The short answer. AI citations break because a language model generates text that looks like a reference; it doesn’t retrieve the reference. In a peer-reviewed test, GPT-3.5 fabricated 55% of the citations it produced and GPT-4 fabricated 18%, and a large share of the real ones contained errors (Walters and Wilder, Scientific Reports, 2023). Research teams avoid this by using AI that answers only from a defined set of source documents, shows which document each answer came from, and keeps a record of how every source was used. That record is data provenance.
Why AI citations break
The model writes references, it doesn't look them up
Walters and Wilder asked GPT-3.5 and GPT-4 to write short literature reviews and then checked all 636 citations. GPT-3.5 fabricated 55% of them. GPT-4 fabricated 18%, and 24% of GPT-4’s real citations contained substantive errors such as wrong dates or page numbers (Scientific Reports, September 2023). The models tested are older than today’s, and newer tools add web retrieval. The underlying behavior hasn’t gone away: a model that isn’t tied to a source will produce something that looks like one.
Fabricated sources reach formal work
The cost isn’t hypothetical. In Mata v. Avianca, a federal judge in New York sanctioned two lawyers in June 2023 after they filed a brief citing court decisions ChatGPT had invented (Mata v. Avianca, Inc., S.D.N.Y. 2023). The same failure in a grant application or a manuscript damages more than one document.
Nobody recorded which version was used
Reproducibility depends on knowing exactly which source informed a result. An AI answer drawn from “the protocol” is useless if the team can’t tell whether it was version 3 or version 5, or whether the document was even in the approved set.
Shadow AI moves research into tools nobody approved
Researchers adopted AI faster than their institutions supported it. Wiley’s 2025 survey of 2,430 researchers found AI use rose from 57% in 2024 to 84% in 2025, with 80% using mainstream tools like ChatGPT. Only 41% felt their organization offered adequate support, and concern about hallucinations and inaccuracies rose from 51% to 64% (Wiley, ExplanAItions 2025).
That gap is shadow AI: unpublished data, draft manuscripts and protocols pasted into general-purpose tools the institution never approved. Two integrity problems follow. The answers aren’t tied to the team’s approved sources, so nobody can trace a claim back to the document behind it. And the source material has left the controlled environment, which matters for anything under a data use agreement or embargo.
Universities face the same gap with student records, covered in our FERPA and AI guide. Banning the tools doesn’t close the gap. Giving researchers an approved tool that answers from the team’s own documents does.
What data provenance means for a research team
Data provenance is the record of where a piece of information came from and how it was used. For AI-assisted research, that breaks into two questions. For a single answer: which document produced this, and can I open it? For the team’s corpus over time: when was each document added, and which answers relied on it?
An AI answer without a source is a lead, not a finding. Treat it the way you’d treat an unattributed quote.
What to look for in a source verification tool
Use these criteria to compare tools, including ours.
| Criterion | What good looks like |
|---|---|
| Closed corpus | The AI answers from documents your team approved, and you can turn off general-knowledge answers |
| Answer-level citations | Every answer names the documents it relied on, and you can open them in one click |
| Confidence signal | The tool tells you how well an answer is supported, and never labels an uncited answer as trusted |
| Conflict handling | When two sources disagree, you see both instead of a blended answer |
| Document registration | Each source carries an identifier and a timestamp |
| Usage record | You can see when each document was cited or viewed as evidence |
| Access control | Collaborators only reach the material they're cleared for |
Where Mithra fits
Mithra is built around those criteria for an organization’s own documents. Files go into permissioned Library spaces and are tokenized on upload, which registers each one with a Token ID, a timestamp, its owner and its metadata. Only tokenized documents are searchable.
Every answer carries a Trusted, Partial or Untrusted label, or General Knowledge when no internal document applies, and an answer with no citation to an internal document can never be labeled Trusted. View Evidence opens each cited source with its Token ID and last-verified timestamp. With “Use only your data sources” turned on, questions the corpus can’t answer are declined rather than filled in from general knowledge. When two documents disagree, Mithra shows the evidence from both. Each document’s Data Lineage view logs when it was tokenized and each time it was cited or viewed as evidence.
Two limits to be clear about. Mithra checks that an answer is tied to a document in your library; it doesn’t check whether that document is right. And it works on the files you upload (PDF, DOCX, TXT, XLSX, XLS and CSV); it doesn’t search external literature databases. Admins can exclude outdated versions by metadata value so a superseded protocol stops answering questions.
Frequently asked questions
Why do AI research tools produce fake citations?
Because a language model generates text that resembles a reference rather than retrieving one from a database. A peer-reviewed test found GPT-4 fabricated 18% of the citations it produced (Walters and Wilder, 2023).
How can researchers verify AI-generated sources?
Open every cited source before relying on it, and prefer tools that answer only from an approved document set and link each answer to the document it used.
What is data provenance in scientific research?
The record of where each piece of information came from and how it has been used, down to the document and version. It’s what lets another researcher trace a result back to its sources.
About Mithra Technologies Inc.
Mithra Technologies Inc. is building the trust infrastructure for enterprise AI. Centered on its cryptographic Single Source of Truth (SSOT®), Mithra enables organizations to establish verifiable AI through trusted enterprise content, cryptographic proof, complete data lineage, provenance and governance. Designed to work across any AI model, Mithra helps enterprises deploy AI that is more transparent, explainable, auditable and resilient while reducing the risks associated with misinformation, unauthorized data usage, cyber threats and the emerging quantum era.