An AI tutor must cite its sources because a language model writes fluent answers whether or not it knows the fact, and a student is the person least equipped to tell the difference. Grounding fixes part of this: the system retrieves the relevant passage from the course material first, generates the answer from that passage, and attaches a reference you can open. The citation turns a confident claim into something you can check before you memorize it.
Why do language models make things up?
Because guessing is rewarded. In Why Language Models Hallucinate, Kalai, Nachum, Vempala and Zhang argue that hallucinations originate as ordinary errors in binary classification: if a model cannot separate false statements from true ones during pretraining, plausible falsehoods arise from statistical pressure alone. They then argue the errors persist because most benchmarks grade a confident wrong answer the same as a blank, so a model that guesses scores higher than one that says "I am not sure". Models are, in their words, optimized to be good test-takers.
The failure is measurable. When Dahl, Magesh, Suzgun and Ho asked GPT-3.5, Llama 2 and PaLM 2 legal questions, hallucination rates ranged from 69% to 88%. They also found what they call contra-factual bias, the tendency to assume that a factual premise in a query is true even if it is flatly wrong, and a tendency toward overconfidence irrespective of actual accuracy.
Translate that to a study session. Ask "why did the 1848 treaty in chapter four fail?" when chapter four has no such treaty, and an ungrounded tutor is likely to explain the failure of a treaty that does not exist.
Why is this worse for a student than for anyone else?
A lawyer can check a citation against a case database. A student is asking precisely because they do not yet know the answer, so the tutor's confidence is the only signal they have, and that confidence is uncorrelated with accuracy.
Students also arrive with a bias in the same direction. In the Roediger and Karpicke testing effect experiments, repeated studying increased students' confidence in their ability to remember the material while producing worse retention than testing on delayed tests. People already feel they know more than they do after rereading. A fluent, wrong explanation feels like more of the same.
The fix is not to stop using a tutor. Practice testing is one of two techniques rated high utility in the 2013 Dunlosky review, so a tutor that quizzes you is doing the right thing. It just has to be right about the answer key.
What is grounding, and how does retrieval-augmented generation work?
Grounding means the model answers from documents placed in front of it rather than from what it absorbed in training. The standard method is retrieval-augmented generation, introduced by Lewis and colleagues in 2020. Their system paired parametric memory, a pretrained sequence-to-sequence model, with non-parametric memory, a dense vector index of Wikipedia accessed with a neural retriever, and produced language that was more specific, diverse and factual than a parametric-only baseline.
For a tutor the pipeline has five steps. The course material is split into passages and each passage is turned into a vector. A question is turned into a vector the same way and the nearest passages are retrieved. The model is instructed to answer from those passages. The answer is returned with references to the passages it used. If nothing relevant is retrieved, the honest output is a refusal rather than a guess.
Grounding is now a product feature at the platform level. Google's Gemini API grounding documentation describes the goal as reducing model hallucinations by basing responses on real-world information and building user trust by showing the sources for the model's claims, and its grounding metadata returns citations that link specific text segments to source URLs. The same structure applies when the source is your own lecture PDF instead of the web: a span of answer, a pointer to the passage that supports it.
Does grounding eliminate hallucinations?
No, it reduces them. In a 2024 study of legal research tools built on RAG, Magesh, Surani, Dahl, Suzgun, Manning and Ho found that Lexis+ AI and Ask Practical Law AI hallucinated more than 17% of the time and Westlaw AI-Assisted Research more than 34% of the time, against 58% to 82% for GPT-4 in their earlier work. Their conclusion is the right one to carry into education: even RAG systems are not hallucination-free, with errors coming from retrieval failures, inapplicable authorities and sycophancy.
Each of those failure modes has a study-room equivalent. Retrieval failure is the tutor pulling the wrong chapter. An inapplicable authority is a passage about a different theorem with a similar name. Sycophancy is the tutor agreeing with the false premise in your question. A citation does not prevent any of them, but it makes all three visible in one click.
What should a citation in a tutor look like?
The test is whether a student can verify the answer in under a minute.
| Weak citation | Strong citation |
|---|---|
| "According to your textbook" | Page number and the passage the answer came from |
| A link to the whole PDF | A pointer that opens the PDF at the page |
| Citations only on explanations | Citations on quiz answers and mock exam keys too |
| An answer when nothing matched | "I could not find this in your material" |
| A source that was never retrieved | Only the passages actually used to generate the answer |
The last row is the one to insist on. A model can write a plausible citation just as easily as a plausible fact, so the reference must be produced by the retrieval step, not by the text generator. This is why StudyOS is built so that every answer cites the page it came from: Orby builds the mock exams and flashcards from your lecture PDFs, and each answer points back to the page in those files.
How do you test any AI tutor in five minutes?
Ask about something you know is not in your notes and see whether it refuses or invents. Ask a question with a wrong premise built in, such as a date or a name that does not appear in the material, and watch for contra-factual bias. Open three citations and check that the page actually says what the tutor said.
Then ask it to quiz you and check the answer key against the source with the same rigor. A tutor that cites its explanations but not its exam answers has left the highest-stakes output unverified.
Frequently asked questions
What does it mean for an AI tutor to be grounded?
A grounded tutor retrieves passages from a defined body of material, such as your lecture PDFs, and generates its answer from those passages rather than from training memory alone.
The approach comes from retrieval-augmented generation, which pairs a language model with a searchable document index and returns references to the passages used.
Can an AI tutor still be wrong if it cites a source?
Yes.
RAG lowers error rates without removing them; purpose-built legal RAG tools still hallucinated on more than 17% to 34% of queries in a 2024 Stanford study.
The citation's value is that it lets you open the page and check, which an uncited answer never allows.
Treat a citation as a checking tool, not a guarantee.
Why not just ask a general chatbot to explain my notes?
Because general models answered legal questions incorrectly 69% to 88% of the time in one study, tended to accept false premises, and were overconfident regardless of accuracy.
Without retrieval from your actual material, there is nothing tying the explanation to what your exam will test, and no way to tell a correct answer from a fluent one.
How can I tell if a citation is real?
Open it.
A real citation resolves to a specific page or passage in the material you supplied, and the text there supports the claim.
Be suspicious of references that name a whole book, that cannot be opened from the answer, or that appear on explanations but never on quiz answers.