Start here
- Deciding about one assignment or tool? Read the three levels and watch them run on one decision. Ten minutes.
- Writing a course, program, or campus policy? Open the decision tool and choose your audience. Link straight to it: faculty, institution.
- Reading the scholarship behind the framework? Go to the framework in context.
AI literacy is three different acts. Describe the system. Analyze what is built into it. Then decide. Most AI literacy programs teach the first and skip to the third, and this page is about the missing middle.
The three acts are three kinds of literacy, conceptual, structural, and operative, and together they set the order in which any decision about AI gets reasoned. The framework names them so faculty development programs, governance bodies, and policy work aim at the right one for the right purpose.
The three levelsTop
The framework names three kinds of literacy, each defined by what a person does.
- Conceptual · Describe. What the systems are, and how they produce text.
- Structural · Analyze. How they are built, and whom they disadvantage.
- Operative · Decide. When to use one, how, and when to refuse, with reasons tied to the task.
Structural and operative are the two levels most often read as one. One test separates them. A structural question has the same answer for everyone who meets the system. An operative question has an answer only once a particular person or committee, in a particular situation, decides.
Ethics appears at both levels and does different work at each. Structural literacy studies the harms a system carries into every use. Operative literacy concerns what one person or body owes in a single decision, including the decision to refuse. Structural analysis informs the decision without making it. Two instructors who agree about a model's bias will still reach different, defensible policies for different courses.
The two words carry these senses. Structural refers to what is built in. Operative is used as in the operative clause of a policy, the part with force.
Select any level panel to expand its definition
Level 1 · DescribeConceptual literacy
- You ask
- What is this system, and whose knowledge does it encode?
- You do
- Name the system type and trace how its data shapes its output
- You produce
- A precise description
Level 2 · AnalyzeStructural literacy
- You ask
- Who controls this system, and whom does it disadvantage?
- You do
- Trace ownership, data sources, and what the system optimizes for
- You produce
- An analysis of the conditions built in
Level 3 · DecideOperative literacy
- You ask
- What does this situation warrant: use, limits, or refusal?
- You do
- Weigh the analysis against the task and decide
- You produce
- A decision with stated reasons, owned by you
Every decision runs all three levels in order. A new system or a new decision starts again at Level 1.
Conceptual literacy is the capacity to recognize, name, and describe AI systems with precision. A practitioner at this level identifies what kind of system a tool is, traces how its training data shapes its output, and keeps apart terms often treated as one: an architecture (such as the transformer), a model trained on it, an application built on the model (a chatbot, a detector), and the category names vendors use to sell it (“generative,” “assistant,” “agent”).
Knowledge at this level is primarily technical and definitional. It makes the vocabulary of AI legible as a site of decision-making. Every term in AI discourse encodes a choice about whose cognitive practices count as intelligence, whose outputs count as correct, and whose contributions to the training set are treated as ground truth.
Conceptual literacy is present when a practitioner reads a system description and identifies what the system does along with whose knowledge the system was built to recognize. At this point the definitional question "what is this?" becomes inseparable from the power question "whose knowledge is encoded?"
In practice. Reading a detector vendor’s description, a practitioner states what the tool measures (statistical features of a text), what it reports (a probability score), and what the score does not establish (what a writer did). The practitioner also asks what writing the tool treats as the human baseline.
Signs it is missing. Treating “AI” as one thing. Reading a detector score as a finding. Taking a vendor’s category names as descriptions.
- For teaching
- Knowledge character
- Vocabulary and concept definition.
- Pedagogy
- Direct instruction, with students describing real tools in their own terms.
- Assessment
- A description is checked against the system and graded for accuracy. The question of whose knowledge is encoded gets argued, and it is assessed with Level 2.
Structural literacy is the capacity to analyze conditions inherent to AI systems as information infrastructure, including the sourcing of training data from dominant-language and dominant-culture texts, the reproduction of existing social hierarchies in model outputs, the concentration of AI development within specific corporate and geopolitical contexts, and the mechanisms by which AI assessment instruments scale linguistic hierarchy as if it were cognitive capacity.
Analysis at this level moves from description to diagnosis. It asks what the system reproduces and for whose benefit.
Structural literacy is present when a practitioner traces the chain of production and control behind a system, reads ownership, optimization targets, and data provenance as analytically prior to any question of deployment or use, and treats those conditions as objects of institutional inquiry.
In practice. Examining the same detector, a practitioner traces who built it, what data trained it, and where it sends student work. The findings hold for every user. Detectors misclassify writing by non-native English writers at high rates (Liang et al., 2023). No tool in a fourteen-tool test proved reliable (Weber-Wulff et al., 2023). The vendor keeps its method undisclosed, and each scan sends student work to a third party.
Signs it is missing. Debating a tool’s features without asking who owns it. Treating bias as a defect the next version will fix. Leaving data flows out of the discussion.
- For teaching
- Knowledge character
- Analysis of inherent system conditions.
- Pedagogy
- Critical analysis, with students tracing ownership, training data, and optimization targets for a tool they use.
- Assessment
- Graded on argumentative quality: the evidence and reasoning behind a claim about a system, including the Level 1 question of whose knowledge it encodes.
Operative literacy is the capacity to exercise contextual, accountable judgment about AI use in specific situations, including the evaluation of outputs, ethical reasoning about deployment, assessment of institutional governance frameworks, and principled refusal where structural conditions warrant it.
Judgment at this level is irreducibly particular. The conditions of a specific situation determine the right action, and the practitioner analyzes those conditions through the conceptual and structural capacities the preceding levels provide.
Operative literacy is present when a practitioner articulates a decision together with the structural analysis grounding it, and accepts institutional accountability for the reasoning.
Refusal theory (Simpson, 2014; Campt, 2017; Benjamin, 2019) names the form operative literacy takes at institutional scale: accountable rejection of conditions which reproduce harm, grounded in analysis of those conditions. Refusal in this sense differs from non-compliance. It carries the same duty to give reasons as any other operative decision.
In practice. Working from the structural analysis, an instructor decides whether detector scores have any place in an academic integrity case, writes the decision into the syllabus, and states the reasons. A decision to refuse carries the same duty to give reasons as a decision to adopt.
Signs it is missing. Deciding by default, adopting a tool because the license exists or banning it because the risk feels high. Decisions with no stated reasons. Reasons nobody owns.
- For teaching
- Knowledge character
- Contextual judgment, principled refusal.
- Pedagogy
- Situated and case-based, with students writing and defending a policy for a real course or task.
- Assessment
- Contextual and accountable: graded on whether the reasons fit the situation, draw on the structural analysis, and name who answers for the decision.
Cole, K. K. (2026). AI literacy is not one thing: Conceptual, structural, and operative levels for higher education governance. Pedagogies: An International Journal, forthcoming. https://cwspwolf.com/ai_literacy_unified.html
One Decision, Three LevelsTop
Should we use an AI-writing detector on student work? The question recurs across higher education, and it reaches every audience on this page: a student whose essay gets flagged, an instructor deciding whether to run one, a writing program setting practice, and a university deciding whether to license one. Vanderbilt University disabled Turnitin's AI detector in August 2023 after months of testing. Here is the same question worked through all three levels.
Level 1 · Describe
A detector is a classifier. It estimates how likely a text is to have come from a language model, using statistical features of the writing, and reports a score. A precise description includes what the score does not mean: a probability estimate is not evidence of what a student did. Describing the tool also means asking what writing it treats as the human baseline, in other words, whose writing it was built to recognize.
Produces: a description of what the tool measures and reports.
Level 2 · Analyze
Liang and colleagues (2023) found detectors misclassified most essays by non-native English writers as AI-generated, while rarely misclassifying essays by native writers. Weber-Wulff and colleagues (2023) tested fourteen tools and found none reliable at identifying AI-generated text. The vendor does not disclose its method, and running the tool sends student work to a third party. These conditions hold for everyone who uses the tool.
Produces: an analysis of the conditions built into the tool.
Level 3 · Decide
The analysis informs the decision without making it. An instructor might decline to treat detector scores as evidence of misconduct. A writing program might limit detectors to research on the tools themselves. A university might disable the feature, as Vanderbilt did, citing accuracy, bias, lack of transparency, and privacy. Each decision states its reasons, and the person or body making it owns them. Refusal here is a reasoned decision grounded in the structural analysis. I have made this decision, and the statement below gives its reasons.
Produces: a decision with stated reasons.
When the vendor releases a new version, or a different tool arrives, the sequence starts again at Level 1.
An operative decision, on the record
Why I do not use AI-writing detectors
I run no AI-writing detectors, in my own courses or in the Campus Writing and Speaking Program, and I advise faculty against using detector scores as evidence in any academic integrity process.
My reasons
Detectors misidentify writers. Liang and colleagues (2023) found GPT detectors classified most essays by non-native English writers as AI-generated, while rarely misclassifying essays by native English writers. A tool with this pattern converts linguistic difference into suspicion of misconduct. The harm falls hardest on multilingual students.
Detectors fail their stated purpose. Weber-Wulff and colleagues (2023) tested fourteen detection tools and found none reliable at identifying AI-generated text.
Detector methods stay hidden. Vendors do not disclose how their tools reach a score. A student has no way to contest a judgment whose basis no one will show them.
Detectors move student writing. Each scan sends a student’s work to a third party, under terms the student did not set and often never sees.
To run a detector on student work is to accuse students on evidence they are not allowed to see, with the errors concentrated among multilingual writers. That is an act I would be choosing, and I decline to choose it.
What I do instead
I help faculty design assignments with a visible writing process: drafts, revision histories, conferences, and reflections on choices. I help faculty write AI policies stating what students are permitted to do, and why. Where an instructor suspects misconduct, the evidence is the student’s process and a conversation with the student.
Accountability
This decision belongs to me. I will revisit it when the evidence changes, and I will state my reasons again if I change course.
Kirsti Cole, Co-Director, Campus Writing and Speaking Program, NC State University. October 2026.
Using the frameworkTop
The detector example shows the order in which any decision about AI gets reasoned, and the order holds for a student choosing a tool for one assignment as much as for a university buying a platform. No one completes the first level and moves on. Each new system, and each new decision, starts the sequence again.
- For one decision. Work through the three questions in order before acting.
- At any scale. The questions stay the same for a student, an instructor, a professional, and an institution. Only the decision changes.
- Where you work. The decision tool and the workbook below apply the framework to assignments, course policies, and purchases.
Tool use has no level of its own. Skill with a tool follows from an operative decision to use it, and the framework asks for the decision first.
Governance Application · By AudienceTop
The three-level framework applies beyond faculty development and holds constant across scales. The same three questions (whose knowledge is encoded, who controls this system, what this situation warrants) apply to an individual student's choice on one assignment, a faculty member's policy for one assignment, a professional's use of a tool on one work product, and an institution's procurement decision. Only the decision changes. Select an audience below to see the same analysis worked through for its decision.
Choose an audience, then answer each checkpoint to open the next level.
The AI Literacies Extension · NC State CWSPTop
Generative AI has changed what asking students to write means. Faculty across disciplines are rethinking assignment design, academic integrity policies, and what authorship requires in their fields. The WOLF AI Literacies extension gives faculty a structured place to work through those questions, with frameworks, discussion prompts, and design tools grounded in communication pedagogy.
The extension advocates no single approach to AI in the classroom. It serves faculty who want to think carefully about what they value in student writing and speaking and to make those values explicit in their assignments, whether AI tools are permitted, restricted, or somewhere in between. It shares WOLF's foundations. Like WOLF, it asks what communication is for, whom it serves, and what disciplinary participation requires. It treats AI as a condition of writing instruction, one demanding more explicit and intentional pedagogy. WOLF has always argued for such pedagogy.
Who this is for
The AI Literacies extension suits faculty who are revising course policies in response to generative AI tools, redesigning assignments to assess disciplinary communication directly, trying to articulate to students (and themselves) what writing and speaking are for in their courses, or interested in teaching AI literacies as a transferable skill within their discipline.
The extension works alone or within an ongoing WOLF partnership. Faculty who complete it alongside a WOLF engagement often find it sharpens the Inquiry and Design phases. The questions AI raises about authorship and assessment expose assumptions about communication which had stayed implicit.
The faculty workbook at the center of the extension runs in three clusters:
The clusters apply the sequence at three scales. The first runs it once, on a discipline. The second applies it to one policy. The third asks where each level appears across a curriculum, and where it is missing.
Liaisons facilitate the extension with the WOLF model: sustained conversation, iterative design work, and documentation of decisions and rationale. The workbook is written for liaison-facilitated use. Faculty are welcome to work through it alone. Contact Kirsti Cole at CWSP to begin or to connect your work to a full WOLF engagement. More information is at cwspwolf.com/ai-literacies (opens in a new window).
The Framework in ContextTop
Two reference panels place the framework against the scholarship it draws on and the terms in common circulation. Full citations appear at the bottom of the page.
In conversation with literacy scholarshipSeven scholars this framework draws on, and where it departs from them
Scribner and Cole (1978), "Unpackaging Literacy"
Studying literacy among the Vai in Liberia, Scribner and Cole found the single word "writing" covered a wide range of activities, each with its own skills and consequences. This framework makes the same argument about AI. "AI literacy" names several different capacities, and a program has to say which one it means.
Scribner (1984), "Literacy in Three Metaphors"
Scribner showed each definition of literacy carries an assumption about what literacy is for: adapting to everyday demands, gaining power, or attaining a valued status. This framework asks the power question at every level. Even the first level, description, includes asking whose knowledge a system encodes.
Street (2003), "What's 'New' in New Literacy Studies?"
Street describes literacy as social practice, varying across times and places and contested within relations of power. He sets this account against what he calls the autonomous model, in which literacy is a neutral skill with effects of its own. This framework follows the social-practice account. It treats AI literacy as practices shaped by who builds, owns, and governs the systems.
Brandt (1998), "Sponsors of Literacy"
Brandt defines sponsors as "any agents, local or distant, concrete or abstract, who enable, support, teach, model, as well as recruit, regulate, suppress, or withhold literacy," and who gain advantage by doing so (p. 166). Structural literacy asks Brandt's question of AI. Who sponsors this literacy, and what do they gain? A company offering free training in its own tools is a sponsor in her sense.
Selber (2004), Multiliteracies for a Digital Age
Selber's three literacies (functional, critical, and rhetorical) remain the closest model in writing studies. Structural literacy corresponds to his critical literacy, which treats computers as cultural artifacts. Operative literacy corresponds to his rhetorical literacy, which joins functional and critical abilities in practice. This framework departs from Selber in two places. It has no functional level, and it treats refusal as a literate act.
Keefe and Copeland (2011), "What Is Literacy? The Power of a Definition"
Keefe and Copeland show how restrictive definitions of literacy shut people out of literacy instruction. Their third core principle reads: “Literacy is not a trait that resides solely in the individual person. It requires and creates a connection (relationship) with others” (p. 97). This framework takes their warning as a design rule. Its levels set the order for reasoning through a decision. They do not sort people by readiness, and no one waits to be judged ready for structural analysis.
Perry (2012), "What Is Literacy? A Critical Overview of Sociocultural Perspectives"
Perry compares three sociocultural approaches (literacy as social practice, multiliteracies, and critical literacy) and argues they share concerns without forming one theory. This framework draws on all three. Perry's caution against merging them is one reason the framework keeps its levels distinct.
Common terms, read against this frameworkSix terms in circulation (literacy, competency, levels, proficiency, fluency, pathway) and what each assumes
Literacy
In policy documents, literacy usually means a baseline of knowledge everyone should have. Here, literacy names three kinds of act (describing, analyzing, and deciding), and a sound decision requires all three.
Skills and competencies
Competency frameworks list what a person should know and do with AI. Ng and colleagues (2021) name four aspects: know and understand, use and apply, evaluate and create, and ethics. Lists like these give use a level of its own and set ethics apart as a separate category. This framework places use inside the operative decision. It treats ethics at two levels, as the harms built into a system and as what one person or body owes in a single decision.
Levels
Many domain literacies arrange their levels by degree. Nutbeam's (2000) health literacy rises from functional to interactive to critical, each level defined by more advanced skill. The levels here differ by the kind of act each requires. An expert describing a new system is doing first-level work.
Proficiency
Proficiency scales rate a person as basic, intermediate, or advanced. Applied to AI, they measure how much a person knows or how well they use a tool. This framework asks which kinds of reasoning went into a decision.
Fluency
The National Research Council's 1999 report, Being Fluent with Information Technology, adopted "fluency" for a deeper understanding than computer literacy, combining skills, concepts, and intellectual capabilities. In current institutional use, fluency marks the top of a scale: use without hesitation. Operative literacy asks for a decision made deliberately, with reasons stated and accountability accepted. Ease of use and accountable judgment are different goals. This framework asks for the second.
Pathway
A pathway assumes a destination set in advance, with learners moving toward it in stages. This framework sets no destination. It asks each person to reason every decision through all three levels, and the outcome, including refusal, belongs to the person or body deciding.