How should middle- and high-school teachers assess students when AI is inevitable?

The assessment-design framework: Claim → Conditions → Evidence → Conversation → Judgment

Use this five-part sequence before creating a task. It works with paper, locked devices, shared documents, labs, performances, and projects.

Multiple kinds of evidence are already a sound assessment principle, not an AI-era novelty. Ontario’s K–12 framework, for example, describes evidence from observations, conversations, and student products. Ontario School Effectiveness Framework The U.S. Department of Education likewise frames formative assessment as evidence used to adapt teaching and learning. Artificial Intelligence and the Future of Teaching and Learning, p. 40

Choose the AI condition by learning target

Do not use one class-wide rule for every assignment. Mark each assessment with one of these conditions.

“AI off” is not “technology off.” If a student’s approved assistive technology is needed for access, remove only the tool capability that invalidates the particular learning claim; do not casually remove the accommodation. U.S. OCR notes that educational institutions must provide equal access to educational benefits and opportunities delivered digitally, and gives examples such as screen readers, captions, voice recognition, and keyboard access. U.S. Department of Education OCR, Technology Accessibility

Build an evidence bundle, not a surveillance system

For each major assessment, select the lightest combination that supports a valid judgment. A bundle need not mean recording every keystroke or collecting every chat.

A reliable default bundle for a major task

Revision history can be useful context: it may show when a student made a major change and give the teacher a concrete place to ask for explanation. It is not proof of authorship. A file can have incomplete history, be shared, be pasted into gradually, or be altered; keyboard-level monitoring creates privacy, equity, and workload costs. Treat a revision trail as a feedback and conversation aid, never as a lie detector.

A short AI-use disclosure students can actually complete

Assessment AI condition: Off / Bounded / Integrated

Tool(s), if any:

What I asked the tool to do:

What I used, changed, or rejected:

How I checked facts, quotations, data, code, or sources:

I can explain the decisions in this submission: Yes / I need help identifying a decision to explain.

The disclosure should reward honest, accurate documentation. It should not reward using AI, nor punish a student who did not use it when its use was optional. When AI provides information or sources, students must verify and cite the underlying sources; a tool’s citations are leads, not evidence by themselves. JCQ guidance on acknowledgement and verification

Assessment patterns that work across subjects

These are patterns, not scripts. Adapt the tool condition to the standard and age group.

Authentic tasks are not automatically AI-resistant. Their value is that they create consequential decisions—local evidence, an audience, constraints, feedback, testing, and revision—that students can explain. Pair authenticity with an individual checkpoint.

A usable 100-point rubric

Use the same categories across a unit where possible, then change the descriptors for the discipline. The weight of “responsible AI use” can be zero when AI is off and higher when AI use itself is an outcome.

Use four performance levels for each row:

Do not grade “human voice,” an expected writing style, compliance with a preferred drafting platform, or the absence of AI when AI was permitted. Grade the stated outcome and the quality of the evidence supporting it.

How to use oral defense without turning it into an accusation

An oral defense works best when it is routine, brief, and designed into the task—not triggered only when a teacher is suspicious. It can be an individual conference, a recorded explanation, a quick whiteboard response, or another comparable method.

For a project, ask the same small set of prompts of everyone or of a pre-announced rotating sample:

• “Show me the most important decision you made and the evidence behind it.”
• “What did you revise after feedback or testing, and why?”
• “If this source/data point/requirement changed, what would change in your conclusion or design?”

Publish the prompts or their categories beforehand. Give the student thinking time and an appropriate alternative mode of response. A student who communicates through AAC, needs processing time, has an anxiety-related accommodation, is still acquiring the classroom language, or has a different access need should not be forced into an impromptu spoken performance unrelated to the intended outcome. Flexible, barrier-aware ways to express learning align with the CAST UDL Guidelines 3.0, while approved accommodations and local policy remain controlling.

If evidence does not support a judgment, say that plainly: “I need another demonstration of this target.” Then offer a normal, documented reassessment: a supervised rewrite of one section, a new comparable problem, an annotated source explanation, or an accessible conference. Do not begin with “prove you did not use AI.”

Homework: make its role honest

Out-of-class work can still matter. What it cannot reliably do on its own is establish independent mastery of a skill that AI can complete.

This is not a reason to give up on homework; it is a reason to stop treating a take-home polished product as the only evidence of individual learning.

Equity, accessibility, privacy, and workload guardrails

Equity and access

• Do not require students to have a personal paid AI subscription, a home device, high-speed internet, or an account that their family cannot create. Offer a comparable no-AI route whenever AI use is optional.
• Keep the cognitive target separate from the method of response when the method is not itself the target. A handwritten essay may measure handwriting stamina, speed, and fine-motor access as well as argument; a rapid oral defense may measure verbal fluency and anxiety tolerance as well as historical thinking.
• Preserve accommodations. Design equivalent evidence, not identical performance: typed, spoken, signed, captioned, screen-reader-compatible, supported, or extended-time responses may all be appropriate depending on the target and plan.
• Teach disclosure explicitly with exemplars. Students unfamiliar with academic conventions should not lose points because “use AI responsibly” was never made concrete.

Privacy and child-safety boundary

Use only school/district-approved tools and accounts. Do not have students put names, grades, disability information, behavioural records, private family details, confidential case material, or identifiable peer work into a public AI service. In U.S. schools, the Department of Education advises teachers to check whether an application is approved and notes conditions applicable when personally identifiable information from education records is disclosed under FERPA’s school-official exception. U.S. Department of Education, student-privacy FAQ

This is not legal advice. Consult the school’s privacy officer/IT team and local requirements before adopting a new service, especially for minors. UNESCO likewise calls for a human-centred, age-appropriate approach and protection of data privacy in education. UNESCO, Guidance for Generative AI in Education and Research

False positives and fair response

Do not use an AI-writing detector to assign a zero, reduce a grade, or initiate discipline. A peer-reviewed 2023 study of seven detectors found a mean false-positive rate of 61.3% for the study’s set of 91 TOEFL essays written by non-native English writers. Liang et al., “GPT detectors are biased against non-native English writers,” Patterns That result does not mean every detector, language, or task has the same error rate; it is enough to show why a detector score cannot establish misconduct fairly.

If a tool flags work, or the work seems inconsistent with available evidence, treat it as a prompt to gather neutral, learning-focused evidence. Ask for a normal conference or comparable additional demonstration, document the evidence and the student’s explanation, and follow the school’s process. Never try to infer guilt from accent, dialect, disability, multilingualism, neatness, vocabulary, typing patterns, or a “vibe.”

Teacher workload

The framework should replace some grading, not create an extra bureaucracy.

• Use an in-class checkpoint for all students, then sample 6–8 short conferences per period on a rotating schedule.
• Score one meaningful revision note rather than every draft.
• Use a common three-question defense bank for a whole unit.
• Ask teams to create a shared product but collect one individual evidence item per student.
• Reallocate points: grade fewer polished take-home products deeply; use low-stakes homework for practice and feedback.
• Keep the disclosure form short. Full chat logs are usually more work, more private, and less instructionally useful than a student’s concise account of decisions.

The risk-management principle is proportionality: select controls that fit the stakes and context, then revisit them as the tool and harms change. NIST, Generative AI Profile (AI 600-1)

A two-week implementation plan

Week 1: establish the routine

1. Audit the next three assessments. For each, write the independent learning claim in one sentence. Flag any task whose only evidence is a take-home final product.
2. Assign an AI condition. Label the task “AI off,” “AI bounded,” or “AI integrated,” and list permitted tools, prohibited uses, accommodations, and disclosure expectations.
3. Add two evidence points. For the next major task, add an in-class plan/annotation/worked-example checkpoint and one brief individual response.
4. Show examples. Walk through one acceptable disclosure, one inappropriate undisclosed use, and one AI-assisted response whose sources must be checked.

Week 2: refine without escalating surveillance

1. Run the task and conference. Ask the same three questions of a rotating sample or everyone when the task is high stakes; offer accessible alternatives in advance.
2. Calibrate the rubric. Mark three anonymized exemplars with colleagues or students. Check whether the rubric accidentally rewards polish over reasoning or disadvantages a response mode unrelated to the target.
3. Use results for teaching. If students cannot explain their final product, reteach or reassess the specific concept; do not treat this automatically as a conduct case.
4. Update one policy sentence. Keep what produced useful evidence. Remove a requirement that did not change instructional decisions.

Assignment language a teacher can adapt

Learning target: I can make and defend a [discipline-specific] claim using [specified evidence/skill].

AI condition: Bounded. You may use an approved AI tool only after completing the in-class evidence plan. You may ask for feedback on clarity or counterarguments. You may not ask it to write the final submission, invent sources/data, or replace your analysis. Verify every claim and cite the source you actually used.

Submit: your in-class plan, final product, two revision notes, and the short AI-use disclosure. Be ready to explain one decision and respond to one small change in the evidence.

Access: Use approved accommodations and contact me before the checkpoint if you need a comparable way to demonstrate the target.

Common failure modes—and better alternatives

Viable alternatives when this framework does not fit

• Short, frequent supervised checks: Best for foundational fluency, computation, reading comprehension, and writing-to-time tasks. Paper is one option; a locked digital environment may be more accessible.
• Portfolio review: Best for long-form creative, research, design, or performance work. Use selected artifacts, reflection, and a final conference rather than every version.
• Open-resource oral or written application: Best when recall is less important than judgment. Permit sources and ask students to apply them to a novel case.
• Performance task with individual accountability: Best for labs, group projects, and career/technical courses. Observe teams, then collect a brief individual explanation or change request.
• AI critique task: Best when tool literacy is a stated outcome. Give all students the same AI output, then ask them to fact-check, annotate limitations, revise it, and explain their choices.

None of these makes deception impossible. The goal is a defensible, humane, instructionally useful judgment—one that does not depend on a teacher winning a technological cat-and-mouse game.

Limitations and what may change

The framework does not prove authorship, guarantee that AI use will be disclosed, or replace the safeguards required for formal examinations. It also requires time to teach routines and to align accommodations. Its central trade-off is deliberate: gather a little more direct evidence of understanding rather than spend more time trying to infer provenance from a finished file.

AI tools, student access, school policies, vendor terms, and assessment rules change quickly. Review the AI conditions, approved-tool list, data-sharing practice, and detector policy at least each term. The U.S. Department of Education’s AI report stresses that educational technology needs human-centred, equitable design and attention to context; it should be read as guidance, not as a product endorsement. U.S. Department of Education, Artificial Intelligence and the Future of Teaching and Learning