How accurate are AI-generated references?

AI language models generate surprisingly convincing academic references—but how many of them are real? In a 2025 evaluation of 400 references generated by eight AI chatbots, only 26.5% were fully correct and 39.8% were fabricated or incorrect (Cabezas-Clavijo & Sidorenko-Bautista, arXiv:2505.18059).

What actually predicts a fabricated reference

Not the model's brand name. A 2026 study of 38 models—8,913 generated references across 24 topics—found that two variables predict whether a model recalls a real paper: the log of its parameter count, and the log of how often the topic appears in the literature. Together they explain about 60% of the variance, rising to 74–94% within a single model family (Smith et al., arXiv:2605.18732). Every reference in that study was checked with SourceVerify using the SVRIS comparison standard, which agreed with human reviewers 94.4% of the time (Cohen's κ = 0.887).

This is why a single "hallucination rate" per model misleads. Averaging across topics overstates reliability on the long tail— precisely where most real research questions sit. The same model can be dependable on a well-covered topic and unreliable on a niche one.

That study scores each generated reference on recall quality: is the reference real, and is it on topic. Three of its findings make the pattern concrete.

• Scale extends reach rather than fixing accuracy. The smallest models recalled only heavily cited work (median ~2,419 citations); the largest reached into far less famous papers (~589–806).
• The smallest models barely try. Llama 3.2 1B returned the same invented "Dahl" reference for all nine political-science topics tested, changing only the year.
• Even a frontier model has a floor. DeepSeek V4 Pro scored 0.90 recall quality on a well-covered topic and 0.43 on a rare one—a mixture-of-experts model outside the study's primary dense fit, but the same pattern. Higher is better on this scale: 0.90 means nearly every reference was both real and on topic.

And the gap does not close with size. Raising the rarest topic tested to 0.90 quality by scale alone would need a model roughly 30× larger than the biggest one evaluated—around 50 trillion parameters. For the long tail, retrieval and verification are the only practical answers.

Estimate recall quality for your own topic and model size →

Why Even "Better" Models Still Fail

The curve that predicts recall is an S-shape, not a straight line. Adding parameters raises the ceiling on topics the model has already seen often enough to resolve, but it barely lifts the floor under rare ones, where the signal from a handful of training mentions stays buried under everything else compressed into the same space (Smith et al., arXiv:2605.18732). A bigger model is a model that reaches further down the tail—not one that stops inventing.

The practical consequence is arithmetic. At the 39.8% fabricated-or-incorrect rate measured by Cabezas-Clavijo & Sidorenko-Bautista (arXiv:2505.18059), a 40-reference bibliography carries roughly 16 entries that need correction before submission.

Types of Citation Errors

AI citation errors fall into several categories:

• Complete fabrication — The paper, journal, and authors don't exist (most common)
• Author confusion — Real author name, wrong paper
• Journal mismatch — Real paper, wrong publication venue
• Year errors — Real paper, wrong publication date
• Title modifications — Similar to a real title but subtly different

Retrieval-Augmented Generation (RAG) Helps—But Isn't Perfect

Tools that use retrieval-augmented generation (RAG) have lower hallucination rates because they search actual databases rather than generating from memory. Smith et al. reach the same conclusion from the other direction: below the frequency at which a topic can be recalled reliably, no amount of extra scale substitutes for looking the reference up (arXiv:2605.18732). But retrieval fails in its own ways. A RAG system can still:

• Return the wrong paper from search results
• Mix metadata from multiple sources
• Misinterpret or truncate citation details

The Solution: Verification

Given that even the best AI models produce unreliable citations, the only safe approach is to verify every reference. This means checking against authoritative databases like OpenAlex, and Google Scholar.

SourceVerify automates this verification using the SVRIS standard—a transparent, auditable method that shows you exactly how each citation was verified. Unlike black-box AI checkers, SVRIS provides auditable results: you can see which fields matched (title, authors, year, venue) and which sources provided evidence.

Summary

AI-generated references are unreliable across every model tested, and how unreliable depends on the model's size and how well covered your topic is—not on which brand you used. In a 2026 study of 38 models, recall of real papers was predictable from those two variables alone (Smith et al., arXiv:2605.18732). No current model can be trusted to produce accurate citations without checking. For academic and professional work, use SVRIS-based verification to catch fabricated references before they damage your credibility.