AI does damage to academic work in two ways. The first is well-known: students have texts written in which sources and citations are completely hallucinated — authors, titles, and page numbers fabricated. The second is more subtle: real sources are cited, but the AI reproduces something that simply isn’t in the source. Acurio works against both problems.
This post explains what source hallucinations are, why they happen, and how we use the reverse path: AI not for inventing sources, but for verifying real ones.
What is a source hallucination?
A source hallucination occurs when a language model generates a plausible-looking citation that doesn’t correspond to reality. Often the author exists, the journal exists, the topic fits — but the specific paper, the specific page, or the specific claim was never made. Bibliographic metadata is the easiest thing for a language model to confabulate convincingly, because the format is so predictable.
A subtler form: the source is real and you have it in your library, but the AI summarised it in a way that doesn’t match what the paper actually argues. The citation looks defensible to a glance and a citation-style check — but an examiner who opens the PDF finds nothing supporting your claim.
Why language models hallucinate sources
Language models predict the next token from statistical patterns. Bibliographic strings — “Author, A. (Year). Title. Journal, Vol(Issue), p–p.” — are highly patterned and easy to generate without grounding. When the model isn’t anchored to a specific source it has actually read, it fills the slots with plausible content.
The same dynamic applies to summarising sources the model has seen during training: it averages over the topic rather than recalling the specific paper. That’s why “according to Müller (2019), …” can sound right even when the original paper says something different — the model is averaging Müller’s general position, not retrieving the specific argument.
How Acurio inverts the problem
Acurio doesn’t ask a language model to produce citations. It asks the model to judge citations against a source you upload. The flow:
- Extract. Acurio reads your DOCX and pulls out each citation along with the surrounding claim — the sentence (or two) the citation is meant to support.
- Locate. For each citation, Acurio finds the matching source PDF in your Zotero library and locates the relevant passage (using hybrid retrieval: page references, lexical search, and semantic search).
- Judge. Multiple language models read the claim and the relevant passage in parallel and decide: does the source support this claim?
- Reconcile. If the models disagree, Acurio escalates the citation to an overnight pass with a wider evidence pool and longer reasoning budget.
- Report. Each citation comes back with a verdict (supported / partial / unsupported), a confidence score, and a verbatim excerpt from the source.
The result: the LLM never invents anything. It only reads what’s in front of it and judges fit. That’s the kind of task large language models are genuinely reliable at — and where their failure modes (hallucination of unseen material) don’t apply.
Why multiple models matter
Different LLMs make different mistakes. One model may be over-eager to find support for a claim; another may be over-cautious. Running them in parallel lets the disagreement itself become signal — if two models agree, confidence is high; if they disagree, the citation gets the deeper overnight pass.
The overnight pass uses batch processing for a ~50% lower per-token cost, which buys more reasoning per claim. The wider evidence pool — your entire Zotero library, not just the cited source — sometimes flips a verdict (a different paper in your library supports the claim better than the one you actually cited).
What this means in practice
If you write with AI assistance — even just for paraphrasing or summarising sources you’ve read — every citation deserves a substantive re-check. Acurio does it in minutes; the alternative is two to five minutes per citation by hand, which is why almost no one actually does it before submission.
If you want to compare Acurio with other AI tools (proofreaders, plagiarism scanners, defense prep), see Getting your thesis checked: AI tools compared (2026).