Define the writing goal before choosing a tool

The first decision is not which chatbot to use, but what learning students should demonstrate. An activity might assess how students develop a thesis, use evidence, organize an argument, revise for a particular audience, or apply the conventions of a discipline. Without a clear goal, allowing or prohibiting a tool can produce a rule that is hard to justify and an assessment that reveals little.

Separate the final product from the process that produces it. If a course aims to help students practice synthesizing readings or drafting paragraphs, outsourcing those tasks can obscure whether they have learned them. If the goal is to critically review proposals or adapt an argument for different audiences, an activity can use AI as material for comparison, provided students must analyze and justify their decisions.

Cornell’s guidance presents flexible policies tailored to an activity: prohibit AI when it would interfere with foundational learning, allow it with attribution when it may be useful, or integrate it deliberately when it supports higher-level goals. This is a framework to adapt, not a universal scale or an institutional rule for other universities. Instructors can align it with learning outcomes and local academic policies.

Before writing the assignment, ask a specific question: what should students be able to do without assistance for the course to consider this goal achieved? The answer helps distinguish between a skill students need to practice themselves and an operation they might partially delegate to focus on a different cognitive task.

Decide what uses are allowed and communicate rules for each activity

A useful policy explains what is allowed, for what purpose, and with what responsibility. Instead of a general statement such as “AI is allowed,” specify whether students may use it to generate ideas, suggest an outline, comment on a draft, correct grammar, or translate for reference. Also state whether the tool may draft text that students submit as their own. Permission for one function does not imply permission for all others.

UC Berkeley’s teaching guidance recommends communicating expectations clearly and connecting them to pedagogical goals. Its examples distinguish between policies that integrate tools, permit limited uses, or prohibit them for particular assignments. These are institutional examples, not rules that automatically apply at other universities. Instructors should check their institution’s academic integrity policy and explain how it applies to each assessment.

A brief instruction can include four elements:

  • Permitted use: for example, asking for suggestions to improve the clarity of a student’s own draft.
  • Prohibited use: for example, asking a system to write the argument being assessed.
  • Disclosure: state which tool was used and for what purpose, in line with course rules.
  • Responsibility: check claims, references, and changes before submitting the work.

Transparency also gives students a chance to ask before they act. If different assignments have different rules, repeat the relevant condition in each assignment rather than relying only on the syllabus. Explain the reason in terms of learning: a restriction on initial drafting may protect practice in formulating ideas, while a revision activity may allow automated feedback for students to evaluate.

Design activities that make the writing process visible

An activity involving AI can assess learning when students must compare, verify, and decide, rather than simply accept a fluent answer. In a revision exercise, for example, each student could first write a paragraph, then ask a tool for feedback on its structure, and submit a revised version with an explanation of which suggestions they accepted or rejected.

A sequence can keep the focus on the writer’s decisions:

  1. Draft a thesis and a paragraph supported by a course reading.
  2. Ask a tool for feedback limited to clarity, organization, or the connection between a claim and its evidence.
  3. Check each comment against the assignment, the source text, and the paragraph’s purpose.
  4. Revise the text and identify what changed, what did not, and why.

This is a proposed activity, not evidence that AI improves writing on its own. Cornell’s report on writing courses describes options such as comparing revisions, assessing a tool’s feedback, or adapting a text for a new audience. It also warns that passive dependence can weaken practice of foundational skills and lead to unsuitable suggestions. The pedagogical value lies in the analysis and revision students are asked to do.

Another option is to require an initial version written without AI, followed by a second version that permits a specific function, such as generating an alternative outline. Students can note what changed between the versions and defend their chosen structure with reference to the assignment. This gives instructors evidence of reasoning and revision without assuming that a tool’s history captures the entire writing process.

Design a writing activity with generative AI

A practical sequence for aligning AI use with learning goals and making students’ decisions visible.

  1. 1. Define the goal

    Name the skill and decide what students must practice themselves.

  2. 2. Explain the rules

    State what AI may do, what it may not do, and what students must disclose.

  3. 3. Show the reasoning

    Ask students to assess feedback and justify what they change or keep.

  4. 4. Assess learning

    Use announced criteria; do not treat detector results as proof of authorship.

  5. 5. Review safeguards

    Consider privacy, accessibility, and whether students can access the tool.

University writing assignments

Adapt the activity and rules to course objectives and institutional policies.

Sources and references5

Assess reasoning without treating detectors as proof

Assessment should reflect the stated learning goal, not a guess about who wrote each sentence. A final text can conceal the decisions and difficulties behind it, so it can be useful to distribute grades across argument, use of evidence, revision, and reflection when these are course objectives. Staged submissions, comments on sources, and a brief explanation of changes can make some of the process visible through explicit criteria.

UC Berkeley Research, Teaching, and Learning reports that the results of an AI-detection pilot were inconsistent and that the institution decided not to enable the feature campus-wide. It also advises against relying on detectors to identify AI use in writing. The practical takeaway is that a detector score does not prove authorship and should not replace clear assessment criteria or applicable institutional procedures.

A rubric can assess verifiable aspects of a submission: whether the thesis answers the question, whether cited sources support the claims, whether the organization serves the intended audience, and whether revisions address identified problems. If the course assesses students’ ability to explain their decisions, a brief reflection or academic conversation can complement the text. These should use announced, consistent criteria—not be improvised as a response to suspicion.

Instructors should not treat every stylistic difference as conclusive evidence of improper use. Assessing learning and handling a possible academic integrity violation are related, but distinct, matters.

Evaluar el razonamiento sin convertir detectores en pruebas La evaluación debe corresponder al aprendizaje declarado, no a una conjetura sobre quién escribió cada frase. Un texto f...
Editorial illustration by BytePreview

Verify generated information and sources

A generative model can produce convincing prose that is incorrect. Cornell warns that these systems can present false claims with apparent confidence, so a bibliography suggested by a tool is not a verified source. In academic work, every reference should be locatable, and each significant claim should be checked against the original document or another relevant source.

Build verification into the assignment rather than leaving it as a general warning. For example, ask students to check a suggested reference in the library catalog, open the original text, and note what claim it actually supports. If a source cannot be found, it should not be kept as if it existed. If a source exists but does not support the claim attributed to it, students should revise the argument and explain the discrepancy.

For a source-based writing exercise, students could use a working table with three columns: the draft’s claim, the passage or data checked in the source, and the revision decision. This table is a learning tool; it does not automatically make every interpretation objective. Students still need to explain why the evidence is relevant and what its limitations are.

UNESCO’s guidance takes a human-centered approach and addresses ethical risks, equity, inclusion, and pedagogical design in education and research. In practice, instructors can make AI literacy part of an assignment where appropriate: teach students to question results, identify errors, and use human judgment about what to claim and what evidence supports it.

Privacy, attribution, accessibility, and course preparation

Do not enter personal data or identifiable student work into an external service without reviewing the tool’s terms and the institution’s policies. UC Berkeley notes that sending student work to third-party detectors can raise privacy and copyright concerns; its guidance recommends assessing security, privacy, and accessibility before adopting tools. The specifics depend on the service, its settings, and institutional agreements, so do not assume a platform protects data without checking.

Attribution should follow course instructions and disciplinary norms. When AI use is permitted, instructors can ask students to identify the system and describe its function—for example, style feedback or outline generation—in addition to citing conventional sources. Disclosure does not replace checking references or transfer responsibility for submitted content to the tool.

Accessibility requires a practical check before making a tool mandatory. Consider whether all students can access it, whether an equivalent alternative is available, and whether its format works with relevant accessibility needs. Cornell and Berkeley include accessibility among the considerations for instructors; that guidance does not guarantee that any particular tool is accessible to every student.

Before publishing the activity, review this checklist:

  • Is the writing goal stated as an observable skill?
  • Does the assignment distinguish permitted, restricted, and prohibited uses?
  • Does it explain the required attribution and who is responsible for the result?
  • Must students verify sources and evaluate the suggestions they receive?
  • Does it avoid entering sensitive information into unapproved services?
  • Is there a reasonable alternative for students who cannot use the tool?
  • Does the rubric assess learning rather than an automated detection score?

A practical starting point is to choose one activity, state what learning it should reveal, and test a specific rule students can understand before they begin. Teaching resources from Cornell, UC Berkeley, and UNESCO offer frameworks and examples to inform that decision; each institution’s academic integrity, privacy, and accessibility policies still govern the course.