Writing programs give students the vocabulary to locate themselves as agents within the cognitive systems they write in. AI policy, institutional and critical alike, names what students should know about AI. It then hands the decision about use to someone else. This page gathers the evidence behind the argument, three questions for programs to ask before the instructor or program decides, and the works cited.

Session materials

The slide deck carries the talk and its speaker notes. The companion document holds the evidence and works cited in an editable form.

Slide deck (opens in a new window) Companion document (opens in a new window)

Room poll

Where does the AI decision sit in your program? Choose the answer closest to what your program does now, whatever the written policy says.

A

A ban

B

A permission chart

C

A detector

D

The individual instructor

E

Nowhere yet

The room voted on Mentimeter. Voting closed on October 15, 2026.

Three levels of AI literacy

The model comes from a scoping review of 250 articles on AI literacy and a reading of university AI governance documents. The levels are different kinds of knowledge, so no checklist teaches all three at once.

Conceptual

What the systems are

How large language models produce text, and what they do not do.

Structural

How they are built and whom they disadvantage

Training data, labor, ownership, and the writers these systems misread.

Operative

When to use one, when to refuse, and why

A decision made inside a specific task, with reasons tied to its purpose. The missing level.

Evidence

Open an entry for the claim, its reading, and a link to the source. Sources verified against primary documents, October 1, 2026.

Distributed cognition

Clark and Chalmers 1998Cognition incorporates its tools

Clark and Chalmers describe the organism and an external resource, pen and paper or language itself, as a coupled system counted as cognitive in its own right.

“The Extended Mind,” Analysis 58 (1): 7–19 (opens in a new window)

Hayles 1999The bounded writer is a construct

Hayles reads the liberal humanist subject as historically specific. Outcomes statements built on a bounded, self-present author inherit an assumption distributed composing has already undone.

How We Became Posthuman, University of Chicago Press (opens in a new window)

Cole 2026 and Cole et al. 2025Students named what AI produced, but not where their own thinking ended

The finding holds across two published assignments in TextGenEd: Continuing Experiments, the Cyborg Composing workshop and the Collaborative Feedback Toolkit.

Cyborg Composing and Posthuman Research Methods (2026) (opens in a new window)
Reflective Multimodal Feedback Practices Across Writing Contexts (2025) (opens in a new window)

Policy delegates the operative level

UChicago Law, July 9, 2026Responsible, effective, and ethical use

One of the strategy’s three themes reads “Teaching the responsible, effective, and ethical use of AI.” The strategy names no one to decide what responsible means in a given course, so the decision falls to the instructor by default.

Alfadel Coloma, “UChicago Law Unveils New AI Strategy” (opens in a new window)

MIT, August 13, 2026The same three terms, the same delegation

Recommendation 3.2.4 reads “Teach effective, responsible, and ethical use of AI,” the same three terms in a different order. Recommendation 3.1.9 advises against relying on AI detectors and warns they “may also mistake the writing of non-native English speakers or neurodivergent students for text generated by AI.” The warning is structural knowledge. It names the writers these systems misread.

Report of MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training (PDF) (opens in a new window)

Goodlad, circulated by AAUP NC, 2026Understand, distinguish, then delegate

In a summary circulated by AAUP North Carolina’s Genuine Intelligence Project, Lauren Goodlad’s first recommendation lists six critical AI literacies. Each asks students to understand or distinguish, which is conceptual and structural work. Her second recommendation allows faculty “to evaluate the learning goals of their own courses and disciplines from the standpoint of generative AI relevance.” The critical side delegates the decision as institutional policy does.

Teaching Critical AI Literacies: recommendations from Lauren Goodlad, circulated by AAUP NC (opens in a new window)

New evidence on the operative level

Marcoccia, Quattrociocchi, Capraro 2026AI advice suppresses “I don’t know”

Across five experiments, access to AI advice nearly eliminated participants’ willingness to suspend judgment, even when the advice was engineered to be wrong. Incentives for accuracy reduced the effect without removing it. Withholding an answer is an operative act, and the study measures its loss. Preprint.

arXiv:2607.13562 (opens in a new window)

Liu et al. 2026Persistence costs follow direct-answer requests

In the second experiment, persistence costs concentrated among participants who had AI solve tasks for them. Participants who used AI for hints persisted at rates close to those without AI. The authors caution this comparison is cross-sectional. Even so, the variable it isolates is how the user decides to engage the tool. Preprint.

arXiv:2604.04721 (opens in a new window)

Strömberg, Lei, Wu 2026Losses concentrate in outsourced homework

Learning losses concentrate among roughly 80 percent of AI users whose short homework times and high homework scores indicate outsourcing. AI users who keep homework time similar to non-users show small losses. Use alone does not predict the outcome. The decision about process does.

CEPR Discussion Paper 21577, June 2026 (opens in a new window)

Detection

Liang et al. 202361.3 percent average false-positive rate

Seven GPT detectors averaged a 61.3 percent false-positive rate on TOEFL essays by non-native English writers. Earlier detectors failed on accuracy, and the failure fell on multilingual writers.

“GPT Detectors Are Biased against Non-Native English Writers,” Patterns (opens in a new window)

Jabarian and Imas 2025Near-zero error moves the problem to premise

Jabarian and Imas find near-zero false positive and false negative rates for Pangram, so the objection shifts. A detector treats authorship as a property of the finished text and hands the authorship judgment to a second AI system, beneath any written policy.

Artificial Writing and Automated Detection, NBER (opens in a new window)

Emig 1977Writing as a mode of learning

Writing studies has located authorship in process since Emig. Drafts, revision histories, and conferences place the question where the learning happens.

“Writing as a Mode of Learning,” CCC 28 (2): 122–28 (opens in a new window)

Brown GAITL Committee, July 2026Students dumbing down their prose to avoid accusations

The committee writes, “the fear of accusations of misuse of GenAI tools may be leading some students to alter their writing, intentionally dumbing down their language and even intentionally including language errors” (p. 18). Detection teaches writers to perform for a machine.

GAITL Committee Final Report and Recommendations (PDF) (opens in a new window)

Outcomes

Boston University English, proposed Sept. 15, 2026Writing generated without the assistance of AI

The outcome sets a condition and gives no reason for it. Outcomes should state the reasoning, because students need it in courses and jobs where no one has banned anything.

BU English, “Proposed Learning Outcomes Announced” (opens in a new window)

Three questions before the instructor or program decides

Ask these of each core assignment before anyone adopts use or refusal. They map onto the three levels.

  1. What the tools are. Which AI tools would a student reach for on this assignment, and what do students need to understand about how those tools produce text?
  2. Whom the tools disadvantage. Whose writing do these tools misread or penalize, such as multilingual writers or writers whose English differs from the standard, and what follows for design and assessment?
  3. When to use them, when to refuse. What decision should a student make about AI on this assignment, and what reasons tie using it, or declining it, to what the assignment is meant to teach?

Refusal becomes a reasoned decision a student learns to make. So does use.

Works cited

MLA 9. Preprints and working papers are marked as such.

  • Alfadel Coloma, Nadia. “UChicago Law Unveils New AI Strategy.” University of Chicago Law School, 9 July 2026, www.law.uchicago.edu/node/115258.
  • Austin, Anne K. “Proposed Learning Outcomes Announced.” Boston University, 15 Sept. 2026, www.bu.edu/english/2026/09/15/new-learning-outcomes/.
  • Brown University Generative AI in Teaching and Learning Committee. Generative AI in Teaching and Learning (GAITL) Committee Final Report and Recommendations. Office of the Provost, Brown University, July 2026.
  • Clark, Andy, and David Chalmers. “The Extended Mind.” Analysis, vol. 58, no. 1, 1998, pp. 7–19, doi.org/10.1093/analys/58.1.7.
  • Cole, Kirsti. “Cyborg Composing and Posthuman Research Methods: A Workshop Assignment for Writing Intensive Courses.” TextGenEd: Continuing Experiments, edited by Carly Schnitzler et al., WAC Clearinghouse, Aug. 2026.
  • Cole, Kirsti, Biven Alexander, Wil Carr, Brody McCurdy, and Bethany Van Scooter. “Reflective Multimodal Feedback Practices Across Writing Contexts.” TextGenEd: Continuing Experiments, edited by Carly Schnitzler et al., WAC Clearinghouse, Aug. 2025.
  • Emi, Bradley, and Max Spero. “Technical Report on the Pangram AI-Generated Text Classifier.” arXiv, 2024, arXiv:2402.14873.
  • Emig, Janet. “Writing as a Mode of Learning.” College Composition and Communication, vol. 28, no. 2, 1977, pp. 122–28, doi.org/10.2307/356095.
  • Goodlad, Lauren M. E. “Teaching Critical AI Literacies.” Google Doc, circulated by the Genuine Intelligence Project, AAUP North Carolina, Sept. 2026, docs.google.com/document/d/1xKMYHzy3JcBMs__qsTM2QKMzM3gAqoqVp739SJ6Pvp4.
  • Goodlad, Lauren M. E., and Sharon Stoerger. “Teaching Critical AI Literacies: Living Document.” Critical AI @ Rutgers, 2023–2026.
  • Hayles, N. Katherine. How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics. U of Chicago P, 1999.
  • Jabarian, Brian, and Alex Imas. “Artificial Writing and Automated Detection.” NBER Working Paper no. 34223, National Bureau of Economic Research, Sept. 2025.
  • Liang, Weixin, et al. “GPT Detectors Are Biased against Non-Native English Writers.” Patterns, vol. 4, no. 7, 2023, 100779, doi.org/10.1016/j.patter.2023.100779.
  • Liu, Grace, et al. “AI Assistance Reduces Persistence and Hurts Independent Performance.” arXiv, Apr. 2026, arXiv:2604.04721. Preprint.
  • Marcoccia, Chiara, Walter Quattrociocchi, and Valerio Capraro. “AI Advice Suppresses People’s Willingness to Say ‘I Don’t Know,’ Even When the Advice Is Wrong and Accuracy Is Incentivized.” arXiv, July 2026, arXiv:2607.13562. Preprint.
  • Massachusetts Institute of Technology. Report of MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training. MIT, 13 Aug. 2026.
  • Strömberg, David, Victor Lei, and Yanhui Wu. “The Generative AI Learning Penalty: Evidence from Chinese Secondary Education.” CEPR Discussion Paper no. 21577, Centre for Economic Policy Research, June 2026.

Kirsti Cole, Professor of English and Co-Director, Campus Writing and Speaking Program, NC State University. Contact kkcole2@ncsu.edu. Related: the AI Policy Lab workshop uses the same three-level framework.