For University Lecturers and Academics

Assessment and Research in the Age of AI: A Guide for University Lecturers

Your assessment methods were designed when student work meant student thinking. Now ChatGPT and Claude produce submissions that look credible but reveal nothing about what your students actually understand. You face a choice: redesign what you measure, or watch your degrees certify outputs rather than judgement.

These are suggestions. Your situation will differ. Use what is useful.

Download printable PDF

Redesign Assessment to Reveal Thinking, Not Output

Multiple choice and essay submission no longer show you who thinks well. You need assessment that forces students to show their reasoning in real time and in ways AI cannot easily simulate. In-class problem solving, live critique of their own work, and explanation of their choices under conditions you control will tell you what they actually know. This is harder to mark than checking an essay, but it is the only way to know if your degree means anything.

›Replace take-home essays with supervised problem sets where students must explain each choice they make
›Use viva voce or recorded defence of work where students justify specific claims from their submission
›Set assessment tasks that require students to evaluate contradictory sources and state which they trust and why, then defend that judgement verbally

Teach Research Methodology as Critique of AI Outputs

Your students will use Elicit, Semantic Scholar AI, and Perplexity for literature review. Stop pretending they will not. Instead, teach them to spot what these tools miss. Show them how AI summarises papers without reading them carefully, how it can smooth over real disagreement between researchers, and how it misses the methodological detail that matters. Make identifying these failures part of the research grade. A literature review that catches where the AI tool missed something valuable demonstrates real research thinking.

›Have students cross-check AI-generated summaries of key papers against the actual text and report discrepancies they find
›Require students to identify at least one limitation or gap in an AI-assisted review and explain why it matters to their research question
›Compare two different AI tools on the same research question and analyse where their outputs diverge and what that tells you about their reliability

Create Checkpoints That Interrupt the Outsourcing of Thinking

When students can paste their entire assignment into Claude, the temptation to let the tool do the thinking becomes very strong. You prevent this by building checkpoints into longer work. Require submission of annotated notes before the essay, a draft outline with your feedback before the final version, and a short explanation of how they changed their thinking in response to feedback. These checkpoints are not about policing use of AI tools. They are about making it impossible to outsource the intellectual struggle that builds competence.

›Request a one-page memo explaining what the student initially thought about a topic and what they changed their mind about, with evidence of this shift in their final work
›Require submission of raw research notes or marked-up sources before accepting an essay, so you see what sources they actually engaged with
›Set a mid-process meeting where students explain their argument so far and you ask them to defend specific choices they cannot defend from their notes

Distinguish Between Research Tool and Research Thinking

Your students will use these tools, and that is acceptable. What matters is whether they are using AI as a spade to dig the research, or as a replacement for knowing how to dig. A student who uses Perplexity to find relevant papers quickly, then reads those papers carefully and builds their own argument has used the tool well. A student who submits a literature review from Claude and calls it research has not. Your role is to make this distinction clear in your marking criteria and to mark based on what thinking the student shows, not whether they used the tool.

›In marking, identify the single most important claim in the work and check whether the student can defend it from primary sources, not AI summaries
›Award high marks only when students show they have read sources deeply enough to spot a contradiction or limitation the AI tool missed
›Ask students to explain why they chose to focus on one paper rather than another, revealing whether they made an active choice or followed an AI ranking

Protect Your Own Judgement Against the Persuasiveness of AI

When you use Claude to draft feedback on assignments or ChatGPT to brainstorm exam questions, you are outsourcing judgement calls that only you can make. AI-generated feedback sounds encouraging and useful but often misses the specific intellectual leap your student needs to make next. AI-generated questions cover material but often miss the underlying skills you are trying to test. Use these tools to handle low-judgement tasks like formatting reading lists or generating multiple versions of a prompt for you to choose from. Keep the judgement work in your hands, even when AI could do it faster.

›When you catch yourself about to use ChatGPT for feedback, stop and write the feedback yourself, noting what the AI version would have missed
›Use AI to generate three versions of an exam question, then rewrite the best one using your own words to ensure it tests what you actually care about
›Treat AI tools as a way to speed up clerical work, not to replace your reading of student work or your decision about what matters in your field

Key principles

1.Assessment must reveal reasoning in conditions where the student cannot hand the work to an AI system.
2.Teaching AI literacy means teaching students to spot what machines miss, not accepting machine outputs as reliable.
3.A degree that certifies skills must prove those skills through work that no tool can do for the student.
4.The intellectual struggle of research and writing builds the competence you are certifying, so protect the conditions that make that struggle necessary.
5.Your judgement about what matters in your field cannot be outsourced to any tool without losing what makes you a teacher.

Key reminders

Record yourself marking one student's work and ask why you gave the grade you did; if your reasons are about polish rather than thinking, your assessment is vulnerable to AI
Spend one seminar showing students how to spot where Elicit or Semantic Scholar AI misread a methodology section in a paper; this teaches research thinking better than any lecture on rigour
When a submitted piece looks too polished, ask the student to explain the most contentious claim in it in conversation; genuine thinking is often easier to speak than AI writing is to fake
Design one assignment per term that requires students to show their working, their doubt, and their changes of mind, not just a finished product
Before using any AI tool to help you teach, ask whether it lets you avoid making a judgement you should be making; if yes, do not use it