Last updated 10 min read
Key takeaways
An AI hallucination is an answer that reads as fluent and confident but is false. The term covers a wide range: a book that was never written, a quotation nobody said, a date off by a decade, a statistic with no source, or a step in a proof that looks right and does not follow from the one before it. What makes hallucinations dangerous for students is not that they happen but that they look exactly like correct answers. The tone is the same, the formatting is the same, and there is no warning label.
They happen because of how large language models work. A model generates text by predicting what is likely to come next, based on patterns in an enormous amount of training text. Usually the likely continuation is also the true one, which is why these tools are so often right. But when the model has little reliable information about something, such as a niche topic, a recent event or the exact wording of a source, the most plausible-sounding continuation can be invented. A citation has a familiar shape: author, year, title, journal. The model can produce that shape without any real article behind it.
Hallucinations are not the model lying, and no single update will remove them. Newer tools tend to hallucinate less on common questions, and tools that answer from a document you supply less still, but none should be treated as a source in its own right.
Some kinds of error come up again and again in study work. Knowing the shape of each one makes it much faster to spot, and tells you exactly where to go to check it.
| Type | What it looks like | How to catch it |
|---|---|---|
| Invented citation | A real-sounding article with plausible authors, a journal name, a year and sometimes a DOI that leads nowhere | Search the exact title in your library catalog or a scholarly search engine, and open the DOI |
| Misattributed or invented quotation | A memorable line credited to a famous writer, or a “quote” from your set text that is not in it | Find the line in your own copy of the text, on the page |
| Wrong specific detail | A date, number, name or formula that is close to right but not right | Check your textbook or lecture notes, not the same AI again |
| Faulty reasoning step | A worked solution where one line does not follow, often near the end | Redo the step yourself or substitute the answer back into the original problem |
| Invented plot or argument detail | A character placed in a scene they are not in, or an argument a philosopher never made | Go to the chapter or passage and read it |
| False claim about your course | Confident statements about what your syllabus covers or how your instructor grades | Check the syllabus or ask your instructor |
You cannot tell a hallucination from its tone, but some situations make one much more likely. When you notice any of the signs listed below, slow down and check before you use the answer.
One quick test: ask a question with a false premise, such as “Why did Darwin reject natural selection?” A reliable answer corrects the premise. One that explains the rejection shows how readily the tool builds on a mistake, including one you make by accident.
Checking does not mean distrusting everything equally. It means separating the claims that matter from the ones that do not, and tracing the ones that matter to something you can open. The routine below takes a few minutes for a typical answer, and much less once it becomes a habit. The guide to finding credible sources covers the wider skill.
For academic sources, knowing how to use Google Scholar effectively makes step three below quick: search the title in quotation marks and see whether the article exists at all.
Suppose you are writing a psychology essay on sleep and memory and ask an AI for three sources. It gives you three tidy references, each with authors, a year, a journal and a one-line summary. Two turn up in a scholarly search immediately. The third has a plausible title and a real journal, and one of its authors is a researcher who does work on sleep, but no search finds the article itself.
This is a typical invented citation: real pieces, a real journal and a real name in the field, assembled into a reference that does not exist. If you cited it, your reader would find nothing, and its one-line summary would be a claim with no evidence behind it.
The fix is to treat the AI's list as a set of search leads, not a bibliography. Search for the real researcher's work on the topic, read the abstracts of what you find, and cite only papers you have opened. The guide to evaluating a source covers what to look for once the real article is in front of you.
The biggest single change you can make is to ask questions the AI can answer from material in front of it. Give it the actual chapter, article or lecture notes, ask it to answer only from that text, to say “that is not in the document” when the answer is not there, and to give a page reference for each claim. A reference does not guarantee accuracy, but it turns checking from a search into a glance. Keep questions narrow, too: “explain the second premise of this argument” leaves less room to drift than “tell me everything about this philosopher.”
Second, use AI for jobs where an error is easy to spot. Explaining a concept you can then check against your textbook, quizzing you on notes you wrote, or suggesting search terms are all low risk. Supplying facts you will repeat without checking, or a bibliography you will paste in directly, is high risk. The guide on using AI for research without being misled goes further on this split.
On Cavua, every course has a 24/7 AI tutor that answers from that course's own syllabus, lessons and tests rather than the open web, whether you reach it by video call, voice call, chat or an audio summary. Because it draws on the same material you are studying, you can check any explanation against the lesson it came from.
For your own reading, the Work space lets you upload PDFs, Word files and more and ask questions across several documents at once, with answers that point to the page. Grounding like this makes errors less likely and much easier to catch, but it does not make them impossible, so the checking habits in this guide still apply. How much is included depends on your plan, and each one is listed on the pricing page.
In most courses, you are responsible for every source you cite and every fact you submit, whatever tool produced it. A fabricated citation in an essay can be treated as an academic integrity problem even if you did not invent it yourself, because the reference claims you consulted a source you never opened. Check your course's policy on AI use, and if it is unclear, ask before you rely on AI for graded work.
The quieter risk is to your own understanding. A wrong date or a muddled definition learned from an AI answer can sit in your memory for weeks and surface in an exam. Tracing claims to your textbook costs a few minutes; unlearning a confident mistake costs more. The habit also teaches you which questions a given tool handles well.
The main AI study tools explained: tutors, summarizers, flashcard apps, document chat and research assistants, what each does well and where each goes wrong.
How to learn a language with AI: what an AI tutor does well, where it falls short, a 30-minute daily routine, role-play speaking practice and an error log.
Can AI replace teachers? No. Here is what AI tutors do well, what only a human teacher can do, and how students can get the most out of having both.
AI prompts for studying that teach rather than tell: a five-part template, weak vs strong examples, and copy-ready prompts for understanding and practice.
AI tutor vs ChatGPT for studying: what each knows before you ask, where each is strongest, and how to use both without getting mixed messages.
How to study with AI the right way: attempt first, use it to question and correct you, check its answers, and measure what you can recall on your own.
AI language models generate text by predicting what is likely to come next, based on patterns learned from huge amounts of training text. When the model lacks reliable information, the most plausible-sounding continuation can be false: a citation with the right shape but no real article, or a date that is close but wrong. It is a side effect of generating language rather than looking facts up, so it happens even in tools that are usually accurate.
Yes. Invented citations are among the most common hallucinations students meet. They often combine real elements, such as a real journal and a researcher who works in the field, into an article that does not exist. Search the exact title in a library catalog or scholarly database, confirm that the authors, year and journal match, and open the article before you cite it.
Not from its tone, since hallucinations sound as confident as correct answers. Look for risk signs instead: precise figures without a source, quotations that fit your argument perfectly, niche or recent topics, and answers that change when you ask again. Then check the claims you plan to use against a source you can open yourself, such as your textbook, the original text or a scholarly database.
It reduces them and makes them easier to catch, but it does not stop them. An AI answering from your document can still blend two passages, misread a table or fill a gap with something plausible. Ask it to answer only from the document, to say when something is not there, and to give a page reference for each claim, then open the page to confirm.
In most courses, you are responsible for the sources in work you submit, whatever tool suggested them. A citation to an article that does not exist can be treated as an academic integrity issue, because it claims you consulted something you did not. Policies differ, so check your institution's rules on AI use and verify every reference before it goes into an assignment.
Every Cavua course comes with a 24/7 tutor you can video call, speak to or chat with — on the material you are actually studying.