By WFY Bureau I Academics
For generations of university students, the essay has been more than an assignment. It has been the principal instrument through which academic understanding is demonstrated. A question is set, a deadline is given and the student disappears into libraries, lecture notes and private thought before returning with a carefully argued piece of writing.
The arrangement depends upon a simple assumption: the work submitted represents the student’s own intellectual effort.
Generative artificial intelligence has unsettled that assumption.
A student can now enter an essay question into an AI system and receive, within seconds, a fluent response containing an introduction, argument, conclusion and apparently academic vocabulary. The answer may be shallow, inaccurate or supported by invented references, but it can also be polished enough to pass a hurried reading. With additional prompting, the student can ask the system to refine its tone, include counterarguments, imitate a particular style or rewrite passages that sound mechanical.
Universities are therefore confronting a question that reaches far beyond cheating. If a student can outsource the production of a convincing academic answer, what exactly is an examination supposed to measure?
The response is already reshaping university assessment. Oral examinations are returning. Students are being asked to defend their work in person. Long essays are being divided into proposals, drafts, annotated sources and reflective commentaries. Supervised writing is reappearing. Practical demonstrations, collaborative projects and continuous assessment are gaining importance.
The essay is not necessarily disappearing. Its position as unquestioned proof of individual learning, however, is ending.
The Moment the Essay Changed
Universities have dealt with academic dishonesty for centuries. Students have copied from classmates, purchased assignments, concealed notes in examination halls and plagiarised published material. Digital technology expanded these possibilities through online essay mills, file-sharing platforms and searchable archives.
Generative AI is different because it can produce new sentences in response to an individual request. There may be no original passage for plagiarism software to locate. Two students entering the same question can receive different answers. The system can rewrite its own response repeatedly, making textual comparison increasingly difficult.
More importantly, AI is not used only to produce complete essays. Students may employ it to generate ideas, simplify difficult readings, create an outline, improve grammar, translate notes, suggest examples, summarise research or challenge an argument. The same technology can function as an unauthorised ghostwriter, a patient tutor, an editor or an accessibility aid.
This makes simple declarations about “AI use” inadequate. Asking whether a student used artificial intelligence is rather like asking whether a student used the internet. The meaningful question is how it was used, for what purpose and under which rules.
The Higher Education Policy Institute’s Student Generative Artificial Intelligence Survey 2026 found that 65 per cent of participating students believed assessment had changed significantly in response to AI. The survey, conducted among UK undergraduates, also found that generative AI had become nearly universal in student life, with 94 per cent reporting some use of it in assessed work.
That figure does not mean 94 per cent submitted machine-written essays. Permitted uses can include searching for explanations, organising material, checking language or generating practice questions. It does, however, show that universities are no longer responding to a marginal activity practised by a small group of technologically adventurous students.
AI has entered the normal academic workflow.
A Crisis of Evidence
An examination is supposed to provide evidence. A medical student must demonstrate that she understands diagnosis. An engineering student must show that he can apply technical principles. A history student must interpret evidence. A law student must construct an argument and identify relevant authority.
The submitted product has traditionally served as proof of that ability. A well-written essay indicates that the student has read, thought, selected evidence and organised an argument.
Generative AI breaks the connection between product and process. A polished final document may no longer reveal how much intellectual work the student performed. Conversely, awkward prose may not indicate weak understanding. It may belong to a student writing in a second or third language who nevertheless has a sophisticated command of the subject.
The problem is not confined to essays. AI can generate computer code, solve mathematical problems, prepare business analyses, suggest architectural concepts, draft legal submissions, produce images and create presentation slides. It can also explain how it arrived at an answer, even when that explanation does not accurately represent any genuine reasoning process.
Universities consequently face a crisis of evidence. They must find ways to establish that graduates possess the knowledge and capabilities their degrees claim to certify.
This is not merely an institutional concern. Patients, employers, clients and the public rely upon qualifications. A university cannot award a nursing, engineering, teaching or legal degree simply because a student has become skilled at obtaining plausible answers from a machine.
At the same time, graduates will enter workplaces where AI use is normal. Preventing students from learning how to use it responsibly would leave them poorly prepared for professional life.
Universities must therefore verify two different capabilities: what a student can do independently and what the student can accomplish intelligently with technological assistance.
The Return of the Oral Examination
One of the oldest forms of academic assessment is returning to prominence: the oral examination.
A student who has submitted a research paper may be invited to explain its central argument, defend the selection of evidence, respond to criticism or apply the same principle to a new problem. The questioning need not be long. Even a brief conversation can reveal whether the student understands the material or merely possesses a document that appears to do so.
Oral assessment tests qualities that a take-home essay can conceal. Can the student think under pressure? Can she explain a complicated idea in plain language? Can he recognise the limitations of his own argument? Can the student revise a position when presented with contrary evidence?
Such examinations resemble the viva voce already used for doctoral theses. The difference is that shorter oral defences are now appearing in undergraduate and postgraduate courses where they were once uncommon.
Their value lies partly in unpredictability. An AI system may help prepare a student, but it cannot silently produce every response during a live conversation. Follow-up questions can be tailored to the individual’s answer. A teacher can move from recollection to interpretation and then to application.
The oral examination is not a perfect solution. Students with anxiety, speech difficulties or certain disabilities may require adjustments. Language differences can affect performance. Cultural expectations influence how confidently students speak to authority figures. Personal questioning also creates room for unconscious bias.
Scale is another obstacle. A lecturer responsible for hundreds of students cannot easily conduct hour-long interviews. Universities may therefore use short, carefully structured conversations, sampled oral checks or group-based defences. Clear rubrics, recorded sessions and multiple assessors can make the process fairer.
Oral assessment should not become an interrogation designed to catch students. Its purpose is to reveal understanding. Used well, it can transform assessment from the inspection of a finished object into an academic conversation.
From One Submission to a Trail of Learning
Another response is to stop assessing only the final product.
A conventional essay may appear at the end of a semester with little evidence of how it developed. A staged assignment makes the learning process visible. The student might first submit a research question, followed by an annotated bibliography, an outline, a draft, peer feedback, a revised argument and a short reflection on the changes made.
AI can still be used during these stages, but the student must demonstrate continuity of thought. Why was one source selected and another rejected? How did feedback alter the argument? Which assumptions proved mistaken? What was changed between drafts, and why?
This approach is sometimes described as process-based assessment. It recognises that serious academic work rarely emerges fully formed. Scholars themselves develop ideas through reading, discussion, revision and criticism.
The method may produce better learning even without the AI challenge. Students who postpone an entire essay until the night before the deadline have fewer opportunities to receive guidance. Staged submissions encourage earlier engagement and allow lecturers to identify confusion before it becomes embedded in the final work.
They can also reduce the temptation to outsource an assignment. A student who must discuss a developing argument over several weeks cannot easily present a purchased or generated essay as a sudden personal achievement.
Yet process-based assessment requires more time from teachers. Every additional draft or conference creates marking, feedback and administrative work. Universities cannot demand increasingly elaborate assessment while reducing teaching staff or expecting lecturers to manage unreasonably large classes.
Assessment reform is therefore not simply a matter of changing questions. It is a matter of institutional resources.
The Supervised Room Returns
Some universities are reviving supervised, in-person writing. Students enter an examination room, receive a question and produce an answer without access to unrestricted digital tools.
This offers a direct way to verify individual knowledge. It may be particularly appropriate when students must demonstrate foundational competence, such as interpreting a text, performing a calculation, diagnosing a problem or constructing a basic argument.
But a complete return to traditional closed-book examinations would be a poor response.
Memory under time pressure is only one form of academic ability. In professional life, people consult documents, use software, work with colleagues and revise their writing. A three-hour handwritten examination may disadvantage students whose understanding is strong but whose handwriting speed, physical condition or examination anxiety affects performance.
There is also a danger that universities will make examinations unnecessarily difficult in an effort to defeat AI. If assessment becomes a technological obstacle course rather than a genuine test of learning, students and teachers will both lose.
The more promising model combines secure and open assessment.
The University of Sydney, for example, has developed a two-lane approach in which secure tasks verify that students have achieved essential learning outcomes, while open assessments allow them to work with AI in controlled, transparent and educationally purposeful ways.
Under such a system, a student might complete a supervised analysis to demonstrate independent understanding and later use AI during a larger project. The student would then be expected to evaluate the machine’s output, identify errors, document its use and remain accountable for the final submission.
This acknowledges reality. Universities need evidence that students can think without AI, but they must also teach them to think well with it.
The Problem with AI Detection
When generative AI first disrupted academic writing, many institutions turned to detection software. These systems analyse linguistic patterns and estimate whether a passage was produced by a machine.
The attraction is obvious. Universities already use plagiarism software, and an AI score appears to offer similar certainty. A suspicious essay can be uploaded, examined and assigned a percentage.
But AI detection is not equivalent to plagiarism matching. Traditional similarity software can point to identifiable language appearing in another source. An AI detector makes a probabilistic judgment based on characteristics of the text. It cannot show the moment at which a student supposedly asked a machine to write a paragraph.
Human writing is also unpredictable. A student may write in short, regular sentences. Another may have been taught to follow a rigid academic template. Writers using English as an additional language may employ vocabulary and structures that an algorithm interprets incorrectly. Heavy grammatical editing can further complicate classification.
AI systems themselves are constantly changing, while paraphrasing tools can alter generated text. Detection becomes a contest in which one technology attempts to identify the output of another.
Even Turnitin cautions that its AI writing model may misidentify both human and machine-generated material. Its own guidance states that an AI score should not be used as the sole basis for adverse action against a student and requires human judgment alongside institutional policy.
That warning is crucial. An automated indicator may justify a conversation or closer examination. It should not function as a verdict.
A false accusation can have serious consequences. A student may face anxiety, reputational damage, delayed graduation, loss of financial support or immigration difficulties. International students can be especially vulnerable because their right to remain in a country may depend upon academic standing.
The burden of proof must remain with the institution. A student should not be expected to prove innocence merely because an opaque system has generated a suspicious percentage.
Fair procedures require disclosure of the evidence, an opportunity to respond, trained decision-makers and a meaningful appeal. Universities must also consider drafts, notes, version histories, source use and the student’s ability to explain the work.
Academic integrity cannot be defended through procedures that are themselves unjust.
The Rules Students Cannot Find
Many disputes arise before an assignment is even written because students do not know what is permitted.
One lecturer may allow AI for brainstorming but prohibit generated sentences. Another may permit language correction. A third may ban all use. A university-wide policy may differ from instructions in a course handbook, while the wording on an individual assignment creates further ambiguity.
Students are then told to act responsibly without being given a stable definition of responsibility.
Consider a student who writes an essay independently and asks an AI tool to improve the grammar. Has the student used an editor or surrendered authorship? What if the tool rewrites entire sentences? What if a translation system is used to convert notes into English? Can AI suggest search terms? Can it summarise a journal article? May it generate a reference list if every citation is checked?
These are not trivial distinctions. They determine whether ordinary academic assistance becomes misconduct.
Universities need rules that are specific to each assessment. A simple system could identify whether AI is prohibited, permitted for limited purposes, required as part of the task or freely available with disclosure. Students should be told how any use must be acknowledged and what records they should retain.
The rules must also be taught. Publishing a policy on a distant webpage is not education. Students need examples, demonstrations and opportunities to practise acceptable use before marks and disciplinary consequences are involved.
UNESCO’s updated guidance on generative AI in education and research emphasises a human-centred approach, including protection of human agency, inclusion and cultural diversity. In 2025, a UNESCO survey found that nearly two-thirds of participating higher-education institutions had developed AI guidance or were in the process of doing so.
Policies are spreading, but clarity remains uneven. The speed of technological change means rules written one year can become inadequate the next.
Unequal Access, Unequal Advantage
AI is often described as freely available. The reality is more complicated.
Many services offer basic versions without charge, but their most capable models, larger usage limits and specialised research functions may require payment. Students with better laptops, faster internet connections and quiet private spaces can use these tools more effectively. Those who have already received digital training know how to formulate prompts, check outputs and combine several systems.
Others may depend on a shared telephone, unreliable connectivity or limited computer access. Students in rural areas and lower-income households can face a new disadvantage that is hidden behind the assumption that everyone has the same technology.
Language creates another division. Generative AI tends to perform more reliably in widely represented languages. Students working with regional languages, minority knowledge traditions or poorly digitised archives may receive weaker results. The technology can reproduce biases in the data on which it was trained, privileging dominant academic voices while marginalising others.
At the same time, AI can improve access. It can explain difficult material in simpler language, create alternative examples, assist students with certain disabilities, support translation and provide help outside normal teaching hours. A first-generation university student who is reluctant to ask a basic question in class may use an AI tutor privately and gain the confidence to participate.
The educational value depends upon access, design and guidance. If universities allow AI without providing equitable access, wealthier students gain an advantage. If they ban it completely, students who need assistive functions may lose valuable support. If institutions provide approved tools, they must also protect personal data and avoid forcing students to surrender their work to commercial platforms without informed consent.
AI equality is not achieved by declaring that a website is available to everyone.
Is This the End of Writing?
The deepest fear is that students will stop learning to write.
Writing is not merely the transcription of completed thought. The effort to organise a sentence, select evidence and connect one idea to another is itself a form of thinking. When students struggle with a paragraph, they often discover that their understanding is incomplete. If a machine performs that struggle for them, the finished prose may conceal an empty space where learning should have occurred.
Universities must therefore preserve writing, not because handwritten essays are sacred, but because sustained composition develops attention, reasoning and intellectual independence.
The solution is to design writing tasks in which thought remains visible. Students can compare an AI-generated answer with scholarly evidence, identify its omissions and rewrite it. They can annotate their own drafts, explain revisions, connect arguments to classroom discussions or apply theory to locally observed situations. They can write for different audiences and defend the choices they make.
A generic question such as “Discuss the causes of the First World War” invites a generic response. A more carefully designed task might require the student to examine conflicting primary sources, evaluate a particular historian’s interpretation and explain how new evidence changed an initial position.
AI may assist, but it cannot replace the student’s responsibility for judgment.
The essay can survive if universities stop treating it as an isolated package of polished prose and begin treating it as evidence within a larger process of inquiry.
The Teacher’s Changing Role
Assessment reform also changes academic work.
Teachers must design tasks that are meaningful, explain AI rules, examine disclosures, conduct oral checks and provide feedback across multiple stages. They need enough familiarity with the technology to recognise both its capabilities and its limitations.
This creates pressure in institutions where staff are already managing large classes, administrative duties, research expectations and insecure employment. It is easy for university leadership to announce “AI-resilient assessment”. It is much harder to provide the staffing and training required to deliver it fairly.
Teachers also disagree about the purpose of higher education. Some see generative AI as a threat to intellectual development. Others regard it as a professional tool that students must learn to command. Many occupy the uncertain middle, experimenting while fearing that a poorly designed policy may either enable dishonesty or punish legitimate learning.
Students should be involved in this redesign. They know where instructions are confusing, which tasks encourage superficial work and how AI is actually being used. Consultation does not mean allowing students to determine academic standards. It means recognising them as participants in the educational system rather than suspects to be managed.
Trust will not be restored by surveillance alone. It must be built through assessment that students consider relevant, rules they can understand and teaching that makes genuine effort worthwhile.
What a Better Examination System Could Measure
The AI era gives universities an opportunity to ask a question that should have been asked long ago: what do we genuinely want students to learn?
If the answer is only the ability to produce 2,000 polished words by a deadline, AI has made the weakness of that objective visible. If the goal is understanding, judgment, curiosity, ethical reasoning, communication and the application of knowledge, assessment can be redesigned around those qualities.
A strong system may include several kinds of evidence: supervised demonstration of foundational knowledge, extended writing developed through drafts, oral defence, practical application, group collaboration and transparent use of digital tools.
Not every assignment needs every element. The design should match the discipline. A journalism student might verify a set of claims, conduct an interview, produce a report and explain editorial decisions. An engineering student might design a solution, test it and defend the assumptions behind it. A literature student might analyse a passage orally before developing a longer written interpretation. A business student could critique an AI-generated market plan using real evidence.
The purpose is not to create assignments that no machine can touch. That contest will become increasingly futile. The purpose is to create assessment in which the student’s decisions, understanding and accountability cannot be removed.
After the Essay
The traditional take-home essay will not vanish from university life. It remains an effective way to develop research, argument and written expression. What is disappearing is the belief that the final document alone can certify who produced the thinking behind it.
Artificial intelligence has exposed weaknesses that were already present in higher education: generic assignments, inconsistent marking, excessive dependence on final examinations, vague integrity rules and insufficient attention to how students actually learn.
Universities can respond by surrounding students with suspicion, increasing surveillance and treating every polished sentence as potential evidence of misconduct. That path may produce fear without restoring confidence.
The more constructive response is to build assessment around evidence of learning. Ask students to show how an idea developed. Invite them to defend it. Require them to test sources, acknowledge tools, explain decisions and demonstrate essential abilities under secure conditions. Allow AI where it strengthens learning and restrict it where independent competence must be verified.
The future examination may contain an essay, but the essay will be only one part of the evidence.
A student may submit the paper, discuss it with a teacher, present the research process, identify how AI was used and apply the argument to an unfamiliar problem. Such an assessment is harder to fake because it asks for more than fluent prose. It asks for ownership of thought.
That may be the unexpected gift of the AI disruption.
For too long, universities have sometimes confused the production of academic language with the achievement of academic understanding. Generative AI has made that confusion impossible to ignore.
The question after the essay is not whether students can still write. It is whether universities are prepared to examine the thinking that writing is meant to reveal.