Beyond AI-Proofing: Rethinking What Counts as Evidence of Learning

Synthesis: Bayraktar (2026) argues that higher education has framed generative AI as an authenticity problem — can this artifact be proven to be the student's own work? — when the more productive question is a validity question: what evidence warrants a claim about what a student knows and can do? The article proposes an AI-Responsive Assessment Framework with four interlocking principles: judgment-centered tasks, process visibility, metacognitive transparency, and iterative human feedback. The aim is not to replace products with process portfolios but to broaden the evidentiary basis so that a polished essay or correct analysis is interpreted alongside evidence of the reasoning that produced it. Rigor, on this account, lies in disciplinary judgment rather than in independent production, and academic integrity becomes intellectual accountability for decisions rather than proof of unassisted authorship.

Key Findings

1. Artifact authenticity is a narrow frame. Detection, bans and in-class rewrites answer whether a product was produced independently; they do not answer whether the student understands the construct the product is supposed to evidence.
2. The four principles supply complementary evidence. Judgment-centered tasks elicit disciplinary decision-making, process visibility shows how reasoning developed, metacognitive transparency makes strategic choices explainable, and iterative human feedback shows whether students can recalibrate judgment against standards.
3. Products remain evidence. The framework does not discard finished work; it treats the artifact as one source among several and asks which additional evidence the construct actually requires.
4. Rigor moves from production conditions to evaluative judgment. The capacity to interpret criteria, compare alternatives, decide and justify is a better account of rigor than the absence of assistance, and it is the capacity AI-mediated work makes most consequential.
5. Implementation is a design and workload problem. Eliciting more evidence costs instructor time, especially in large courses, and centers for teaching and learning are better placed to help faculty translate the principles into discipline-appropriate designs than to focus on tools or policy alone.

Why artifact authenticity is the wrong frame

The article begins with the AI-proofing moment: institutions responding to generative AI with detection software, revised integrity policies and assignments redesigned to prevent AI use. Bayraktar does not deny that authorship matters, but argues that when authenticity becomes the organizing question, assessment narrows to a verification exercise. The interesting question — whether the evidence collected supports the inference the instructor wants to make — drops out.

The alternative is to treat the problem as one of construct validity. An assessment is valid when the evidence it produces warrants the claim being made. If the claim is "this student can construct and defend an argument," a polished essay produced with AI may be weak evidence; a revision portfolio with explanations of what changed and why may be stronger. If the claim is "this student can perform a procedure independently," then independent performance is the construct and should be assessed directly. The framework is therefore a design heuristic, not a mandate to add portfolios to every course.

The four principles

Judgment-centered tasks ask students to apply disciplinary criteria, choose among plausible approaches, and justify decisions. The evidence lies less in which option was chosen than in the reasoning that made it defensible.

Process visibility supplements the artifact with evidence of how it developed: evolving drafts, explanations of substantive revisions, records of alternatives considered, responses to feedback. The question shifts from "can the learner prove independent production?" to "what evidence does the learner provide of the thinking through which understanding developed?"

Metacognitive transparency asks students to explain how they approached the task, where AI contributed, where its output was inadequate, and why suggestions were accepted, modified or rejected. Disclosure alone establishes provenance; explanation shows whether the student can monitor and evaluate their own thinking.

Iterative human feedback treats feedback as an evidentiary opportunity as well as an instructional one. How a student interprets comments, decides what to act on and justifies departures from advice reveals developing judgment in a way that fluent AI feedback cannot.

Three disciplinary illustrations

In advanced writing, a researched argument assignment adds a revision portfolio: each draft carries a short account of what changed, what was rejected and why, in relation to rhetorical criteria discussed in class, plus a conference response to instructor feedback. In clinical social work, a case conceptualization is accompanied by a process record of frameworks considered and rejected, points of interpretive uncertainty, and an examination of where an AI tool's framing converged with or diverged from the student's clinical reasoning, followed by supervision dialogue. In graduate research methods, an empirical project adds an analytic decision log recording operationalization choices, approaches considered, missing-data handling, and annotations of any AI-generated code explaining what each step does and where it was verified or corrected.

The three cases are chosen to show that the same principles translate differently. What counts as judgment, process and feedback in composition is not what counts in clinical supervision or statistical analysis, which is why the framework asks for disciplinary translation rather than wholesale adoption.

What this means for practice

• Start from the construct, not the countermeasure. Ask what learning the assessment is meant to evidence and whether the current evidence supports that inference before redesigning around AI.
• Add evidence selectively. A single added artifact — a decision log, a short revision rationale, a conference response — can close an evidentiary gap without turning every course into a portfolio course.
• Make AI use an object of judgment. Asking students to evaluate where a tool's output was wrong or unhelpful produces better evidence than a disclosure statement.
• Treat feedback as data. What students do with comments is evidence of evaluative judgment; design at least one occasion where that response is visible.
• Resource the implementation. The framework increases interpretive work; without teaching-center support and workload recognition it will be adopted rhetorically and abandoned in practice.

Limitations

• The article is a conceptual framework, not an empirical study; it offers no data on whether the proposed evidence improves learning or agreement among raters.
• The three illustrations are constructed examples, not implementations with students, and the article does not report feasibility, cost or student response.
• The framework depends on instructors' disciplinary judgment about which evidence is sufficient, and it offers no threshold or decision rule for when a product alone is adequate.
• Added evidence raises privacy and equity questions — process records and metacognitive accounts can disadvantage students with less time or different cultural expectations of self-disclosure — that the article acknowledges only in passing.
• Whether students can produce good metacognitive accounts without coaching, and whether such accounts can be gamed, remains untested.

Citation

Bayraktar, B. (2026). Beyond AI-Proofing: Rethinking What Counts as Evidence of Learning. The Interdisciplinary Journal for Research on Pedagogy and Learning Design, 1(1).