Take-home exams, AI, and the Question beneath the Question

Lia Daniels

Take-home exams have long occupied an uneasy middle ground in higher education — less artificial than timed, in-person tests, yet more vulnerable than tightly proctored assessments. The rise of generative AI has intensified this unease, prompting urgent questions about integrity, authenticity, and fairness. But as our interview with CTL scholar Lia Daniels makes clear, the current debate is not really about a single assessment format. It is about confidence, specifically, instructor confidence that student work completed outside the classroom still functions as credible evidence of learning.

That erosion of confidence is what feels destabilizing.

From an assessment perspective, AI acts less like a rupture and more like a spotlight. It exposes vulnerabilities that were already present in unsupervised assessment. Daniels reminds us that students have always been able to consult peers, family members, notes, calculators, or online resources. “Instructors have been comfortable with that level of risk, but now we’re not sure. We might have designed the assessment so that talking to a peer was okay, or so that getting a neighbour's opinion was developing thinking. But then we trusted that students would sit down and write the essay or take the test on their own. We didn’t mind too much that ambiguity because the student was still expected to produce the work themselves, and if they didn’t, that was a clear breach of academic integrity.” 

Daniels emphasizes that the problem is not simply access to help, but how easy it has become for AI to complete an entire task: “Now I think we as instructors are wrestling with a space in which we don’t have the same confidence in the assessment to be an indicator of student learning. Many instructors are no longer comfortable with take-home exams or take-home assignments because it's just really easy to have AI do the whole thing.” Where previous supports nudged or scaffolded student thinking, generative AI can now generate full responses, essays, solutions, explanations, and even self-checks. 

She also points out that AI is increasingly unavoidable. Some search engines now surface AI-generated summaries by default, and writing tools routinely offer automated suggestions. Even students who believe they are playing by the rules may still be influenced by AI embedded in everyday tools. This normalization makes simple prohibitions—“don’t use AI”—far harder to enforce in practice: “I think it is hard to say, I will just tell them not to do it, and they won't do it, and that will work.” In this context, the question shifts from enforcement to intention: what does meaningful learning look like when AI is already woven into the tools students use every day?

A common argument suggests that if AI can complete a take-home exam, then the exam must never have measured meaningful learning. Daniels explicitly rejects this claim. 

She argues that many learning outcomes—such as solving mathematical equations or explaining disciplinary concepts—remain valid, even if AI can now perform them well. A take-home exam may have aligned perfectly with such outcomes prior to AI, and nothing about the outcome itself has inherently changed.

The tension arises because AI can now replicate the process that instructors once trusted students to perform. In this moment, Daniels suggests, instructors face a real decision point: “You change the learner outcome, or you change what the assessment looks like. That's where you sort of land.” 

Beyond logistics, AI raises deeper questions about originality and independence.

Daniels is unequivocal that AI cannot hold authorship: “AI can't be held accountable. I think at the core, anyone who uses AI for any academic purpose has to accept accountability for the work that is produced. That's just to me non-negotiable.” 

At the same time, Daniels expresses ambivalence about assessing students without support. Drawing an analogy to anxiety accommodations, she questions whether removing tools reveals something more authentic, or simply creates artificial conditions that no longer reflect how learning and thinking actually occur: “there’s part of me that thinks there is a time and a place where the way in which students develop their ideas in cooperation with an AI is what it is, not something that functions as an extra or an add-on.”

Daniels does not offer easy solutions, but she does offer grounded guidance for instructors who may want to keep their take-home exams. 

“If you're keeping a multiple-choice take-home exam, you're going to want to use all of the proctoring services, all the lockdowns, all the exam security, and really enhance your commitment and conversations around the Code of Conduct.” Having students actively check a box and agree not to use AI creates a psychological tension that may discourage misuse, even if it cannot eliminate it: “there is something very psychologically compelling about agreeing to something and then intentionally doing something different,” says Daniels. 

For open-ended or long-answer exams, she suggests either requiring students to affirm non-use of AI or asking them to explain how AI was used: “ If students have permission to use AI, adding a requirement for students to describe how it was used in completing this particular take-home is a good idea.” While instructors may not always audit AI outputs, these practices help establish transparent norms.

In both cases, Daniels suggests that instructors may also want to reconsider the weighting for at-home assessments. Returning to confidence, she urges instructors to reflect: “Am I comfortable giving 30% of students’ grades to AI if they decide to use it to basically do the work?” If the answer is no, another option is to consider reducing the weight of the assessment. 

Crucially, Daniels cautions against designing assessments around what AI “can’t do yet.” That window is closing rapidly.

She concludes soberly yet clearly: there is no guaranteed AI-free take-home assessment. Instead, instructors must decide where their own boundaries lie: “we both have to build up our understanding of responsible use  and also have our own sense of where the boundary is—where we’re like, ‘This is too uncomfortable - I don’t know if it still represents students’ learning.’ It’s really tricky.”

From our conversation with Lia Daniels, it becomes clear that what AI ultimately demands is not perfect control but pedagogical clarity.