Critical AI Engagement Framework, Versions 1.0 and 1.4 (scroll down…)
NB: This page was original published 24 March 2026, and has since been updated on 21 May 2026 to reflect the latest framework iteration.
This framework draws on Hosseini's (2026) forthcoming paper in The Geographical Journal, which examines how generative AI tools reproduce racial, gendered, and class-based representations through algorithmic coloniality (Mohamed et al., 2020). That work demonstrates how seemingly neutral prompts encode dominant cultural assumptions, producing outputs that reflect and reinforce existing inequalities.
The Critical AI Engagement Framework extends this analysis into a practical pedagogical tool, mapping how individuals engage with generative AI across two axes: epistemic posture and structural consciousness. It is designed for use across educational contexts, supporting educators and learners in moving beyond prompt refinement toward critical, collective, and structurally informed engagement with AI systems.
April 2026 update: the paper informing this framework has now been published and is available below
Hosseini, D. D. 2026. “Generative AI: A Problematic Illustration of the Intersections of Race, Gender and Class.” The Geographical Journal 192, no. 2: e70085. https://doi.org/10.1111/geoj.70085.
Critical AI Engagement Framework
Grounded in Mohamed et al. (2020), Benjamin (2019), Noble (2018), Zembylas (2023), Crenshaw (1991)
Generative AI tools are now embedded in higher education, used to draft text, generate images, produce code, and synthesise research. Yet these tools are not neutral. They encode the assumptions, hierarchies, and exclusions of the data they were trained on, reproducing patterns of racial, gendered, and class-based harm even as their outputs become more sophisticated (Benjamin, 2019; Mohamed et al., 2020; Noble, 2018). Understanding how students engage with these tools, and what that engagement does or does not interrogate, is therefore an urgent task for educators.
The Critical AI Engagement Framework maps educational engagement with generative AI across two axes. The horizontal axis describes an individual’s epistemic posture toward AI: from treating its outputs as authoritative, through questioning them, to recognising the colonial and structural conditions that shape what AI knows, whose knowledge it centres, and what it silences (Hosseini, 2026; Maalsen, 2023). The vertical axis describes how far an individual recognises AI as a socio-technical artefact shaped by relations of power, rather than simply a tool with fixable flaws (Benjamin, 2019; Quijano, 2000; Zembylas, 2023). The positions individuals occupy are not freely chosen but reflect the institutional, curricular, and social conditions that shape how they learn and work. The framework’s aspiration is not a more capable individual user but collective action: sustained, community-grounded engagement that works toward structural change (Camacho Felix, 2025; Mohamed et al., 2020).
Epistemic deference
AI as neutral oracle
Critical interrogation
Outputs questioned
Epistemic agency
Whose knowledge?
Collective / relational
Critique with others
Individualized
engagement
AI as personal tool,
neutral & apolitical
neutral & apolitical
The Uncritical Receiver
Accepts outputs; AI naturalized as neutral knowledge source
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): student treats AI outputs as objective, without recognizing that GenAI encodes "opaque, inconsistent cultural assumptions" shaped by historically racist and sexist training data (Keshishi & Hosseini, 2023; Benjamin, 2019).
For educators
Make the assumption of neutrality visible through practical experiments. Ask: whose perspective does a GenAI output reflect by default — in an image, a summary, a clinical recommendation, or a generated essay? Drawing on Day & Esson (2025) and Hosseini (2026), show how seemingly neutral prompts reproduce skewed cultural defaults across output types.
For students
The student experiences AI outputs as "natural" rather than constructed. As Benjamin (2019) argues, socio-technical artefacts are not static reflections — they are shaped by the feedback and values of those who built them. Students need a framework to see this, not just permission to question.
The Cautious Pragmatist
Checks outputs; AI still framed as neutral instrument
Theoretical anchor
The student audits outputs for factual errors but not for the cultural assumptions encoded in them. As Day & Esson (2025) show, even anomalous outputs require users to "remain vigilant to the opaque and shifting nature of generative AI tools" — vigilance the Cautious Pragmatist applies technically but not epistemically.
For educators
Shift from "is this accurate?" to "whose accuracy?" Introduce temporal dynamism (Kleinman, 2024): improved outputs — whether images, text, or code — do not mean underlying biases have been addressed. Use Spennemann & Oddone's (2025) technique of asking GenAI to explain its own outputs as a critical exercise.
For students
May believe that better prompting solves the problem. However, prompt refinement "would not address the underlying biases within the datasets themselves" (Hosseini, 2026). The student needs to move from refining inputs to interrogating the training data and the colonial logics embedded within it.
The Epistemically Alert
Interrogates whose knowledge is centered; notices silences
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): student recognizes that AI systems embed a "dominant, Eurocentric worldview" that upholds hierarchical, racialized, and gendered ways of knowing. Connects to Noble's (2018) algorithms of oppression and Maalsen's (2023) algorithmic epistemologies.
For educators
Move from naming bias to interrogating its origin. Use Quijano's (2000) coloniality of power to show that AI's racial and gender defaults are not errors but expressions of colonial hierarchies embedded in training data. Ask: what would an AI trained on non-Eurocentric datasets produce differently?
For students
May feel isolated, especially when institutional AI guidance frames the issue as a technical problem. Wilby & Esson's (2024) call for "capabilities, caveats, and criticality" provides legitimizing language. Connect to communities of practice doing this work.
The Isolated Disruptor
Critiques AI alone; change without solidarity
Theoretical anchor
Individual critique of algorithmic coloniality, however sophisticated, cannot address structural problems in proprietary and inaccessible training datasets (Amoore et al., 2024). Mohamed et al. (2020) are explicit: structural change requires "political coalitions and communities," not individual actors.
For educators
Connect students to collective and cross-disciplinary action. Addressing algorithmic coloniality requires breaking down "disciplinary and departmental silos" (Hosseini, 2026; Maalsen, 2023). Individual insight without structural leverage changes nothing about the datasets or systems producing harmful outputs.
For students
Risk of cynicism or disengagement when individual critique runs up against inaccessible, proprietary datasets and opaque systems. As Hosseini (2026) demonstrates, surface improvements in GenAI outputs can mask rather than resolve the underlying colonial logics — students need community and strategy, not just analysis.
Partial structural
awareness
Senses bias or harm,
lacks systemic account
lacks systemic account
The Uneasy Adapter
Senses something wrong; lacks language to name it
Theoretical anchor
Pre-conceptual awareness of algorithmic harm: student senses that something is "off" in AI outputs — perhaps noticing racial or gender skew — but has not yet encountered the theoretical vocabulary to name it. This is the moment described by Day & Esson (2025) when outputs produce "surprising results."
For educators
This is a threshold moment. Offer concepts — algorithmic coloniality (Mohamed et al., 2020), algorithms of oppression (Noble, 2018), socio-technical artefacts (Benjamin, 2019) — as language for what is already felt. Hosseini's (2026) method of prompting GenAI and critically analyzing outputs is a replicable pedagogical entry point adaptable across text, image, and code generation.
For students
High potential. Already doing affective critical work. Avoid rushing to resolution — the unease is epistemically productive. GenAI outputs should be approached "not [as] surprising, but as symptomatic of racialised and gendered logics" (Hosseini, 2026) embedded in training data across all output modalities.
The Informed Skeptic
Identifies bias in outputs; most common profile
Theoretical anchor
Can identify racial and gender skew in outputs — consistent with quantitative evidence (Cheong et al., 2024; Currie et al., 2024, 2025) — but frames it as a dataset problem rather than an expression of algorithmic coloniality (Mohamed et al., 2020). The systemic account is absent.
For educators
Move from "bias as glitch" to "bias as design." Use Benjamin's (2019, p. 59) argument that training datasets carry "the prejudices of the individuals who compiled them." Ask: why does a GenAI default encode particular assumptions about race, class, gender, or expertise — whether producing an image, drafting a clinical summary, or generating a curriculum resource?
For students
May believe that surface improvements — more realistic outputs, more diverse teams, better prompts — will resolve the issue. Hosseini (2026) demonstrates directly that successive GenAI model versions produced aesthetically improved outputs while reproducing the same racial and gendered logic. The technical fix does not address colonial logics in the training data.
The Structural Analyst
Names AI harms systemically; connects to power
Theoretical anchor
Understands AI as a socio-technical artefact (Benjamin, 2019) shaped by Silicon Valley's role as "part of the United States, a global hegemon and a successor to European colonial powers" (Keshishi & Hosseini, 2023). Connects algorithmic coloniality (Mohamed et al., 2020) to concrete outputs.
For educators
Deepen from analysis to action. Introduce reparative description (Parry, 2023): how might geographers work with public image repositories to revise false past categorizations? Introduce Zembylas's (2023) strategies for "undoing the ethics of digital neocolonialism."
For students
May become frustrated that structural analysis does not translate into change. Channel into cross-disciplinary collaboration. Addressing problematic training data requires collective action and "relational approaches that emphasise the spatial and political contexts of algorithms" (Maalsen, 2023; Hosseini, 2026). Analysis without community and outlet risks paralysis.
The Emerging Ally
Seeks solidarity; building shared critical vocabulary
Theoretical anchor
Transitional position between individual and collective consciousness (Freire). Recognises that critique must be collective but lacks the structural analysis to ground it yet.
For educators
Facilitate cross-disciplinary collaboration explicitly. Addressing algorithmic harm requires breaking "disciplinary and departmental silos" (Hosseini, 2026; Maalsen, 2023) — across education, geography, data science, and activism. Connect emerging allies to existing coalitions and communities of practice doing this work.
For students
Motivated by justice but may lack the analytical vocabulary to sustain critique under institutional pressure. Pairing with theoretically grounded peers — including those with lived experience of the harms being analyzed (cf. acknowledgments in Keshishi & Hosseini, 2023) — is more generative than educator-only support.
Structural
consciousness
AI as site of
coloniality & harm
coloniality & harm
Conscientized but Constrained
Sees the system; defers under institutional pressure
Theoretical anchor
Understands algorithmic coloniality and its harms but operates in institutional systems — curriculum, assessment, professional bodies — that have not caught up with the critique. Within many national contexts "there are nascent discussions on the ethical issues of using Gen AI technologies within tertiary education" (Hosseini, 2026) — the institutional conversation is beginning but remains uneven.
For educators
Name the institutional lag explicitly. Developing "algorithmic literacy as part of wider digital literacy initiatives" (Kong et al., 2023; Zembylas, 2023; Hosseini, 2026) is a growing expectation — the conversation is beginning, and students can actively contribute to shaping it rather than waiting for institutions to catch up.
For students
Risk of internalizing structural constraint as personal inadequacy. The student's tension is not a sign of failure — it is evidence of structural contradictions that institutions have not yet resolved. Validate the critique while building pathways to act within and against institutional constraints.
The Critical Refuser
Refuses metaphorical framing; acts on structural critique
Theoretical anchor
Tuck & Yang (2012): decolonization is not a metaphor. Student refuses cosmetic diversity framings and demands structural change to what AI produces and whom it serves.
For educators
Support with Mohamed et al.'s (2020) practical recommendations: identifying sites of coloniality in AI systems, understanding where and how algorithms are made, engaging in reparative description (Parry, 2023), and developing local and national policy challenges to colonial algorithmic logics.
For students
May encounter resistance from colleagues who frame AI critique as technophobia or obstructionism. Documentation and publication — as Hosseini (2026) demonstrates — transforms resistant practice into sharable pedagogical resource. Connect to communities doing this work across disciplines; the argument gains force collectively.
The Critical Collaborator
Challenges AI's epistemic order; builds alternatives
Theoretical anchor
Actively participates in co-creating "instructional materials that transcend boundaries" (Hosseini, 2026) — resources that make algorithmic coloniality visible and addressable across GenAI modalities. Draws on intersectionality (Crenshaw, 1991; Hill Collins, 2019) to hold race, gender, and class in simultaneous analysis rather than treating each as a separate problem.
For educators
Commission rather than assess. Meaningful critique of algorithmic coloniality requires centering those with lived expertise in the harms being analyzed — not as informants but as co-authors (Hosseini, 2026). This student's contribution should shape pedagogy, not merely illustrate it. Invite co-authorship, co-design, and co-delivery.
For students
Risk of co-option — being absorbed as institutional evidence of diversity without structural change. Hosseini's (2026) reflexive positioning — centering colleagues with lived expertise in racial and gender inequity — models how genuine co-production differs from performative consultation. Support students to name and resist this distinction.
The Praxis Collective aspirational*
Reflection + action with others; pluriversal praxis
Theoretical anchor
Camacho Felix's (2025) decolonial imaginations and collective imagination — "unveiling different possibilities for addressing injustices" through relational, mutual aid. Mohamed et al.'s (2020) political coalitions. Benjamin's (2019) abolitionist tools for dismantling the New Jim Code in AI systems.
For educators
Collective praxis around GenAI requires institutional conditions: time, resource, partnership, and willingness to redistribute epistemic authority. It demands cross-disciplinary collaboration, reparative dataset work (Parry, 2023), and policy advocacy (Hosseini, 2026; Mohamed et al., 2020) — none of which individual pedagogy alone can produce. Educators must build the structures, not just model the position.
For students
Students here are co-researchers and co-educators. Hosseini (2026) models this directly: conducting experiments, publishing findings, and encouraging readers to replicate and extend the work with a critical eye. Sustain rather than assess — the goal is ongoing collective action that outlasts the course, not a demonstration of competence for a grade.
← epistemic deference
collective / relational agency →
Movement across these axes is non-linear — students may hold multiple positions simultaneously across different contexts and knowledge domains
Theoretical grounding
Horizontal axis: Mohamed et al. (2020) — algorithmic coloniality; Noble (2018) — algorithms of oppression; Maalsen (2023) — algorithmic epistemologies and situated knowledge ·
Vertical axis: Benjamin (2019) — socio-technical artefacts encoding racial inequity; Zembylas (2023) — decolonial AI in HE; Quijano (2000) — coloniality of power; Camacho Felix (2025) — decolonial imaginations and collective action
Critical AI Engagement Framework Version 1.4
This version adds in the researcher lens and concepts such as epistemic accountability and the research lifecycle.
Critical AI Engagement Framework
Grounded in Mohamed et al. (2020), Benjamin (2019), Noble (2018), Zembylas (2023), Crenshaw (1991)
Generative AI tools are now embedded in higher education, used to draft text, generate images, produce code, and synthesise research. Yet these tools are not neutral. They encode the assumptions, hierarchies, and exclusions of the data they were trained on, reproducing patterns of racial, gendered, and class-based harm even as their outputs become more sophisticated (Benjamin, 2019; Mohamed et al., 2020; Noble, 2018). Understanding how students engage with these tools, and what that engagement does or does not interrogate, is therefore an urgent task for educators.
The Critical AI Engagement Framework maps educational engagement with generative AI across two axes. The horizontal axis describes an individual’s epistemic posture toward AI: from treating its outputs as authoritative, through questioning them, to recognizing the colonial and structural conditions that shape what AI knows, whose knowledge it centers, and what it silences (Hosseini, 2026; Maalsen, 2023). The vertical axis describes how far an individual recognizes AI as a socio-technical artefact shaped by relations of power, rather than simply a tool with fixable flaws (Benjamin, 2019; Quijano, 2000; Zembylas, 2023). The positions individuals occupy are not freely chosen but reflect the institutional, curricular, and social conditions that shape how they learn and work. The framework’s aspiration is not a more capable individual user but collective action: sustained, community-grounded engagement that works toward structural change (Camacho Felix, 2025; Mohamed et al., 2020).
Epistemic deference
AI as neutral oracle
Critical interrogation
Outputs questioned
Epistemic agency
Whose knowledge?
Collective / relational
Critique with others
Individualized
engagement
AI as personal tool,
neutral & apolitical
neutral & apolitical
The Uncritical Receiver
Accepts outputs; AI naturalized as neutral knowledge source
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): student treats AI outputs as objective, without recognizing that GenAI encodes "opaque, inconsistent cultural assumptions" shaped by historically racist and sexist training data (Keshishi & Hosseini, 2023; Benjamin, 2019).
For educators
Make the assumption of neutrality visible through practical experiments. Ask: whose perspective does a GenAI output reflect by default — in an image, a summary, a clinical recommendation, or a generated essay? Drawing on Day & Esson (2025) and Hosseini (2026), show how seemingly neutral prompts reproduce skewed cultural defaults across output types.
For students
The student experiences AI outputs as "natural" rather than constructed. As Benjamin (2019) argues, socio-technical artefacts are not static reflections — they are shaped by the feedback and values of those who built them. Students need a framework to see this, not just permission to question.
The Cautious Pragmatist
Checks outputs; AI still framed as neutral instrument
Theoretical anchor
The student audits outputs for factual errors but not for the cultural assumptions encoded in them. As Day & Esson (2025) show, even anomalous outputs require users to "remain vigilant to the opaque and shifting nature of generative AI tools" — vigilance the Cautious Pragmatist applies technically but not epistemically.
For educators
Shift from "is this accurate?" to "whose accuracy?" Note that improved outputs do not mean underlying biases have been addressed (Hosseini, 2026). Use Spennemann & Oddone's (2025) technique of asking GenAI to explain its own outputs as a critical exercise.
For students
May believe that better prompting solves the problem. However, prompt refinement "would not address the underlying biases within the datasets themselves" (Hosseini, 2026). The student needs to move from refining inputs to interrogating the training data and the colonial logics embedded within it.
The Epistemically Alert
Interrogates whose knowledge is centered; notices silences
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): student recognizes that AI systems embed a "dominant, Eurocentric worldview" that upholds hierarchical, racialized, and gendered ways of knowing. Connects to Noble's (2018) algorithms of oppression and Maalsen's (2023) algorithmic epistemologies.
For educators
Move from naming bias to interrogating its origin. Use Quijano's (2000) coloniality of power to show that AI's racial and gender defaults are not errors but expressions of colonial hierarchies embedded in training data. Ask: what would an AI trained on non-Eurocentric datasets produce differently?
For students
May feel isolated, especially when institutional AI guidance frames the issue as a technical problem. Wilby & Esson's (2024) call for "capabilities, caveats, and criticality" provides legitimizing language. Connect to communities of practice doing this work.
The Isolated Disruptor
Critiques AI alone; change without solidarity
Theoretical anchor
Individual critique of algorithmic coloniality, however sophisticated, cannot address structural problems in proprietary and inaccessible training datasets (Amoore et al., 2024). Mohamed et al. (2020) are explicit: structural change requires "political coalitions and communities," not individual actors.
For educators
Connect students to collective and cross-disciplinary action. Addressing algorithmic coloniality requires breaking down "disciplinary and departmental silos" (Hosseini, 2026; Maalsen, 2023). Individual insight without structural leverage changes nothing about the datasets or systems producing harmful outputs.
For students
Risk of cynicism or disengagement when individual critique runs up against inaccessible, proprietary datasets. As Hosseini (2026) demonstrates, surface improvements in GenAI outputs can mask rather than resolve underlying colonial logics — students need community and strategy, not just analysis.
Partial structural
awareness
Senses bias or harm,
lacks systemic account
lacks systemic account
The Uneasy Adapter
Senses something wrong; lacks language to name it
Theoretical anchor
Pre-conceptual awareness of algorithmic harm: student senses that something is "off" in AI outputs — perhaps noticing racial or gender skew — but has not yet encountered the theoretical vocabulary to name it. This is the moment described by Day & Esson (2025) when outputs produce "surprising results."
For educators
This is a threshold moment. Offer concepts — algorithmic coloniality (Mohamed et al., 2020), algorithms of oppression (Noble, 2018), socio-technical artefacts (Benjamin, 2019) — as language for what is already felt. Hosseini's (2026) method of prompting GenAI and critically analyzing outputs is a replicable pedagogical entry point.
For students
High potential. Already doing affective critical work. Avoid rushing to resolution — the unease is epistemically productive. GenAI outputs should be approached "not [as] surprising, but as symptomatic of racialised and gendered logics" (Hosseini, 2026) embedded in training data across all output modalities.
The Informed Skeptic
Identifies bias in outputs; most common profile
Theoretical anchor
Can identify racial and gender skew in outputs — consistent with quantitative evidence (Cheong et al., 2024; Currie et al., 2024, 2025) — but frames it as a dataset problem rather than an expression of algorithmic coloniality (Mohamed et al., 2020). The systemic account is absent.
For educators
Move from "bias as glitch" to "bias as design." Use Benjamin's (2019, p. 59) argument that training datasets carry "the prejudices of the individuals who compiled them." Ask: why does a GenAI default encode particular assumptions about race, class, gender, or expertise?
For students
May believe that surface improvements will resolve the issue. Hosseini (2026) demonstrates directly that successive GenAI model versions produced aesthetically improved outputs while reproducing the same racial and gendered logic. The technical fix does not address colonial logics in the training data.
The Structural Analyst
Names AI harms systemically; connects to power
Theoretical anchor
Understands AI as a socio-technical artefact (Benjamin, 2019) shaped by Silicon Valley's role as "part of the United States, a global hegemon and a successor to European colonial powers" (Keshishi & Hosseini, 2023). Connects algorithmic coloniality (Mohamed et al., 2020) to concrete outputs.
For educators
Deepen from analysis to action. Introduce reparative description (Parry, 2023): how might geographers work with public image repositories to revise false past categorizations? Introduce Zembylas's (2023) strategies for "undoing the ethics of digital neocolonialism."
For students
May become frustrated that structural analysis does not translate into change. Channel into cross-disciplinary collaboration. Addressing problematic training data requires collective action and "relational approaches that emphasize the spatial and political contexts of algorithms" (Maalsen, 2023; Hosseini, 2026).
The Emerging Ally
Seeks solidarity; building shared critical vocabulary
Theoretical anchor
Transitional position between individual and collective consciousness (Freire). Recognizes that critique must be collective but lacks the structural analysis to ground it yet.
For educators
Facilitate cross-disciplinary collaboration explicitly. Addressing algorithmic harm requires breaking "disciplinary and departmental silos" (Hosseini, 2026; Maalsen, 2023). Connect emerging allies to existing coalitions and communities of practice doing this work.
For students
Motivated by justice but may lack the analytical vocabulary to sustain critique under institutional pressure. Pairing with theoretically grounded peers — including those with lived experience of the harms being analyzed — is more generative than educator-only support.
Structural
consciousness
AI as site of
coloniality & harm
coloniality & harm
Conscientized but Constrained
Sees the system; defers under institutional pressure
Theoretical anchor
Understands algorithmic coloniality and its harms but operates in institutional systems — curriculum, assessment, professional bodies — that have not caught up with the critique. Within many national contexts "there are nascent discussions on the ethical issues of using Gen AI technologies within tertiary education" (Hosseini, 2026).
For educators
Name the institutional lag explicitly. Developing "algorithmic literacy as part of wider digital literacy initiatives" (Kong et al., 2023; Zembylas, 2023; Hosseini, 2026) is a growing expectation — students can actively contribute to shaping it rather than waiting for institutions to catch up.
For students
Risk of internalizing structural constraint as personal inadequacy. The student's tension is not a sign of failure — it is evidence of structural contradictions that institutions have not yet resolved. Validate the critique while building pathways to act within and against institutional constraints.
The Critical Refuser
Refuses metaphorical framing; acts on structural critique
Theoretical anchor
Tuck & Yang (2012): decolonization is not a metaphor. Student refuses cosmetic diversity framings and demands structural change to what AI produces and whom it serves.
For educators
Support with Mohamed et al.'s (2020) practical recommendations: identifying sites of coloniality in AI systems, understanding where and how algorithms are made, engaging in reparative description (Parry, 2023), and developing local and national policy challenges to colonial algorithmic logics.
For students
May encounter resistance from colleagues who frame AI critique as technophobia. Documentation and publication — as Hosseini (2026) demonstrates — transforms resistant practice into sharable pedagogical resource. Connect to communities doing this work across disciplines.
The Critical Collaborator
Challenges AI's epistemic order; builds alternatives
Theoretical anchor
Actively participates in co-creating "instructional materials that transcend boundaries" (Hosseini, 2026). Draws on intersectionality (Crenshaw, 1991; Hill Collins, 2019) to hold race, gender, and class in simultaneous analysis rather than treating each as a separate problem.
For educators
Commission rather than assess. Meaningful critique of algorithmic coloniality requires centering those with lived expertise in the harms being analyzed — not as informants but as co-authors (Hosseini, 2026). Invite co-authorship, co-design, and co-delivery.
For students
Risk of co-option — being absorbed as institutional evidence of diversity without structural change. Hosseini's (2026) reflexive positioning models how genuine co-production differs from performative consultation. Support students to name and resist this distinction.
The Praxis Collective aspirational*
Reflection + action with others; pluriversal praxis
Theoretical anchor
Camacho Felix's (2025) decolonial imaginations and collective imagination. Mohamed et al.'s (2020) political coalitions. Benjamin's (2019) abolitionist tools for dismantling the New Jim Code in AI systems.
For educators
Collective praxis around GenAI requires institutional conditions: time, resource, partnership, and willingness to redistribute epistemic authority. It demands cross-disciplinary collaboration, reparative dataset work (Parry, 2023), and policy advocacy — none of which individual pedagogy alone can produce.
For students
Students here are co-researchers and co-educators. Hosseini (2026) models this directly: conducting experiments, publishing findings, and encouraging readers to replicate and extend the work with a critical eye. Sustain rather than assess.
← epistemic deferencecollective / relational agency →
Movement across these axes is non-linear — individuals may hold multiple positions simultaneously across different contexts and knowledge domains
Theoretical grounding
Horizontal axis: Mohamed et al. (2020) — algorithmic coloniality; Noble (2018) — algorithms of oppression; Maalsen (2023) — algorithmic epistemologies · Vertical axis: Benjamin (2019) — socio-technical artefacts encoding racial inequity; Zembylas (2023) — decolonial AI in HE; Quijano (2000) — coloniality of power; Camacho Felix (2025) — decolonial imaginations and collective action
This view reframes the framework for researchers engaging with generative AI across the research process. The two applied sections in each cell address the researcher's own practice and the dimension of epistemic accountability — a concept that holds together citational justice (the obligation to credit ideas and intellectual labour whether or not they appear in a formal reference list) and research integrity (the obligation to be transparent about the conditions, limitations, and role of AI in knowledge production). Together they ask: is this knowledge claim accountable — to its sources, its methods, and the communities it draws on?
Positions are not fixed. A researcher may occupy different cells across different phases of the research lifecycle, and across different relationships — with their own field, with collaborators, with communities whose knowledge they draw on.
Epistemic deference
AI as neutral oracle
Critical interrogation
Outputs questioned
Epistemic agency
Whose knowledge?
Collective / relational
Critique with others
Individualized
engagement
AI as personal tool,
neutral & apolitical
neutral & apolitical
The Uncritical Receiver
Accepts AI-generated syntheses as authoritative; does not interrogate sources
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): researcher treats AI literature syntheses, methodology suggestions, and summaries as objective, without recognising that GenAI encodes dominant epistemological hierarchies in what it surfaces and what it buries.
For researchers
The most immediate risk is invisible citational erasure — AI synthesises ideas without naming their origins, and the researcher reproduces that synthesis without tracing it back. This is not plagiarism in the conventional sense; it is the structural absorption of others' intellectual labour without acknowledgment. Check: could you trace every conceptual move in an AI-assisted draft back to a named source?
Epistemic accountability
AI-generated text disproportionately surfaces and centres Western, English-language, and institutionally prestigious scholarship. Accepting this output without interrogation reproduces existing patterns of citational exclusion at scale — particularly for Indigenous, Global South, and community-based knowledge traditions. Whose ideas are being absorbed here, and are they being named?
The Cautious Pragmatist
Verifies factual accuracy; does not interrogate epistemological defaults
Theoretical anchor
The researcher fact-checks AI outputs but accepts their epistemological framing — which fields count as relevant, whose scholars are cited, what counts as evidence. Improved fluency in AI outputs does not mean the underlying knowledge hierarchies have shifted (Hosseini, 2026).
For researchers
Shift from "is this reference real?" to "whose scholarship does this AI treat as foundational?" Run the same research question through AI and check: which traditions, languages, and geographic contexts are centred in the output? Whose frameworks are presented as universal? This is the citational audit that fact-checking alone does not perform.
Epistemic accountability
The Cautious Pragmatist may catch hallucinated citations but will not notice that the scholars being cited are systematically drawn from particular traditions. AI reproduces the citational habits of its training data — which over-represents certain journals, languages, and institutional contexts. Active remediation requires deliberately seeking out and naming scholarship that AI sidelines.
The Epistemically Alert
Interrogates whose knowledge AI centres; actively traces silences
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): researcher recognises that AI embeds a dominant, Eurocentric worldview in what it treats as relevant, credible, and foundational. Connects to Maalsen's (2023) algorithmic epistemologies: AI does not just reflect knowledge hierarchies — it reproduces and legitimises them.
For researchers
Use AI-generated literature maps as a diagnostic, not a foundation. Where are the gaps? Which fields, languages, and traditions are absent? Supplement systematically rather than incidentally. Consider naming the epistemological limits of any AI-assisted literature review explicitly in your methodology — this is a form of reflexive transparency that strengthens rather than weakens the work.
Epistemic accountability
Epistemic alertness is the precondition for citational justice. It is not enough to add diversity to a reference list; the question is whether the conceptual architecture of the work — its framing, its definitions, its theoretical moves — draws on a genuinely plural set of traditions. This researcher is positioned to ask that question; the task is to act on it consistently.
The Isolated Disruptor
Critiques AI's epistemic defaults alone; lacks relational infrastructure
Theoretical anchor
Individual critique of algorithmic coloniality, however rigorous, cannot change proprietary training datasets or the citational norms of major journals (Amoore et al., 2024). Mohamed et al. (2020): structural change requires political coalitions, not individual scholarly positioning.
For researchers
The risk here is that critique becomes a scholarly identity rather than a lever for change. Publishing papers that name AI's citational failures is necessary but not sufficient if those papers circulate only within already-convinced communities. The question is: what collective structures — journal policies, funder requirements, disciplinary norms — could be shifted, and who are the coalitions that could shift them?
Epistemic accountability
Citational justice enacted alone — by one researcher choosing to diversify their own reference lists — is valuable but limited. It does not change the structural conditions that produce citational inequality: impact factor systems, language hierarchies in publishing, paywalled access to scholarship from the Global South. Connecting individual practice to structural advocacy is the work this position points toward.
Partial structural
awareness
Senses bias or harm,
lacks systemic account
lacks systemic account
The Uneasy Adapter
Senses something off in AI-assisted research; lacks conceptual language
Theoretical anchor
Pre-conceptual awareness of algorithmic harm: the researcher senses that AI-assisted literature searches, writing suggestions, or methodology outputs are skewed — but lacks the vocabulary to name what the skew is or where it comes from (Fricker, 2007, hermeneutical injustice).
For researchers
The unease is epistemically productive — it is evidence that the researcher is already doing critical work at an affective level. The conceptual vocabulary that names this unease includes algorithmic coloniality (Mohamed et al., 2020), citational injustice (Ahmed, 2017), and epistemic extractivism — the process by which AI absorbs and repackages knowledge without crediting its origins. Naming the problem is the first step to acting on it.
Epistemic accountability
The sense that "something is missing" from an AI-generated literature map is often precisely citational: the scholars whose work would reframe the question are absent. This researcher is close to articulating that absence — what is needed is the vocabulary to name which traditions are missing and why their exclusion is not accidental.
The Informed Skeptic
Identifies specific AI biases; does not yet connect to structural account
Theoretical anchor
Can identify specific instances where AI outputs reflect racial, gendered, or disciplinary bias — but frames this as a data quality problem rather than an expression of algorithmic coloniality (Mohamed et al., 2020). Instance-level critique without structural account.
For researchers
The next move is from "this AI output is biased" to "this AI output reflects a system trained on scholarship that was already biased toward particular traditions, and my use of it without intervention reproduces that bias in my work." Benjamin's (2019) account of how socio-technical systems encode the assumptions of their makers provides the structural bridge this position needs.
Epistemic accountability
The Informed Skeptic notices citational gaps but treats them as omissions to be corrected individually rather than as expressions of structural inequality. The shift required is recognising that AI's citational habits are not random: they reflect and reinforce the hierarchies of academic publishing — impact factors, language, geography, institutional prestige — that already disadvantage particular scholars and traditions.
The Structural Analyst
Accounts for AI's epistemic harms systemically in research practice
Theoretical anchor
Understands AI as a socio-technical artefact (Benjamin, 2019) whose outputs in research contexts — literature syntheses, methodology suggestions, analytical framings — encode the colonial hierarchies of their training data. Connects Mohamed et al.'s (2020) algorithmic coloniality directly to their own methodological choices.
For researchers
Analysis must connect to action in the research process itself: not just naming AI's structural harms in a methods section, but making methodological choices that actively counter them — using AI as one input among many, supplementing systematically from excluded traditions, and being explicit in publications about what AI cannot see. Describing and contextualising AI outputs in ways that make their structural conditions visible to others offers a model for how this transparency can be enacted.
Epistemic accountability
Structural analysis of AI's citational defaults should translate into active citational practice: naming the scholars whose ideas inform the work even when they are not formally cited, acknowledging intellectual debts to communities and interlocutors, and refusing the convention that only peer-reviewed, institutionally affiliated sources count as citable knowledge. Sara Ahmed's (2017) citational politics offers a principled framework for this.
The Emerging Ally
Building shared critical research practice; seeking community
Theoretical anchor
Transitional position between individual and collective critical consciousness (Freire, 1970). Recognises that AI's structural harms cannot be addressed by individual researchers alone — but has not yet developed the theoretical grounding or the relational infrastructure to act collectively.
For researchers
The institutional structures for collective critical research practice around AI are still nascent — but they exist: research networks, journal special issues, funding calls, and disciplinary working groups focused on AI ethics and decolonisation. Connecting to these structures — research networks, disciplinary working groups, and journal special issues focused on AI ethics and decolonisation — is more generative than developing critique in isolation.
Epistemic accountability
Building shared critical citational practice is more effective than individual reform. Research groups that collectively commit to citational justice — explicitly discussing whose scholarship they are drawing on, whose they are missing, and why — create the conditions for structural change that individual practice cannot. This researcher is positioned to initiate those conversations.
Structural
consciousness
AI as site of
coloniality & harm
coloniality & harm
Conscientized but Constrained
Understands AI's structural harms; institutional pressures limit action
Theoretical anchor
Understands algorithmic coloniality and its research implications but operates within institutional structures — REF metrics, funder requirements, journal conventions, impact factor pressures — that have not caught up with the critique. The constraint is structural, not personal.
For researchers
Identify the specific points in the research lifecycle where structural consciousness can be enacted without institutional penalty — and the points where it cannot yet. Methodology transparency, supplementary citational notes, and collaborative publications are lower-risk entry points. The longer-term task is building the disciplinary arguments that shift what counts as rigorous and responsible AI use in research.
Epistemic accountability
The pressure to cite high-impact, English-language, institutionally prestigious scholarship is real and career-consequential. Citational justice in this position requires identifying which aspects of that pressure can be resisted now — in acknowledgments sections, in working papers, in teaching materials — and building the disciplinary case for broader reform over time.
The Critical Refuser
Refuses AI's epistemic defaults; enacts structural critique in research practice
Theoretical anchor
Tuck & Yang (2012): decolonization is not a metaphor. The researcher refuses to treat AI as a neutral research tool and refuses the cosmetic diversity framings — diverse prompt outputs, diverse research team images — that leave colonial data structures intact.
For researchers
Refusal in research practice is not disengagement from AI but a theoretically grounded decision about what it will and will not be used for. Mohamed et al.'s (2020) practical framework — identifying sites of coloniality, understanding where algorithms are made, engaging in reparative description — provides the action vocabulary this position needs. Document and publish these choices: refusal is more powerful as shared practice than as individual stance.
Epistemic accountability
Refusing AI's citational defaults means refusing to let AI determine whose scholarship is foundational to the work. It means actively tracing ideas to their origins — including unpublished work, conference presentations, community knowledge, and conversations — and naming those origins in the text, not just in the references. Ahmed's (2017) practice of making the politics of citation explicit is the model here.
The Critical Collaborator
Co-produces research that challenges AI's epistemic order; centers lived expertise
Theoretical anchor
Draws on intersectionality (Crenshaw, 1991; Hill Collins, 2019) to hold race, gender, and class in simultaneous analysis. Hosseini (2026) models this directly: centering colleagues with lived expertise in the harms being analyzed — not as data sources or acknowledgment entries but as co-authors and co-thinkers whose intellectual contributions shape the work.
For researchers
The question this position raises is not just who is on the research team but whose ideas are structuring the research — and whether that is reflected in authorship, acknowledgment, and citation. Co-production is not adding diverse voices to an existing framework; it is allowing those voices to reshape the framework itself. This is the distinction between consultation and genuine epistemic collaboration.
Epistemic accountability
Citational justice at this position means that intellectual contributions shape authorship decisions — not just acknowledgment entries. It also means naming the ideas of collaborators, interlocutors, and community members in the text of the work, even where formal co-authorship is not possible. The goal is to make the relational and collective character of knowledge production visible, against the convention of the solo scholarly voice.
The Praxis Collective aspirational*
Collective research action; pluriversal knowledge production
Theoretical anchor
Camacho Felix's (2025) decolonial imaginations: the collective envisioning and construction of research practices that do not reproduce colonial knowledge hierarchies. Freire's (1970) praxis: reflection and action undertaken with others, not on behalf of them. The goal is not one critical research community but the recognition that multiple knowledge traditions have equal standing (Mignolo & Walsh, 2018).
For researchers
Collective praxis in research requires institutional conditions that most researchers do not currently have: protected time for collaborative methodological reflection, funding structures that recognise community knowledge, and publishing conventions that can accommodate collective and non-Western authorship. Building those conditions is itself a research and advocacy task — one that this position is equipped to pursue.
Epistemic accountability
At this position, citational justice is not a corrective add-on but constitutive of the research itself. The work is designed from the outset to make its knowledge relations visible: who contributed what, whose frameworks are being used and from where, what intellectual debts are owed and to whom. This is scholarship that models the knowledge relations it advocates for — pluriversal, relational, and transparent about the conditions of its own production.
← epistemic deferencecollective / relational agency →
Movement across these axes is non-linear — researchers may hold different positions across different phases of the research lifecycle and different collaborative relationships
Theoretical grounding
Horizontal axis: Mohamed et al. (2020) — algorithmic coloniality; Noble (2018) — algorithms of oppression; Maalsen (2023) — algorithmic epistemologies · Vertical axis: Benjamin (2019) — socio-technical artefacts; Zembylas (2023) — decolonial AI in HE; Quijano (2000) — coloniality of power; Camacho Felix (2025) — decolonial imaginations · Epistemic accountability: Ahmed (2017) — citational politics and naming intellectual genealogies; Fricker (2007) — hermeneutical injustice as structural condition of not having vocabulary to name epistemic harm; Tuck & Yang (2012) — decolonization as structural and political; UK Research Integrity Office / Singapore Statement (2010) — research integrity as transparency about conditions of knowledge production
This view adds epistemic accountability as a fourth dimension alongside the original three sections. Epistemic accountability holds together two obligations that AI use makes newly urgent: citational justice — recognising whose ideas are absorbed through AI-mediated synthesis and naming intellectual contributions that fall outside formal citation conventions — and research integrity — being transparent about the conditions, limitations, and role of AI in knowledge production. Neither subordinates the other. Citational justice asks whose labour is acknowledged; research integrity asks whether the conditions of production are honestly disclosed. Together they ask: is this knowledge claim accountable — to its sources, its methods, and the communities it draws on?
This view preserves the educational framing (theoretical anchor, for educators, for students) and adds epistemic accountability as a cross-cutting dimension that applies regardless of whether the user is a researcher, educator, or student. AI use in educational contexts raises the same questions of intellectual credit, attribution transparency, and relational accountability as it does in formal research.
Epistemic deference
AI as neutral oracle
Critical interrogation
Outputs questioned
Epistemic agency
Whose knowledge?
Collective / relational
Critique with others
Individualized
engagement
AI as personal tool,
neutral & apolitical
neutral & apolitical
The Uncritical Receiver
Accepts outputs; AI naturalized as neutral knowledge source
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): treats AI outputs as objective, without recognising that GenAI encodes dominant cultural assumptions shaped by historically racist and sexist training data (Benjamin, 2019).
For educators
Make the assumption of neutrality visible. Ask: whose perspective does a GenAI output reflect by default? Show how neutral prompts reproduce skewed cultural defaults across output types (Day & Esson, 2025; Hosseini, 2026).
For students
AI outputs feel "natural" rather than constructed. Socio-technical artefacts are shaped by the values of those who built them (Benjamin, 2019). Students need a framework to see this, not just permission to question.
Epistemic accountability
At this position, ideas absorbed through AI are unlikely to be traced to their origins at all. This is invisible intellectual debt — the uncritical receiver inherits AI's citational erasures without knowing it. The first step is simply asking: where did this idea come from, and who first articulated it?
The Cautious Pragmatist
Checks outputs; AI still framed as neutral instrument
Theoretical anchor
Audits outputs for factual errors but not cultural assumptions. Improved outputs do not mean the underlying biases have been addressed (Hosseini, 2026).
For educators
Shift from "is this accurate?" to "whose accuracy?" Ask students to run the same question through AI and identify which scholarly traditions dominate the output and which are absent.
For students
Prompt refinement "would not address the underlying biases within the datasets themselves" (Hosseini, 2026). Move from refining inputs to interrogating the colonial logics embedded in training data.
Epistemic accountability
Verifying that citations are real is not the same as asking whose scholarship the AI treats as foundational. The Cautious Pragmatist may catch hallucinated references but miss the systematic over-representation of Western, English-language scholarship. Active remediation means deliberately supplementing AI outputs from excluded traditions and naming those sources.
The Epistemically Alert
Interrogates whose knowledge is centered; notices silences
Theoretical anchor
Algorithmic coloniality (Mohamed et al., 2020): AI embeds a dominant, Eurocentric worldview in what it treats as relevant, credible, and foundational (Noble, 2018; Maalsen, 2023).
For educators
Move from naming bias to interrogating its origin. Use Quijano's (2000) coloniality of power to show that AI's defaults are not errors but expressions of colonial hierarchies in training data.
For students
May feel isolated when institutional AI guidance frames the issue as technical. Wilby & Esson's (2024) call for "capabilities, caveats, and criticality" provides legitimizing language. Connect to communities doing this work.
Epistemic accountability
Epistemic alertness is the precondition for citational justice. This position asks not just "whose knowledge is missing?" but "whose intellectual labour has been absorbed without credit?" — including ideas circulating in communities, conferences, and conversations that predate their appearance in peer-reviewed form.
The Isolated Disruptor
Critiques AI alone; change without solidarity
Theoretical anchor
Individual critique cannot address structural problems in proprietary datasets (Amoore et al., 2024). Structural change requires political coalitions, not individual actors (Mohamed et al., 2020).
For educators
Connect to collective and cross-disciplinary action. Individual insight without structural leverage changes nothing about the datasets producing harmful outputs (Hosseini, 2026; Maalsen, 2023).
For students
Risk of cynicism when individual critique runs up against inaccessible, proprietary systems. Students need community and strategy, not just analysis (Hosseini, 2026).
Epistemic accountability
Diversifying one's own reference list in isolation does not change the structural conditions that produce citational inequality. Collective citational politics — research groups, journals, funders — is what changes norms. This position points toward the need for that collective, even if it has not yet been built.
Partial structural
awareness
Senses bias or harm,
lacks systemic account
lacks systemic account
The Uneasy Adapter
Senses something wrong; lacks language to name it
Theoretical anchor
Pre-conceptual awareness of algorithmic harm: senses skew but lacks vocabulary to name it — Fricker's (2007) hermeneutical injustice at the level of AI engagement.
For educators
Threshold moment. Offer concepts — algorithmic coloniality (Mohamed et al., 2020), algorithms of oppression (Noble, 2018) — as language for what is already felt (Hosseini, 2026).
For students
High potential. The unease is epistemically productive. GenAI outputs are "not surprising, but symptomatic of racialised and gendered logics" (Hosseini, 2026). Avoid rushing to resolution.
Epistemic accountability
The sense that something is missing from AI outputs is often precisely citational — the scholars who would reframe the question are absent. This position is close to naming that absence; the vocabulary of citational injustice (Ahmed, 2017) and epistemic extractivism can give it form.
The Informed Skeptic
Identifies bias in outputs; lacks structural account
Theoretical anchor
Identifies racial and gender skew (Cheong et al., 2024; Currie et al., 2024) but frames it as a dataset problem rather than algorithmic coloniality (Mohamed et al., 2020).
For educators
Move from "bias as glitch" to "bias as design." Training datasets carry "the prejudices of the individuals who compiled them" (Benjamin, 2019, p. 59). Ask: why does GenAI encode particular assumptions about race, class, and expertise?
For students
Surface improvements do not resolve colonial logics. Hosseini (2026) demonstrates that successive model versions reproduced the same racial and gendered logic despite aesthetic improvement.
Epistemic accountability
The Informed Skeptic notices citational gaps but treats them as individual omissions rather than structural patterns. The shift required is recognising that AI's citational habits reflect publishing hierarchies — impact factors, language, geography, prestige — that systematically disadvantage certain scholars. Correction requires structural awareness, not just addition.
The Structural Analyst
Names AI harms systemically; connects to power
Theoretical anchor
Understands AI as a socio-technical artefact (Benjamin, 2019) shaped by colonial power structures (Keshishi & Hosseini, 2023). Connects algorithmic coloniality (Mohamed et al., 2020) to concrete outputs.
For educators
Deepen from analysis to action. Introduce reparative description (Parry, 2023) and Zembylas's (2023) strategies for "undoing the ethics of digital neocolonialism."
For students
May be frustrated that structural analysis does not translate into change. Channel into cross-disciplinary collaboration (Maalsen, 2023; Hosseini, 2026). Analysis without community and outlet risks paralysis.
Epistemic accountability
Structural analysis should translate into active citational practice: naming scholars whose ideas inform the work even outside formal citation, acknowledging intellectual debts to communities, and refusing the convention that only peer-reviewed institutional sources count as knowledge. Ahmed's (2017) citational politics is the principled framework here.
The Emerging Ally
Seeks solidarity; building shared critical vocabulary
Theoretical anchor
Transitional between individual and collective consciousness (Freire). Recognises critique must be collective but lacks structural grounding and relational infrastructure yet.
For educators
Facilitate cross-disciplinary collaboration explicitly. Addressing algorithmic harm requires breaking disciplinary silos (Hosseini, 2026; Maalsen, 2023). Connect to existing coalitions and communities of practice.
For students
Motivated by justice but may lack analytical vocabulary to sustain critique under institutional pressure. Pairing with theoretically grounded peers — including those with lived experience — is more generative than educator-only support.
Epistemic accountability
Building shared critical citational practice within a group — explicitly discussing whose scholarship is being drawn on and whose is absent — creates conditions for structural change that individual practice cannot. This researcher is positioned to initiate those conversations and model collective accountability in citation.
Structural
consciousness
AI as site of
coloniality & harm
coloniality & harm
Conscientized but Constrained
Sees the system; defers under institutional pressure
Theoretical anchor
Understands algorithmic coloniality and its harms but operates within institutional systems that have not caught up with the critique (Hosseini, 2026). The constraint is structural, not personal.
For educators
Name the institutional lag explicitly. "Algorithmic literacy as part of wider digital literacy initiatives" is a growing expectation (Kong et al., 2023; Zembylas, 2023) — students can contribute to shaping it rather than waiting for institutions to catch up.
For students
The tension is not personal failure — it is evidence of structural contradictions institutions have not resolved. Validate the critique while building pathways to act within and against institutional constraints.
Epistemic accountability
The pressure to cite high-impact, English-language scholarship is real and career-consequential. Citational justice in this position means identifying where it can be enacted now — in acknowledgments, working papers, teaching materials — and building the disciplinary argument for broader reform over time, without waiting for permission.
The Critical Refuser
Refuses metaphorical framing; acts on structural critique
Theoretical anchor
Tuck & Yang (2012): decolonization is not a metaphor. Refuses cosmetic diversity framings and demands structural change to what AI produces and whom it serves.
For educators
Support with Mohamed et al.'s (2020) practical recommendations: identifying sites of coloniality, engaging in reparative description (Parry, 2023), and developing policy challenges to colonial algorithmic logics.
For students
May encounter resistance framing AI critique as technophobia. Documentation and publication transforms resistant practice into shareable resource (Hosseini, 2026). Connect to communities doing this work across disciplines.
Epistemic accountability
Refusing AI's citational defaults means actively tracing ideas to their origins — including unpublished work, conference presentations, community knowledge, and conversations — and naming those origins in the text, not just the reference list. The politics of citation are made explicit, not hidden in a bibliography (Ahmed, 2017).
The Critical Collaborator
Challenges AI's epistemic order; builds alternatives with others
Theoretical anchor
Intersectionality (Crenshaw, 1991; Hill Collins, 2019). Hosseini (2026): centering colleagues with lived expertise as co-authors and co-thinkers whose contributions shape the work, not merely validate it.
For educators
Commission rather than assess. Centering lived expertise means co-authorship, co-design, and co-delivery — not consultation. This student's contribution should shape pedagogy, not merely illustrate it (Hosseini, 2026).
For students
Risk of co-option — absorbed as institutional evidence of diversity without structural change. Hosseini's (2026) reflexive positioning models how genuine co-production differs from performative consultation. Support students to name and resist this distinction.
Epistemic accountability
At this position, citational justice means that intellectual contributions shape authorship — not just acknowledgment entries. It means naming collaborators', interlocutors', and community members' ideas in the body of the work. The relational and collective character of knowledge production is made visible, against the convention of the solo scholarly voice.
The Praxis Collective aspirational*
Reflection + action with others; pluriversal praxis
Theoretical anchor
Camacho Felix (2025): decolonial imaginations and collective action. Mohamed et al. (2020): political coalitions. Freire (1970): praxis undertaken with others, not on behalf of them. Mignolo & Walsh (2018): pluriversality — multiple knowledge traditions with equal standing.
For educators
Collective praxis requires institutional conditions: time, resource, partnership, willingness to redistribute epistemic authority, reparative dataset work (Parry, 2023), and policy advocacy. Educators must build the structures, not just model the position.
For students
Students here are co-researchers and co-educators. Hosseini (2026) models this: conducting experiments, publishing findings, encouraging replication. The goal is ongoing collective action that outlasts the course.
Epistemic accountability
At this position, citational justice is constitutive of the work — not an add-on. The research is designed from the outset to make its knowledge relations visible: who contributed what, whose frameworks are being used and from where, what intellectual debts are owed and to whom. This is scholarship that enacts the knowledge relations it advocates for — pluriversal, relational, and transparent about the conditions of its own production.
← epistemic deferencecollective / relational agency →
Movement across these axes is non-linear — individuals may hold multiple positions simultaneously across different contexts, relationships, and knowledge domains
Theoretical grounding
Horizontal axis: Mohamed et al. (2020); Noble (2018); Maalsen (2023) · Vertical axis: Benjamin (2019); Zembylas (2023); Quijano (2000); Camacho Felix (2025) · Epistemic accountability: Ahmed, S. (2017). Living a Feminist Life. Duke University Press — the practice of naming intellectual genealogies explicitly in the text, including debts outside formal citation conventions. Epistemic accountability extends this to research integrity: transparent disclosure of AI use, its limitations, and the structural conditions that shaped what could be known — following the Singapore Statement on Research Integrity (2010) and Fricker's (2007) account of hermeneutical injustice as the structural absence of vocabulary to name epistemic harm.
This view maps AI engagement across the research lifecycle, tracing how epistemic posture — from deference to collective agency — manifests differently at each phase of research practice. Each phase carries distinct risks and possibilities: the literature and framing phase is where AI most aggressively shapes what counts as relevant scholarship; methodology and design is where colonial defaults in AI recommendations become embedded in the research architecture; data and evidence is where AI's training hierarchies most directly affect what gets surfaced and what gets buried; analysis and interpretation is where AI can flatten the conceptual moves that distinguish critical from descriptive work; writing and synthesis is where citational erasure is most invisible; and publication and dissemination is where structural change is most possible — and most resisted.
The epistemic accountability dimension (gold) runs through all phases but intensifies toward dissemination. It addresses two obligations simultaneously: citational justice — whose intellectual labour is acknowledged, including named authors, unnamed collaborators, community knowledge, conversations, and ideas that circulate before they are published — and research integrity — transparent disclosure of how AI was used, what it could not see, and what structural conditions shaped the knowledge produced. Both are present at every phase; the balance between them shifts as the research progresses.
Epistemic deference
AI as oracle
Critical interrogation
Outputs questioned
Epistemic agency
Whose knowledge?
Collective / relational
Critique with others
1. Literature & framing
Identifying what is known and from where
The phase where AI most aggressively determines whose scholarship counts as foundational. Epistemic deference here shapes the entire conceptual architecture of the research.
Delegates framing to AI
Uses AI-generated literature maps as the basis for research framing without interrogating which traditions, languages, or epistemological positions are absent from the output.
Epistemic accountability
Inherits AI's citational exclusions wholesale. Scholars from the Global South, Indigenous traditions, and community-based knowledge are systematically absent — and this absence shapes what the research can ask and how it frames its answers.Fact-checks; accepts framing
Verifies that AI-cited sources are real and accurate but accepts the epistemological frame AI provides — which fields count as relevant, whose scholars are foundational, what counts as a gap in the literature.
Epistemic accountability
Catches hallucinated references but misses systematic citational bias. The canon AI constructs is accepted as the canon; active supplementation from excluded traditions requires a deliberate second pass.Interrogates the canon AI constructs
Uses AI-generated literature maps as a diagnostic: where are the gaps? Which languages, disciplines, and geographic traditions are absent? Treats AI output as a partial and interested map, not a neutral survey of the field.
Epistemic accountability
Actively seeks out scholarship that AI sidelines. Names these sources explicitly in the literature review, including discussion of why they were absent from AI outputs — making the politics of knowledge production visible in the text itself (Ahmed, 2017).Co-constructs the literature frame
Works with collaborators, communities, and interlocutors to identify what counts as relevant literature — rather than delegating that determination to AI. AI is one input among many, weighted by a collectively agreed epistemological framework.
Epistemic accountability
The literature frame itself reflects a collective decision about whose knowledge matters. Intellectual contributions from collaborators and communities that fall outside formal publication are named in the text, not just in acknowledgments.2. Methodology & design
How the research will be conducted and why
AI methodology recommendations embed epistemological defaults — which methods count as rigorous, whose frameworks are treated as standard — that shape what the research can find before data collection begins.
Follows AI methodology suggestions
Treats AI-recommended methods as neutral best practice, without interrogating which epistemological tradition they come from or whose research contexts they were designed for.
Epistemic accountability
Methodology chapters that draw on AI-recommended frameworks without attribution embed intellectual debts that are never acknowledged. Which scholars developed the approaches being used, and are they named?Questions specific method recommendations
Identifies when AI-recommended methods seem ill-suited to the research context, population, or question — but does not yet connect this to the colonial and epistemological conditions that produced those defaults.
Epistemic accountability
Begins to notice that the methodological literature AI surfaces over-represents particular contexts and traditions. Starts to supplement deliberately, but may not yet name the structural reasons for the gap.Designs methodology against AI's defaults
Explicitly asks: what epistemological tradition does this AI-recommended method come from, and is it appropriate for this research? Draws on methodological frameworks from outside the defaults AI provides — including participatory, decolonial, and community-based approaches.
Epistemic accountability
Names the scholars who developed the methodological approaches being used, particularly where those scholars come from traditions AI sidelines. Methodological transparency includes an account of what AI recommended and why those recommendations were departed from.Co-designs methodology with communities
Methodology is co-designed with research participants, communities, or collaborators rather than determined by the researcher alone. AI may be consulted but its suggestions are filtered through a collectively agreed epistemological framework.
Epistemic accountability
Methodological contributions from community collaborators are attributed in the methods section, not just the acknowledgments. Co-design is named as intellectual contribution, not just procedural participation.3. Data & evidence
Gathering, generating, and appraising material
AI's training hierarchies directly affect what gets surfaced and what gets buried. At this phase, algorithmic coloniality is most materially consequential — shaping whose experiences, voices, and data are treated as valid evidence.
Accepts AI-mediated data as representative
Treats AI-generated, AI-retrieved, or AI-summarised data as representative of the phenomenon being studied, without interrogating whose experiences and knowledge are encoded in the training data.
Epistemic accountability
Data generated or retrieved through AI carries the citational exclusions of its training data. Whose voices, experiences, and knowledge are structurally absent from what AI can find or generate — and how does that shape what the research can claim?Supplements AI data with additional sources
Recognises that AI-mediated data has gaps and actively supplements, but frames the gaps as practical limitations rather than expressions of algorithmic coloniality.
Epistemic accountability
Supplementary sources are added but their origins may not be named or their structural exclusion from AI outputs explained. The gap between AI data and supplementary data is a citational and political gap, not just an empirical one.Interrogates what AI cannot see
Treats AI's data limitations as evidence of algorithmic coloniality in action, not as a technical problem. Asks systematically: what kinds of evidence, experience, and knowledge are structurally excluded from what AI can retrieve or generate, and why?
Epistemic accountability
The sources AI cannot see are named and sought out. Evidence from community archives, oral traditions, unpublished reports, and non-English sources is actively included and attributed. The process of retrieval is documented in the methods.Generates data with communities, not about them
Data is generated through relational and participatory processes in which communities are co-producers of evidence rather than subjects of AI-mediated analysis. AI tools are used, if at all, within a framework agreed with research partners.
Epistemic accountability
Community-generated evidence is attributed to its sources. Participants and partners who shape data collection and interpretation are named in the research — not anonymised by default, but with agency over how their contributions appear in the final work.4. Analysis & interpretation
Making sense of what has been found
AI analysis flattens complexity and forecloses interpretive possibilities. The conceptual moves that distinguish critical from descriptive research are most at risk here — as AI tends toward synthesis, consensus, and the resolution of tension.
Accepts AI interpretive frames
Uses AI-generated summaries, thematic analyses, and interpretive framings as the basis for findings, without interrogating the epistemological assumptions that structure them.
Epistemic accountability
Interpretive frameworks absorbed from AI without interrogation carry unattributed intellectual debts. Which theoretical traditions are shaping the analysis — and are their originators named?Checks AI analysis against own reading
Uses AI analysis as a first pass but checks key interpretations against primary material. Identifies when AI over-simplifies or resolves tensions the researcher considers important — but without a structural account of why AI does this.
Epistemic accountability
Begins to notice that AI analysis tends to foreground certain theoretical traditions and background others. May add citations from excluded traditions but does not yet name why they were excluded from AI outputs.Refuses AI's interpretive foreclosures
Treats AI's tendency to synthesise and resolve as an epistemological problem, not a feature. Uses AI analysis as a diagnostic of what dominant frameworks would say, then deliberately works against those framings using theoretical resources AI sidelines.
Epistemic accountability
The theoretical frameworks driving the analysis are named and traced — including frameworks from traditions AI sidelines. Interpretive choices are documented and their intellectual origins acknowledged, including where those origins are conversations and collaborations rather than published texts.Interprets collectively and relationally
Analysis is developed through dialogue — with co-researchers, community partners, and interlocutors — rather than produced by the researcher alone with AI assistance. AI may contribute but does not determine the interpretive frame.
Epistemic accountability
Interpretive contributions from collaborators are attributed in the analysis section, not just in acknowledgments. The collectively produced character of the analysis is made visible — including the names and perspectives of those who shaped how the data was read.5. Writing & synthesis
Producing the scholarly text
Citational erasure is most invisible at this phase. AI-assisted writing absorbs ideas, framings, and formulations without tracing them to their origins — and researchers may not notice because the output reads as their own.
Uses AI-generated text directly
Incorporates AI-drafted sections into the manuscript with minimal modification, accepting the epistemological framing, conceptual vocabulary, and citational choices embedded in the output.
Epistemic accountability
AI-generated text carries absorbed intellectual debts that are invisible in the output. The ideas, formulations, and framings in AI writing originated somewhere — in scholarship, in discourse, in communities — and those origins are erased by default. This is the deepest citational risk in AI-assisted research.Edits AI drafts; adds own voice
Uses AI drafts as scaffolding but revises substantially, adding citations, correcting framing, and introducing theoretical moves AI did not make. The epistemological defaults in AI drafts are partially but not systematically interrogated.
Epistemic accountability
Adding citations to an AI draft is not the same as tracing the intellectual genealogy of the ideas in the text. The question is not only whether references are accurate but whether the scholars whose thinking shaped the argument — directly or indirectly — are named.Writes against AI's framings
Uses AI drafts as a foil: a representation of what the dominant epistemological position would say, against which the researcher's own critical contribution is defined. AI may assist with structure or clarity but does not determine the conceptual moves.
Epistemic accountability
Makes the politics of citation explicit in the text, following Ahmed (2017): naming whose work has shaped the argument, including works that would not appear in a standard literature review, and acknowledging intellectual debts to collaborators and interlocutors who are not formal co-authors.Writes with others; collective voice
The manuscript is produced through collective writing processes that make the relational character of the work visible — co-authored with communities, collaborators, and interlocutors, with AI as one tool among many rather than a ghostwriter.
Epistemic accountability
The text names all those whose intellectual contributions shaped it — whether or not they are listed as authors. Acknowledges ideas that emerged in conversation, community engagement, or collaborative processes that predate the formal research. The acknowledgments section does substantive intellectual work, not just courtesy.6. Publication & dissemination
Sharing findings and building knowledge
The phase where structural change is most possible and most resisted. Publishing choices — venue, format, access, co-authorship conventions — are themselves political acts that reproduce or challenge the knowledge hierarchies AI encodes.
Publishes within existing norms
Targets high-impact journals, follows conventional authorship and citation practices, uses AI-assisted writing to meet volume and speed demands without questioning the structural conditions those demands reproduce.
Epistemic accountability
Publishing in high-impact, subscription-access venues while using AI-generated text that absorbs without crediting scholarship from the Global South enacts a double extraction: taking knowledge and returning it inaccessibly, with origins erased.Questions specific publication norms
Aware that publication conventions — impact factors, language requirements, authorship rules — reproduce inequalities, but navigates them pragmatically rather than challenging them structurally.
Epistemic accountability
May choose open-access venues or add supplementary attribution notes, but does not yet connect these choices to a structural account of how publishing conventions reproduce the same hierarchies as AI's training data.Makes publishing choices politically
Treats publication venue, access model, co-authorship conventions, and dissemination format as political decisions — not just career calculations. Uses Hosseini's (2026) model of publishing findings that encourage replication and community engagement rather than gatekeeping.
Epistemic accountability
Names the intellectual genealogy of the work in the published text — not just in the reference list but in the body of the argument. Advocates within disciplinary and journal contexts for citational norms that recognise non-Western, non-English, and community-based scholarship. Documents and publishes the citational politics of the research itself.Disseminates with communities, not to them
Findings are shared with and through the communities whose knowledge and experience informed the research — before, not after, academic publication where possible. Dissemination is designed as part of the research, not an afterthought. Multiple formats, languages, and venues are used to reach the communities the work is accountable to.
Epistemic accountability
Publications credit all intellectual contributions — including those of community partners and interlocutors — in formats those contributors can access and share. The goal is not only to acknowledge intellectual debt but to return knowledge to its sources in forms that are useful, accessible, and accountable. This is the fullest enactment of epistemic accountability — pluriversal, relational, and structurally honest about the conditions of knowledge production.The lifecycle view maps positions available at each phase — not a required sequence. Researchers will occupy different positions at different phases, and movement is non-linear. The epistemic accountability dimension (gold) is present throughout but intensifies toward publication, where structural inequalities in knowledge recognition are most consequential and most addressable.
Theoretical grounding — lifecycle view
Research lifecycle framing: Mohamed et al. (2020) — algorithmic coloniality across research practice; Maalsen (2023) — algorithmic epistemologies in geographic and social research · Epistemic accountability: Ahmed, S. (2017) — citational politics and naming intellectual genealogies; Singapore Statement on Research Integrity (2010) — transparency about conditions and limitations of knowledge production; Fricker (2007) — hermeneutical injustice as structural absence of vocabulary to name epistemic harm; UK Research Integrity Office guidance on AI use in research (2024) — disclosure of AI role as a dimension of methodological honesty · Relational knowledge production: Camacho Felix (2025) — decolonial imaginations; Tuck & Yang (2012) — decolonization as structural and collective; Hosseini (2026) — reflexive positioning and centering of lived expertise in research
Listening, Speaking, Learning: On Verbal Feedback and (Re)Humanizing Assessment
Sunset over Queens Park pond, Glasgow, UK
Recently, I listened to a podcast from 2022, Educatalks: Reflective Practice featuring Professor Melaine Coward, a professor of medical education reflect on her career and her commitment to reflective practice. Medical Educatalks is a podcast created by the Developing Medical Educators Group (DMEG) at the Academy of Medical Educators. Toward the end of the conversation, she described her decision to give students verbal feedback on their assessments. The interviewer sounded genuinely surprised, he hadn’t encountered that approach before.
I paused.
Not because it felt novel, but because it felt familiar.
In 2015, while teaching at a University of London institution, I experimented with providing verbal feedback on written assignments. At the time, our digital marking platform enabled tutors to attach audio recordings directly to students’ scripts, so feedback could be posted alongside the written work itself. Students could either listen asynchronously or book a short follow-up slot to discuss it further. I would have their script in front of me as I recorded or spoke with them, talking through strengths, misunderstandings, and next steps. It was dialogic, immediate, and relational, but it was not universally welcomed.
The pushback was swift and couched in procedural language:
How could this be standardized?
How could it be moderated?
Where was the audit trail?
Ironically, the digital system did generate an artefact. The audio file was stored alongside the script. There was a record. And yet the discomfort persisted. What seemed to trouble colleagues was not the absence of documentation, but the presence of voice, with its tone, inflection, and spontaneity. Feedback had become less easily reduced to a static text block. The underlying concern was not simply technical. It was cultural. Feedback, in this framing, was not primarily a pedagogical encounter, it was a compliance mechanism.
Listening to Coward years later, I realized something I could not fully articulate back then: verbal feedback is not merely a technique. It is an epistemological stance. It is a small but meaningful act of (re)humanising assessment.
“But what I found was, so they weren’t reading the comments that I’d spent ages putting on marking there, because I do spend time, it matters that I give good feedback. When I did recorded feedback, I found I had a lot more follow-up from students, because they had had to listen to my feedback, and it was a very clear message of, I really enjoyed this, something for you to think about. I would be quite structured in how I recorded it, so I had notes, so it was formulaic in that sense, but not rehearsed.”
Feedback as encounter, not transmission
Higher education assessment cultures are deeply shaped by what Paulo Freire famously critiqued as the “banking model” of education in Pedagogy of the Oppressed. In that model, knowledge is deposited; feedback becomes a written correction of deficits; learning is framed as remediation.
Written feedback, of course, can be thoughtful and transformative. But it often operates within systems that prioritize defensibility over dialogue. Comments are calibrated for external examiners. Language becomes cautious. Tone becomes formal and neutralized. The student becomes a case.
Audio feedback, even when delivered asynchronously through a digital platform, subtly shifts that dynamic. Students hear emphasis. They hear encouragement. They hear uncertainty where appropriate. Meaning is shaped not only by what is said, but how it is said.
And when audio is paired with optional follow-up conversation, feedback becomes dialogic in a deeper sense. Students can respond, query, reinterpret.
This resonates with Freire’s insistence on dialogue as the foundation of emancipatory education. It also aligns with bell hooks’ vision of engaged pedagogy in Teaching to Transgress, where teaching and learning are relational acts rather than one-way transmissions.
When we speak with students rather than at them, feedback becomes less about surveillance and more about growth. Voice, literal voice, reintroduces presence into assessment by (re)humanizing it.
The standardization question
The resistance I encountered in 2015 revolved around standardisation. Written comments were seen as stable, recordable, and therefore fair. Audio feedback, even though stored and retrievable, was viewed as potentially variable.
But here is the uncomfortable truth: standardization is not synonymous with justice.
Critical and decolonial scholars have long questioned whose norms assessment criteria encode. Ngũgĩ wa Thiong'o, in Decolonising the Mind, reminds us that language and evaluation are never neutral; they are embedded within colonial power structures. Similarly, scholars of antiracist pedagogy argue that assessment practices often privilege dominant linguistic and epistemic norms and performances.
Audio feedback can surface some of this hidden curriculum. It allows educators to unpack what we mean by “criticality” or “coherence” in accessible, responsive ways. It can soften deficit framings by conveying nuance and care. It can make tacit expectations explicit.
For neurodivergent students, multilingual students, or those unfamiliar with disciplinary conventions, hearing feedback, with tone and pacing, can support comprehension in ways that dense written comments may not.
Uniform delivery formats may be easier to audit. But equity sometimes requires responsiveness.
Reflective practice and professional identity
Coward’s framing of verbal feedback emerged from reflective practice, a concept often associated with Donald Schön and his work The Reflective Practitioner. Reflection is not merely about improving technique; it is about interrogating the assumptions that underpin our actions.
Looking back, I can see that my 2015 experience exposed a tension between two logics and a clash of paradigms:
Assessment as pedagogical relationship / Was feedback a compliance mechanism or a pedagogical relationship?
Assessment as quality assurance infrastructure / Was my role to produce defensible documentation or to cultivate understanding?
The digital tool itself was neutral. It could host text or voice. The debate was about what counted as legitimate academic labor and legitimate evidence of fairness.
Reflective practice asks us to interrogate not only how we teach, but why certain practices are normalized while others are treated as suspect.
(Re)humanizing assessment in digital spaces
In my current work, including conversations around decolonizing curricula and rethinking assessment, I often return to a simple question:
What would assessment look like if we centered humanity rather than auditability?
This is not an argument to abandon rigor or documentation. Rather, it is a call to re-balance priorities.
(Re)humanizing assessment might include:
Dialogic feedback conversations alongside written summaries
Audio or video feedback that conveys tone and relational presence
Opportunities for students to respond to feedback
Co-constructed criteria discussions
Assessment designs that value multiple ways of knowing
These moves resonate with broader critical pedagogical commitments: resisting neoliberal metrics, challenging deficit framings, and recognizing students as co-participants in knowledge production. These moves further resonate with critical pedagogy’s insistence on dialogue, with antiracist commitments to challenging hidden norms, and with decolonial calls to unsettle inherited hierarchies of knowledge.
They also align with emerging scholarship on compassionate pedagogy and relational assessment cultures within higher education.
“Hearing someone talk about what you’ve done, the tone and voice to highlight praise, concern, and you can add in a more questioning tone ... They loved it. They loved it because they could hear from my touch. ”
An epiphany, years later: are our systems human enough?
Listening to Coward describe her practice, I felt both affirmed and reflective. The surprise expressed by the podcast interviewer revealed how deeply entrenched written, standardized feedback remains. Yet the fact that such practices continue to surface across disciplines, from medical education to the humanities, suggests a quiet shift. I also felt less concerned with whether verbal feedback is innovative and more interested in what it reveals. What I once framed defensively as “innovative feedback” now feels more clearly like a small act of resistance against depersonalized academic systems.
Even when captured and archived in a digital platform, voice unsettles the fantasy that assessment can be entirely standardised and neutral. It reintroduces tone, care, and relational accountability.
Perhaps the question is not whether audio feedback can be moderated.
Perhaps the more urgent question is whether our assessment cultures allow space for humanity, for dialogue, for nuance, for recognition.
If critical, antiracist, and decolonial pedagogies ask us to re-centre people rather than processes, then even something as simple as attaching a recorded voice note to a script can become a quietly radical act.
Suggested further reading
Pedagogy of the Oppressed – Paulo Freire
Teaching to Transgress – bell hooks
Decolonising the Mind – Ngũgĩ wa Thiong'o
A Handbook of Reflective and Experiential Learning – Jennifer A. Moon
The Reflective Practitioner – Donald Schön
And, of course, I would recommend listening to Educatalks: Reflective Practice featuring Melaine Coward, not because verbal feedback is revolutionary, but because reflective conversations about practice remind us that teaching is, at its heart, relational work.
In a sector increasingly governed by metrics, that reminder feels quietly radical.