When AI Detectors Fail: Designing More Honest Assessments in Higher Education

The question many educators have been asking is deceptively simple:

Did this student use AI to complete the assignment?

In August 2026, that question is becoming less useful: not because academic integrity no longer matters, but because AI detectors cannot provide the certainty that high-stakes decisions require.

A recent Inside Higher Ed report describes a growing number of institutions prohibiting or discouraging reliance on AI detectors. The reasons include false positives, bias against some writers, inconsistent results, and an increasingly familiar “cat and mouse” cycle in which AI-generated text is modified to evade detection.

The emerging alternative is not to ignore AI use. It is to design assessments that make learning more visible.

That shift has implications for faculty, instructional designers, educational developers, and student-support staff. It asks institutions to move from surveillance toward clarity, process visibility, critical AI literacy, and shared responsibility.

The problem with treating a detector score as evidence

AI detectors produce estimates. They do not provide a verified account of how a student completed an assignment.

That distinction matters. A detector may flag writing because of its style, structure, or word choices. It may fail to identify AI-assisted writing that has been substantially revised. It may also perform differently for multilingual writers, students with different linguistic patterns, or assignments that require highly conventional technical language.

A score can therefore become an accusation without establishing what actually happened.

This creates several risks:

• A student’s work may be questioned based on an opaque output they cannot meaningfully challenge.

• Faculty may be asked to interpret a tool they did not select and cannot validate.

• Students may become more anxious about sounding “too polished” or “too unusual.”

• Time and attention may be diverted from teaching toward forensic investigation.

• Institutions may unintentionally weaken trust with the very learners they are trying to support.

The California Community Colleges’ August 2026 memorandum on building local AI frameworks reflects this broader direction. It encourages colleges to focus on meaningful evidence of learning, while preserving faculty judgment and local decision-making. The memorandum specifically points toward assessments that ask students to explain reasoning, document processes, defend conclusions, verify claims, and apply knowledge in authentic contexts.

That is an important distinction: the goal is not to prove that students did not use AI. The goal is to assess whether students achieved the intended learning outcomes.

What more honest assessment can look like

Not every assignment needs to be rebuilt from the ground up. A more sustainable approach is to identify selected assessments where the current format does not provide enough evidence of learning, then redesign those tasks with greater process visibility.

Here are four practical moves educators can make now.

1. Clarify permitted AI use at the assignment level

A general syllabus statement is rarely enough. Students need to know what AI use means for this assignment, in this course, at this stage of learning.

For each major assessment, clarify:

• Whether AI tools may be used for brainstorming, outlining, translation, coding, accessibility, revision, or feedback.

• Which parts of the task must represent the student’s own analysis, judgment, performance, or reflection.

• Whether AI use must be disclosed, cited, or documented.

• Whether students may use a particular tool: or whether they should not be required to use one.

• What students should do if they are uncertain about whether a use is permitted.

The CCCCO guidance emphasizes that institutional frameworks should establish guardrails while faculty retain responsibility for course- and assignment-level expectations. That balance is important. A college can provide shared language and privacy standards without imposing one universal rule on every discipline.

A useful policy is specific enough to guide action and flexible enough to reflect the learning outcome.

For example:

AI may be used to generate possible questions and identify areas for further research. Your final argument, source evaluation, interpretation, and reflection must be your own. Include a brief process statement describing whether and how you used AI, what you verified, and which suggestions you rejected.

This is clearer: and more educational: than simply writing “AI prohibited” or “AI allowed.”

2. Redesign selected assessments to reveal reasoning and process

A polished final product is only one possible form of evidence. Add small, purposeful checkpoints that help students show how they arrived at their work.

Depending on the discipline, this might include:

• A proposal or problem-framing memo.

• An annotated bibliography or source-verification log.

• Drafts with feedback and revision notes.

• A short explanation of a method, calculation, design choice, or coding decision.

• A reflection on what changed between drafts.

• A brief oral walkthrough or demonstration.

• A critique of an AI-generated response, including errors, omissions, bias, or unsupported claims.

• A portfolio showing development over time.

These elements do not need to become a second, hidden assignment. They can be brief and aligned with the existing rubric.

For instance, a research paper might be assessed through the paper itself, a one-page source-and-decision log, and a 300-word revision reflection. A software project might include a code walkthrough in which the student explains one design choice and one debugging decision. A public speaking assignment might ask students to evaluate how AI affected audience analysis, message development, or ethical communication.

The point is not to force students to produce evidence for surveillance. It is to assess the thinking the assignment was intended to develop.

3. Invite students into norm-setting

Students are not only recipients of AI policy. They are also living with its consequences, testing its boundaries, and developing professional habits in real time.

A recent case study in Compass: Journal of Learning and Teaching in Higher Education describes students working as co-designers and co-evaluators of an AI literacy framework. Their participation helped staff understand student perspectives while giving students meaningful experience in research, collaboration, communication, and leadership.

Student consultation can be scaled to fit your context. You might:

• Ask students what language in an AI policy is confusing or difficult to interpret.

• Discuss examples of assistance versus substitution.

• Invite students to identify privacy or access concerns.

• Co-create a short set of class norms.

• Ask students to review whether a process requirement is proportionate and accessible.

• Revisit the norms after an assignment and discuss what worked.

This does not mean every classroom decision becomes a vote. Faculty still have responsibility for learning outcomes and assessment. But participation can make expectations more understandable, realistic, and trusted.

4. Teach critical AI literacy: not just tool use

AI literacy is broader than knowing how to write a prompt.

Students need opportunities to examine how AI systems produce outputs, where those outputs can fail, how bias and power shape automated systems, and what information should not be entered into a tool. They also need to understand attribution, verification, privacy, accessibility, and professional responsibility.

This work is already moving into curriculum design. Penn State’s open educational resource, Beyond the Podium: AI, Speech, and Civic Voice, integrates AI literacy into public speaking through questions of ethical communication, misinformation, authorship, attribution, and audience engagement.

Cornell has also expanded its AI Critical Literacy Program for incoming students. Its modules address what generative AI is, ethical questions, AI and learning, and personal AI-use policies.

These examples suggest a useful principle: AI literacy should not be isolated as a one-time orientation. It can be embedded in the practices students are already learning.

A biology course might ask students to verify an AI-generated explanation against primary sources. A nursing program might examine when AI-supported recommendations require human review. A communications course might compare AI-generated claims with audience needs and ethical obligations. A library session might focus on source evaluation, privacy, and information integrity.

The goal is not to make students suspicious of every tool. It is to help them develop judgment.

Protect privacy, access, and faculty autonomy

Assessment redesign should not create new inequities.

If AI use is required, students need equitable access to an institutionally supported tool. Requiring personal subscriptions or consumer platforms can introduce cost, privacy, and data-protection concerns. The CCCCO memorandum also emphasizes accessibility review, Universal Design for Learning, student rights, and human oversight.

Faculty should be supported rather than left to redesign every assignment alone. Institutions can provide:

• Shared examples and adaptable assignment templates.

• Time for faculty learning communities.

• Instructional design and educational development support.

• Clear processes for reviewing tools for privacy and accessibility.

• Student-support guidance for questions and appeals.

• Opportunities to pilot and evaluate changes before scaling them.

This is where a care-centered approach matters. If the only institutional response is “use a detector” or “redesign everything,” educators and students are both placed under unnecessary pressure.

A better question is: What is the smallest meaningful change that will give us better evidence of learning?

A practical starting point for this term

If you are revising a course now, begin with one assessment.

1. Identify the learning outcome it is meant to measure.

2. Ask whether the current submission provides credible evidence of that outcome.

3. Write clear, assignment-specific AI guidance.

4. Add one process artifact, such as a draft, decision log, reflection, or walkthrough.

5. Invite students to discuss the expectations before submitting.

6. Review what you learned and adjust the assessment next time.

You do not need to eliminate every risk. You need an assessment system that is transparent, proportionate, and connected to learning.

Designing the next step with support

The move away from unreliable detection is not a move away from rigor. It is an opportunity to make rigor more visible, more equitable, and more educationally meaningful.

At The Open Practice Academy, Custom Learning Pathways can support faculty and teams as they develop AI literacy, redesign learning experiences, and help people adopt new practices under real institutional constraints. For a focused need: such as revising an assessment, clarifying AI-use guidance, or building a faculty workshop: Specialist Support may be the right place to begin.

AI will continue to change. Your assessment principles can remain steady: clarity, evidence, human judgment, accessibility, and care.