How generative AI is changing university assessment this academic year

Generative AI is shifting university assessment away from grading the final written product and towards evaluating the student’s process, critical thinking, and ability to verify AI outputs. Institutions are replacing traditional essays and unsupervised coding assignments with oral vivas, in-class supervised coding tasks, and exams that require students to critique AI-generated solutions.

For students sitting artificial intelligence and machine learning modules this academic year, this shift means you are no longer just expected to write code or recall definitions from a textbook. You are expected to demonstrate how you formulate problems, evaluate models, and navigate the technical limitations of the tools you use. The days of submitting a completely unsupervised take-home coursework assignment as your sole grade are rapidly disappearing. Instead, examiners want to see the cognitive work that happens before you write the code and the validation work that happens after the code is compiled. This structural change fundamentally alters how you need to revise for your upcoming exams and approach your practical assignments.

The move from code generation to code critique

Previously, a standard machine learning assignment might ask you to build a convolutional neural network to classify an image dataset and submit your Python script for marking. Now, large language models can generate that exact boilerplate code in seconds. Because generating standard scripts is trivial, university examiners have updated their rubrics. They now test your ability to read, debug, and critique code that an AI has already written. You will likely face exam questions presenting a flawed AI-generated script and asking you to identify the security vulnerabilities, the statistical biases, or the mathematical errors within it.

Consider a worked example you might see in a mid-term paper. The prompt provides a snippet of Python code generated by an AI designed to train a model using the scikit-learn library. The exam question asks you to find the logical error in the pipeline. A student relying purely on memorisation might struggle to find the fault. A student who understands the mechanics of data leakage will immediately spot that the AI applied the StandardScaler to the entire dataset before applying the train_test_split function. This means information from the test set has leaked into the training set, rendering the evaluation metrics completely invalid.

This type of assessment tests your architectural understanding rather than your typing speed. It proves to the examiner that you understand the underlying mathematics of machine learning and the specific rules of the framework you are operating in. To prepare for this, your revision needs to include reading other people’s code, breaking it down line by line, and deliberately inserting logical errors into functioning scripts to see how the system fails. You must train yourself to spot hallucinations, such as an AI calling a library function with a parameter that does not actually exist.

Authentic assessment and the rise of the viva voce

Another major change this academic year is the widespread introduction of oral examinations, commonly known as vivas, for undergraduate and master’s level modules. While vivas have traditionally been reserved for PhD defences, they are now becoming a standard method to verify that a student actually understands the coursework they have submitted. During these sessions, an examiner will sit with you for fifteen to thirty minutes and ask you to explain specific decisions you made in your machine learning project.

For example, if you submitted a natural language processing model that classifies document sentiment, the examiner might ask why you chose a complex transformer architecture over a much simpler term frequency-inverse document frequency approach for this specific dataset. They might ask you to explain how changing the learning rate would impact your validation loss, or ask you to draw the architecture of your neural network on a whiteboard. If a language model wrote your coursework and you did not bother to understand it, you will fail to answer these highly specific, contextual questions. The assessment is the conversation itself, not just the submitted code file.

Revising for a viva requires a completely different strategy than revising for a written paper. You need to practise speaking about technical concepts out loud. Gather a group of course mates and take turns interrogating each other’s project repositories. Force yourself to explain complex concepts like attention mechanisms, cross-entropy loss, or vanishing gradients without looking at your notes. This active recall and verbal reasoning will build the technical fluency you need to satisfy an examiner during a high-pressure oral assessment.

Evaluating the process through version control and prompt logs

Universities are increasingly interested in how you arrived at your final answer, heavily weighting your methodology over your final output. In practical terms, this means you will be assessed on your Git commit history, your architectural planning diagrams, and your prompt engineering logs. If you use generative AI to assist with your coursework, many modules now require you to submit an appendix detailing exactly what prompts you used, what the AI returned, and how you modified that output to suit your specific engineering problem.

A concrete example of this is the requirement to document your hyperparameter tuning process. An examiner wants to see the systematic steps you took to improve your model’s accuracy. If your Git history shows a single massive commit containing a perfectly tuned model with no prior experimental branches, this will immediately trigger academic integrity flags. Conversely, a commit history showing initial baseline models, experimental tweaks, failed attempts, and incremental improvements demonstrates genuine engagement with the scientific method.

Therefore, you must treat your process documentation as a core part of your examinable material. Start logging your work from day one of the assignment. When you ask a language model to explain a mathematical concept or help you debug a specific dependency error, save that interaction. In your exam or coursework report, you can then critically reflect on the AI’s suggestions, noting exactly where it provided bad advice and how you corrected it. This type of critical reflection proves your competence and is highly rewarded in current university mark schemes.

Integrating AI as a permitted tool in controlled environments

Rather than banning generative tools entirely, forward-thinking university computer science departments are integrating them into supervised, timed examinations. You may find yourself sitting in a computer lab with restricted internet access, explicitly permitted to use a specific, university-hosted local language model. The exam question will require you to use this tool to solve a complex data science problem within a strict two-hour time limit.

In this scenario, the AI acts as an advanced calculator. The assessment tests your problem formulation and systems engineering skills. If you feed the model vague instructions, it will return vague or hallucinated code, and you will run out of time trying to fix it. If you break the problem down into clear, modular steps and direct the AI effectively, you will succeed. You are being tested on your ability to manage an intelligent but highly unreliable digital assistant under strict time pressure.

To excel in these environments, you must practise the skill of constraint-based prompting. Work on specifying your inputs precisely, defining the expected data types, and setting strict boundaries for the output. Understanding the limitations of current generation models is crucial here. If you know that a particular model struggles with complex tensor reshaping operations in PyTorch, you will know to write that specific part of the code yourself rather than wasting valuable exam time arguing with the AI to fix its own mistakes.

How to answer this in an exam

When you face an exam question asking you to evaluate an AI-generated solution or discuss the limitations of a machine learning architecture, examiners are looking for critical detachment. Mark schemes reward students who can move beyond simply describing what the code does and instead analyse why it was generated that way. You need to explicitly name the underlying principles. This means citing the specific statistical bias present in the training data or identifying the exact mathematical flaw in the AI’s logic.

Top marks are awarded to students who demonstrate advanced verification skills. If you are asked to correct an AI hallucination in a piece of code, you must state clearly how you would test the corrected code to prove it works. You should mention writing unit tests, checking edge cases, and validating the model against a known holdout set. The Full Marks Press AI Ethics and Responsible AI guide covers this with worked exam questions, showing you exactly how to structure these critical evaluations to pick up every available mark on the rubric.

Frequently asked questions

Will universities completely ban generative AI in exams? Most universities are moving away from blanket bans because they are impossible to enforce fairly. Instead, they are designing assessments like oral vivas and supervised lab tests, or they are explicitly permitting AI use under strictly documented conditions to test your verification skills.

How can I prove my code is my own work? Maintain a detailed version control history with frequent, descriptive commits. Documenting your failed experiments and the logical steps you took to fix bugs provides a clear audit trail that examiners look for to verify your authorship.

What should I do if an AI generates the correct answer for my coursework? Acknowledge the use of the tool in your methodology section, include the exact prompts you used in an appendix, and spend your word count critically analysing why the answer is correct and what limitations the model might have in a real production environment.

Why are oral examinations becoming more common in computer science? Oral examinations allow examiners to instantly verify your depth of knowledge. They prevent students from passing off AI-generated text as their own by forcing them to defend their technical design choices verbally in real time.

For more practitioner-led revision strategies and complete module breakdowns, you can download a free copy at fullmarkspress.com/free.