From Assignment to Assurance
Student assessment in a GenAI-infused environment has to answer a question that a polished assignment cannot answer by itself: what can this student understand, explain, judge, and do when the system is no longer carrying the central work? This final chapter brings together the performance-learning distinction, Evidence-Centred Design, authentic practice, conversational observation, and human oversight to give that question a practical form. The goal is not to make assessment AI-proof. It is to make the claim behind a grade defensible.
The unit of certification is the human capability
When AI can produce an essay, a calculation, or a working program, the final object is still part of the evidence. It may show communication, technical execution, or the quality of a solution. But it no longer supports a broad conclusion about the student on its own. The relevant unit of certification is the capability the course or qualification promises: forming an argument, applying a concept, evaluating evidence, debugging a method, collaborating, communicating, or using a tool responsibly.
That distinction gives a precise answer to the opening problem. If AI can produce the assignment, certify the human capability through observations that require the student to make meaningful judgments and demonstrate what remains after assistance. This does not mean that an AI-assisted product is illegitimate. A student may be expected to work with AI in a professional setting. It means the assessment must show whether the student selected the tool appropriately, checked its output, understood the result, adapted it to the situation, and can perform the relevant reasoning at the level the qualification claims.
Evidence-Centred Design makes the logic explicit. Begin with a claim about the learner. Design a task that can elicit the claim under stated conditions. Decide which observations would support the inference. Then record the boundary: what the evidence does not establish. A student who submits a correct diagnosis has provided product evidence. A student who asks relevant questions, explains why those questions matter, and revises a provisional conclusion has provided process and explanation evidence. A later case that requires the same reasoning in a new context provides evidence about transfer. These observations are not interchangeable, but together they can support a more credible judgment.
The same architecture applies to an AI-assisted writing task. The product may demonstrate a finished argument. The process may show how the student selected sources, reacted to criticism, or changed the thesis. Explanation may show whether the student can defend the central reasoning and identify weaknesses. A transfer task may show whether the capability survives outside the exact prompt. The assessment does not need to record every prompt or keystroke. It needs enough distinct evidence to resolve the ambiguity created when a system could have supplied the target work.
This is also why authenticity and validity must be designed together. Authentic practice may include documentation, search, collaboration, and AI. Removing those tools can make an assessment less like the work students are preparing to do. But realistic conditions do not excuse an institution from observing the student’s contribution. A workplace-like project with an AI-generated solution can be authentic as a simulation and still weak as evidence of the student’s reasoning. Authenticity supplies relevance; process and performance evidence supply the basis for the learner inference.
The conversational exam offers one workable expression of this principle. Students in an engineering-computations course code in a familiar notebook environment while explaining their reasoning. Prepared questions are linked to learning objectives, and hints, simplified tasks, or elevated concept checks give the examiner structured ways to probe understanding. A rubric separates technical execution, conceptual understanding, and problem solving or communication. The format does not derive validity from proving that students never used a tool. It derives it from requiring a live demonstration and an explanation that a submitted file could not provide.
Yet the case also changes the meaning of “human presence.” Presence is not simply placing an instructor in the room. It is the disciplined interpretation of evidence. The examiner has to know what the task claims, what a response demonstrates, what a hint changes, and when an error is meaningful or incidental.