanding, far transfer, and independent evaluative judgment.

A third issue is the unit of intervention. Formative assessment operates in individual cognition, classroom dialogue, teacher professional decision making, and institutionally governed information systems. A test of message content does not directly establish a school improvement strategy, and a school-level reform trial cannot uniquely identify the best sentence a teacher should write on an essay. Theoretical integration requires related studies across levels with explicit causal links. Bennett (2011) and Kingston & Nash (2011) show why clear definitions and rigorous evidence remain essential.

A fourth issue concerns artificial feedback quality and accountability. Adaptive systems may improve access to explanatory information, but the field lacks sufficiently broad evidence about long-term dependence, independent reasoning, undetected errors, fairness across linguistic and cultural populations, and the consequences of replacing human judgment with algorithmic scores. Huang et al. (2026) identify positive achievement effects in an emerging literature while acknowledging high heterogeneity and weak evidence for some noncognitive and metacognitive outcomes. Longitudinal, preregistered, multicenter trials with instructor-supervised comparison groups would clarify which configurations support durable learning.

Formative Assessment in the Artificial Era

The Artificial Era extends the educational environment beyond exclusively human interactions. Artificial systems can participate in the production of questions, interpretation of responses, generation of explanations, and selection of subsequent tasks. The scientific object remains a relation among observable evidence, inferential judgments, instructional changes, and independently demonstrated learning. Whether a system is experienced as a teacher, tutor, collaborator, or tool influences trust and feedback uptake, but the legitimacy of its instructional claim depends on verifiable performance and transparent methods. Terminological questions about Artificial Cognition and Artificial Agency should be operationalized through competence and behavior rather than used to imply empirically established internal states.

Human–artificial learning relations create new experimental contrasts. Researchers can compare human-only feedback, artificial-only feedback, and joint configurations while holding response time, practice opportunities, and task difficulty constant. They can examine whether an instructor's verification changes student trust, whether students become more able to detect generated errors, and whether independent performance deteriorates when automated assistance is withdrawn. Feedback design could also be evaluated for its effects on the learner's capacity to question the system. These are empirical hypotheses, and the corresponding educational processes must be measured rather than assumed from the technological form of the interaction.

The institutional challenge is to preserve the epistemic chain from observation to action. An automatically generated comment should be traceable to the student's actual performance and the relevant curricular standard; the explanation should be checkable; and the learner should have a means to challenge or improve it. Automated errors and disparities should be recorded as part of system evaluation. This framework allows artificial participants to contribute useful informational operations while placing the validity of each operation under scientific scrutiny. It also positions formative assessment as a significant area for investigating evolving relationships between Homo and Artificial in education.

Conclusion

Formative assessment and instructional feedback constitute a recurring evidence-guided process through which teachers and learners identify aspects of understanding, interpret information, choose responses, and investigate the consequences for subsequent learning. Their scientific significance arises from the connection between measurement and pedagogical action. The same quiz, comment, rubric, or algorithm can have different effects depending on the construct assessed, the interpretation offered, the opportunity for revision, and the learner's capacity to respond. A defensible definition therefore identifies the use of evidence within a learning cycle rather than equating assessment with a particular instrument.

The historical literature provides a coherent sequence of contributions. Early performance-feedback and evaluative-format experiments revealed the importance of attention and motivation; Sadler's theory established the centrality of standards and evaluative judgment; classroom formative-assessment scholarship examined the contingency of instruction; models of self-regulation and feedback literacy emphasized the recipient's active use of information. Modern meta-analyses establish that feedback can benefit learning and that its effects are highly heterogeneous. These findings strengthen the field when expressed as conditional causal claims, rather than general promises attached to the words “formative” or “feedback.”

Future progress depends on sharper diagnosis, independent learning measures, component-level experiments, long-term follow-up, and transparent reporting of implementation and harms. Artificial systems expand the possible scale and form of feedback, making these requirements more pressing. The enduring research question is how reliable evidence about partial understanding can be transformed into the particular actions that produce durable, transferable, and increasingly self-regulated competence.

References

  • Ba, S., Yang, L., Yan, Z., Looi, C. K., & Gašević, D. (2025). Unraveling the mechanisms and effectiveness of AI-assisted feedback in education: A systematic literature review. Computers and Education Open, 9, Article 100284. https://doi.org/10.1016/j.caeo.2025.100284
  • Bangert-Drowns, R. L., Kulik, C.-L. C., Kulik, J. A., & Morgan, M. T. (1991). The instructional effect of feedback in test-like events. Review of Educational Research, 61(2), 213–238. https://doi.org/10.3102/00346543061002213
  • Bennett, R. E. (2011). Formative assessment: A critical review. Assessment in Education: Principles, Policy & Practice, 18(1), 5–25. https://doi.org/10.1080/0969594X.2010.513678
  • Black, P., & Wiliam, D. (1998). Assessment and classroom learning. Assessment in Education: Principles, Policy & Practice, 5(1), 7–74. https://doi.org/10.1080/0969595980050102
  • Black, P., & Wiliam, D. (2009). Developing the theory of formative assessment. Educational Assessment, Evaluation and Accountability, 21(1), 5–31. https://doi.org/10.1007/s11092-008-9068-5
  • Butler, R. (1988). Enhancing and undermining intrinsic motivation: The effects of task-involving and ego-involving evaluation on interest and performance. British Journal of Educational Psychology, 58(1), 1–14. https://doi.org/10.1111/j.2044-8279.1988.tb00874.x
  • Carless, D., & Boud, D. (2018). The development of student feedback literacy: Enabling uptake of feedback. Assessment & Evaluation in Higher Education, 43(8), 1315–1325. https://doi.org/10.1080/02602938.2018.1463354
  • Dawson, P., Henderson, M., Mahoney, P., Phillips, M., Ryan, T., Boud, D., & Molloy, E. (2019). What makes for effective feedback: Staff and student perspectives. Assessment & Evaluation in Higher Education, 44(1), 25–36. https://doi.org/10.1080/02602938.2018.1467877
  • Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
  • Furtak, E. M., & Ruiz-Primo, M. A. (2008). Making students' thinking explicit in writing and discussion: An analysis of formative assessment prompts. Science Education, 92(5), 799–824. https://doi.org/10.1002/sce.20270
  • Hattie, J., & Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81–112. https://doi.org/10.3102/003465430298487
  • Huang, Y., Chen, S., Zhang, W., & Chen, M. (2026). Can generative AI feedback effectively enhance learning outcomes? A meta-analysis of 36 experimental and quasi-experimental studies. Education Sciences, 16(6), Article 816. https://doi.org/10.3390/educsci16060816
  • Kingston, N., & Nash, B. (2011). Formative assessment: A meta-analysis and a call for research. Educational Measurement: Issues and Practice, 30(4), 28–37. https://doi.org/10.1111/j.1745-3992.2011.00220.x
  • Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284. https://doi.org/10.1037/0033-2909.119.2.254
  • Lipnevich, A. A., & Panadero, E. (2021). A review of feedback models and theories: Descriptions, definitions, and conclusions. Frontiers in Education, 6, Article 720195. https://doi.org/10.3389/feduc.2021.720195
  • Nicol, D. J., & Macfarlane-Dick, D. (2006). Formative assessment and self-regulated learning: A model and seven principles of good feedback practice. Studies in Higher Education, 31(2), 199–218. https://doi.org/10.1080/03075070600572090
  • Panadero, E., & Jonsson, A. (2013). The use of scoring rubrics for formative assessment purposes revisited: A review. Educational Research Review, 9, 129–144. https://doi.org/10.1016/j.edurev.2013.01.002
  • Panadero, E., & Lipnevich, A. A. (2022). A review of feedback models and typologies: Towards an integrative model of feedback elements. Educational Research Review, 35, Article 100416. https://doi.org/10.1016/j.edurev.2021.100416
  • Ruiz-Primo, M. A., & Furtak, E. M. (2007). Exploring teachers' informal formative assessment practices and students' understanding in the context of scientific inquiry. Journal of Research in Science Teaching, 44(1), 57–84. https://doi.org/10.1002/tea.20163
  • Sadler, D. R. (1989). Formative assessment and the design of instructional systems. Instructional Science, 18(2), 119–144. https://doi.org/10.1007/BF00117714
  • Shute, V. J. (2008). Focus on formative feedback. Review of Educational Research, 78(1), 153–189. https://doi.org/10.3102/0034654307313795
  • Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176–199. https://doi.org/10.1177/1745691615569000
  • Tunstall, P., & Gipps, C. (1996). 'How does your teacher help you to make your work better?' Children's understanding of formative assessment. The Curriculum Journal, 7(2), 185–203. https://doi.org/10.1080/0958517960070205
  • van der Kleij, F. M., Feskens, R. C. W., & Eggen, T. J. H. M. (2015). Effects of feedback in a computer-based learning environment on students' learning outcomes: A meta-analysis. Review of Educational Research, 85(4), 475–511. https://doi.org/10.3102/0034654314564881
  • Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, Article 3087. https://doi.org/10.3389/fpsyg.2019.03087