Marked Pedagogies: Bias and Algorithmic Feedback

• Marked pedagogies are instructional approaches that explicitly highlight learner identities and characteristics, using algorithmic and human-mediated cues to customize feedback.
• They integrate insights from sociolinguistics, educational justice, and computational pedagogy to reveal and address systemic biases in teaching and assessment.
• Empirical studies using large language models demonstrate that marked pedagogies can both reinforce stereotypes and promote more inclusive, transparent curricular strategies.

Marked pedagogies refer to instructional designs and algorithmic feedback practices that systematically foreground or respond differentially to learner attributes—such as race, gender, linguistic background, or presumed achievement—shaping educational experiences in ways that depart from a neutral or “unmarked” baseline. The term synthesizes insights from sociolinguistics, educational justice theory, and computational pedagogy, designating both observable surface features of feedback (lexical, structural, and tonal) and deeper “instructional orientations” that can reinforce or disrupt systemic biases. Marked pedagogies are instantiated in both human- and AI-mediated settings, with contemporary research identifying particular risks and affordances in the era of LLMs and AI-augmented courseware (Tan et al., 12 Mar 2026, Mayr, 7 Apr 2026, Ardila-Mantilla, 2020).

1. Theoretical Foundations and Definitions

Marked pedagogies extend the classical sociolinguistic notion of markedness—in which “unmarked” categories (e.g., White, male, monolingual) are constructed as normative and “marked” categories (e.g., Black, female, English language learners) as departures from this norm—into the domain of instructional action. Markedness historically referred to how linguistic expressions encode difference; in marked pedagogies, the principle is applied to feedback, assessment, and curricular structuring. Tan et al. define marked pedagogies as the systematic ways in which LLMs and algorithmic feedback tools adjust their focus, evaluative stance, and address based on discrete cues about presumed student identity, achievement, or disposition (Tan et al., 12 Mar 2026). In parallel, Mayr operationalizes “knowledge markers”—a coarse-grained tagging of course content by knowledge type (Application, Structure, Procedure)—as a design-level instantiation of marked pedagogical structuring, making the affordances of educational artifacts visible and actionable (Mayr, 7 Apr 2026).

2. Empirical Investigation in Automated Feedback

Tan et al. conducted a large-scale empirical study examining the presence of marked pedagogies in LLM-generated feedback. Using 600 standardized essays from the PERSUADE corpus, the authors manipulated only the minimal student attribute in prompts (e.g., English language learner, Black student, low-achieving), leaving the essay content unchanged. Four LLMs (GPT-4o, GPT-3.5-turbo, Llama-3.3-70B, Llama-3.1-8B) were tested in a zero-shot instruction setting. Feedback was generated for each essay under baseline, marked, and comparative identity conditions, allowing for isolation of lexical and structural shifts attributable solely to these cues (Tan et al., 12 Mar 2026).

Analytically, the study leveraged an adapted Marked Words framework, employing log-odds ratios and Dirichlet priors to determine which words were significantly distinctive of feedback given under marked versus comparative conditions. The aggregation metric $C_s(F)$, quantifying the percentage of feedback tokens corresponding to the top-20 distinctive words for each attribute set, was regressed on prompt condition, controlling for within-essay covariates.

The findings indicated:

• Systematic Stereotype-Alignment: LLMs provided less substantive critique, more praise, and often tacitly lowered expectations for marked categories (e.g., feedback for ELL or Black students pivoted to language adequacy, while White students received more abstract or structural guidance).
• Linguistic Modulation: Distinctive address and affect—e.g., first-person empathy for female students (“I love your confidence”), upbeat hedges for presumed unmotivated students, terse corrections for low-achievement markers.
• Statistical Significance: Condition effects were robust ($p<0.001$). For instance, ELL prompts yielded a +3.829 percentage point increase over baseline in distinctive feedback, and Black identity markers produced a +180% increase in identity-relevant terms.

These patterns persisted across open and closed source models, with even purely “comparative” unmarked prompts (e.g., specifying the student as White) producing discernible and non-negligible lexicon shifts.

3. Markedness in Human-Designed Pedagogy and Community Practices

Beyond the computational domain, marked pedagogies have been advanced as intentional design strategies in human-facilitated classrooms. Ardila-Mantilla’s account of “Todxs cuentan” frames the first day of a university mathematics class as a sequence of explicit marking protocols: centering music and culture, building community agreements, and welcoming full student humanity. Drawing on Lorde’s imperative to treat difference as a resource, hooks’s pedagogy of communal presence, and Gutiérrez’s call for “rehumanization,” this approach enacts protocols that operationalize student input, co-authorship, and reflective adaptation (Ardila-Mantilla, 2020). Each classroom activity—musical introductions, group agreements, transparent assessment dialogue—is designed to mark difference as generative, not deficit, and to flatten hierarchical instructor-student relations.

The process is algorithmic in nature: For $t=1...6$ (spanning environment, introductions, cultural engagement, agreement, assessment, and community-building notecards), the protocol iterates enactment, solicitation of feedback, and visible integration of student input.

4. Knowledge Markers and Structural Markedness in Course Design

Marked pedagogies can manifest structurally in curriculum and artifact design. Mayr introduces “knowledge markers” (A/S/P) as lightweight, tool-agnostic labels for instructional units: Application, Structure, and Procedure (Mayr, 7 Apr 2026). Each marker is mapped to Bloom’s taxonomy and Krathwohl’s revision, rendering explicit the intended cognitive emphasis of each instructional segment. Table extraction from Mayr demonstrates this mapping:

Marker	Emphasis	Taxonomic Mapping
(A)	Implementation	Remember, Apply, Create (factual/procedural)
(S)	Mental models	Understand, Analyze (conceptual)
(P)	Systematic methods	Apply, Evaluate, Create (procedural/metacognitive)

This structural markedness allows instructors and students to visualize and balance the distribution of cognitive demands, operationalize interleaving, and prevent the occlusion of critical knowledge types (e.g., overemphasis on application at the expense of conceptual or evaluative scaffolding). In AI-permissive environments—where generative tools can produce plausible artifacts without human understanding—these markers make learning intent transparent and counteract superficial learning.

5. Analytical Frameworks and Quantification

Quantification of marked pedagogies leverages both corpus-level linguistic indicators and structural cartography. The Marked Words framework utilizes log-odds ratio with an informative Dirichlet prior to stabilize low-frequency estimates. For corpus vocabularies $P$ (marked) and $Q$ (comparative), relative entropy ($D_{KL}(P||Q)$) captures overall distributional divergence.

Concentration metrics—such as $C_s(F)$—express the degree to which feedback depends on attribute-distinctive lexicon. These are suitable for regression modeling with fixed effects to isolate independent pedagogical variables. In structural analysis, distributional ratios of A/S/P markers across course units, sections, and time can reveal implicit value systems embedded in curricular design, providing a formal mechanism to inspect (and remediate) unintended imbalances.

6. Implications, Limitations, and Critical Perspectives

Marked pedagogies, whether algorithmically induced or consciously engineered, pose risks and affordances for educational equity. Empirical evidence demonstrates that LLM-based personalization, absent critical accountability, systematically amplifies stereotypes and withholds substantive feedback from marginalized identities (Tan et al., 12 Mar 2026). Even overtly inclusive pedagogies risk essentializing difference or setting up hierarchical “inclusion” as solely the work of the dominant culture unless student agency is continually foregrounded (Ardila-Mantilla, 2020).

Methodological limitations in structural marker schemes include subjectivity in unit-level annotation and lack of empirical outcome data (learning gains, long-term retention) (Mayr, 7 Apr 2026). Community calibration and periodic re-inspection of markedness are necessary.

The field remains open to further empirical studies quantifying student navigation, verifying impacts of intervention, and developing semi-automated support for large-scale marker assignment. A plausible implication is that without deliberate transparency and iterative adaptation, both algorithmic and human pedagogical systems risk scaling and entrenching the very inequities they claim to resolve.