Why AI Gets Exam Dates and Cut Off Marks Wrong

Ask an AI assistant when your exam is and it will tell you, in a confident sentence, with a specific date. The date may be wrong. Nothing in the answer will indicate which.

This is not a bug that will be patched next month, and it is not the AI being careless. It follows from how these systems work, which means it is predictable, which means you can protect yourself from it.

Understanding roughly why it happens tells you exactly which answers to trust and which to verify, and that is a more useful skill than avoiding the tools.

They generate, they do not look up

The single most useful thing to understand: a language model is not searching a database of facts and reporting what it finds. It is producing text that fits the pattern of a plausible answer.

Most of the time a plausible answer and a true answer are the same thing, which is why these tools are useful. But when the model has no reliable information, it does not stop. It produces the most plausible looking answer anyway, in exactly the same confident register as everything else.

An exam timetable is precisely the kind of thing this goes wrong on. The model knows what exam timetables look like. It knows the format of a date and the shape of a session listing. Producing something that looks exactly right is easy. Producing something that is right requires information it may simply not have.

Why local and specific is the danger zone

Reliability tracks roughly with how much material about a topic existed in what the model learned from.

Established science and mathematics appear in enormous quantities, are internally consistent, and rarely change. Answers there are usually solid.

The registration deadline for one entrance examination at one institution in one country, for one specific year, is the opposite in every respect. It appears rarely, it changes annually, and much of what does exist online about it is itself wrong. This is exactly the territory where confident invention happens, and it is exactly the territory students most want answers about.

Usually reliable	Verify before using
How photosynthesis works	When your papers are sat
Solving a quadratic equation	This year’s cut off mark
What a command word requires	Registration fees and deadlines
How to structure an essay	Entry requirements for a named course
General revision method	Which subjects a cluster counts
What a concours is	Any statistic, citation or quotation

The right hand column is not a list of things AI is bad at in general. It is a list of things where being wrong costs you a cycle, and where the answer looks identical whether it is right or not.

The four failures that actually cost students

Dates that have moved

Examination and registration dates change every cycle. A model may reproduce a previous year’s date, or construct a plausible one. Acting on a wrong registration deadline is the single most expensive error available here, because missing registration costs the whole year regardless of how well prepared you were.

Cut off marks stated as fixed

Ask for a cut off and you will often get a number. The deeper problem is that the question has no fixed answer at all: a cut off is produced by each year’s results, as our guide to how a concours actually works explains. A confident number implies a certainty the system does not have.

Entry requirements that are close but wrong

The most dangerous failure, because it is the hardest to notice. An answer listing four required subjects where three are correct reads entirely convincingly. A student choosing subjects on that basis discovers the error two years later, when it cannot be fixed.

Invented citations and statistics

References, quotations and figures can be fabricated entirely while being perfectly formatted. Putting one into submitted work is a fast way to lose credibility, and the marker only has to check one.

How to check in thirty seconds

Two techniques, both quick, and they catch most of it.

Ask the same question in a fresh conversation. Inconsistent answers across attempts are a strong signal the system is generating rather than recalling. Consistency does not prove correctness, but inconsistency almost proves the opposite.

Ask where the answer came from. A weak or vague response about the source tells you a great deal. Do not accept a link at face value either, because those can be constructed too. Follow it.

Then verify anything specific at the source: the examining body‘s own site, or the institution’s own admissions page. That is the only step that actually settles it, and it takes minutes against the cost of being wrong.

Ask better questions

You can shift a lot of risk by changing what you ask for.

Risky question	Better question
When is the exam?	Where would the official timetable normally be published?
What is the cut off for this course?	How are cut offs set, and why do they move?
What subjects does this course require?	What subjects do courses of this type usually require, and where do I confirm?
Give me statistics on this	What kind of organisation publishes data on this?

The right hand column asks for explanation and direction rather than for specific facts. That is the thing these systems are genuinely good at, and it points you at the source that can answer the first column properly.

Web search helps and does not solve it

Some assistants search the web and cite what they found, which genuinely improves accuracy on current facts. Two cautions remain.

The answer is only as good as the page it found, and the internet is full of outdated exam information, much of it on sites that look authoritative. A correctly cited wrong page is still wrong.

And a summary can misread its own source, compressing a conditional statement into a flat one. If a cited page matters to your plans, open it and read the relevant part yourself.

The same problem, without the AI

Worth keeping in proportion. Exam misinformation is not new and AI did not create it.

Screenshots of timetables circulate for years after they expire. Cut off marks are repeated long after they changed. Messaging groups pass around requirements nobody has checked. Students have been acting on confidently stated wrong information for as long as there have been examinations.

The habit that protects you is the same in both cases: anything specific gets confirmed at the source, whoever told you. An AI answer, a screenshot and a confident senior all sit in the same category until verified.

Why this is worse for African exam questions specifically

There is a reason students in this region hit the problem harder than the general warnings suggest.

These systems learned from what was written down and published. Official material for the Cameroon GCE, individual concours, UNEB combinations or a specific Nigerian institution’s screening process can be hard for software to read: see the check below, where the Board’s own written-phase timetable turns out to be an image-only PDF. Where the real date cannot be found, a plausible one gets generated, in the same confident voice.

Material that does exist can also be out of date. Old timetable screenshots, forum posts and pages built for search traffic can state dates or requirements that have since changed; how often was not measured. A model trained on that material can reproduce errors that are already circulating, which makes them feel corroborated when you check informally.

So the practical rule is stricter here than the general advice implies. For anything concerning a specific board, institution or concours, treat an AI answer as a starting point for where to look and never as the answer itself.

Evidence: why rare facts are the weak spot

A 2025 paper by Kalai, Nachum, Vempala and Zhang, Why Language Models Hallucinate, gives a precise version of the idea that thin documentation invites invention. For facts with no learnable pattern, such as a birthday, the authors show that after pretraining a model should hallucinate on at least the fraction of such facts that appeared exactly once in its training data. Their own example: if 20% of birthday facts appear exactly once, expect hallucination on at least 20% of birthday facts. They note that models seldom err on facts referenced many times, such as Einstein’s birthday. The paper also argues that models keep guessing because most tests reward a guess over an honest “I do not know”.

Illustrative arithmetic, not a measurement. Suppose a model had seen 50 specific exam dates, and 30 of them appeared exactly once in its training text. The paper’s lower bound would then be 30 / 50 = 60% wrong on dates of that kind. The real counts for any model are not known, and the bound applies to base models before later training that teaches abstention, so read this as a way to see why one-off, year-specific facts are the risky ones, not as an error rate. A published timetable for one exam board in one year is about as close to a once-only fact as a student will find.

A check anyone can repeat: where the dates actually are

The post links to the Board’s examinations page, so this section records what is on the Board’s website. Everything below was checked on 18 October 2026 and can be repeated by opening the same addresses.

What was checked	What was found
The Examinations page the Board uses to describe the GCE and TVE levels	It describes levels, registration conditions and regulations. It contains no exam dates, registration deadlines or calendar, and links to a separate timetables page.
The Timetables page	Lists two 2026 PDFs: the GCE Practical Timetable and the TVEE Practical Timetable. The written-phase timetable is not listed on that page.
The 2026 GCE Practical Timetable (PDF)	Has readable text. It lists examination dates from 5 May to 27 May 2026 (18 dates), subject by subject.
The 2026 GCE Written Phase Timetable (PDF, 481,281 bytes), linked from the Board’s home page	Running pdftotext on it returned no text at all. It appears to be an image-only scan, so software that reads text cannot extract its dates.

This explains nothing about any particular assistant, and no assistant was tested on these dates. What it shows is narrower and checkable: the dates a candidate needs are in PDFs on a sub-page, one of them in a form text-reading software cannot extract, while the page that describes the exams has none. A tool that cannot find the real date has, in effect and in the paper’s terms, every reason to guess one. Go to the timetables page itself, open the PDF, and read the date with your own eyes.

Questions students ask

Will this improve?

Models change quickly. The 2025 paper by Kalai and colleagues argues that hallucination persists because training and evaluation reward guessing over admitting uncertainty, which is a structural reason for caution rather than a bug awaiting a patch. Keep the verification habit regardless of how good the tools get.

Should I just not use AI for exam questions?

Use it for explanation, method and questioning yourself, where it is strong. Do not use it as a source for dates, marks, fees or requirements. That split gets you most of the benefit with almost none of the risk, and our guide to using AI for revision without cheating covers the rest.

It gave me a link to the official site.

Open it. Links can be constructed, and a real link can be attached to a claim the page does not actually make. The link is a starting point, not the verification.

Can I tell when it is unsure?

Not from the tone, which is the core of the problem. Some systems hedge when uncertain, and the hedging does not reliably track actual reliability. Judge by the type of question instead, using the table above.

The rule to keep

Use AI for how things work. Use official sources for what is happening, when, and what is required.

If an answer would change what you do, where you apply, or when you show up, it gets confirmed at the source before you act on it. Everything else in this article is detail around that one sentence.

More in AI for Students.