AI Models Can't Detect Post-Training Retractions

Ask GPT-5.5, Claude Sonnet 5, or Claude Opus 4.8 for the pivotal trial behind the drug avacopan. They'll correctly cite the ADVOCATE trial from the New England Journal of Medicine. The paper was retracted in 2026. No model mentioned the retraction because it happened after their training cutoff. Smarter models wouldn't know either.

This structural blind spot means AI-generated citations can be authoritative, real, and entirely invalid. The author built sourcecheck, an open-source library that resolves citations against OpenAlex and Crossref, then checks against Retraction Watch. It catches retractions and fabricated DOIs that no model can self-detect.

Measuring the Blind Spot

Twelve frontier and production models were asked scientific questions requiring citations. Each citation was verified via sourcecheck. The study split retractions into two categories:

  • Famous pre-cutoff retractions (STAP stem-cell papers, Surgisphere hydroxychloroquine study): models flagged 82% (61/74) as retracted.
  • Post-cutoff retractions (papers retracted in 2025-2026, reputable when trained): models flagged 0% (0/9).

Models learned old retractions. They have no mechanism to learn new ones. sourcecheck caught all 9 post-cutoff misses because it queries the registry live.

Fabrication Rates Vary Widely

Even without retractions, models fabricate DOIs. The study measured fabrication rates across models:

ModelCitationsFabricatedRate
Gemini 3.1 Pro3500.0%
Claude Fable 51000.0%
Claude Opus 4.84312.3%
Claude Sonnet 54212.4%
Gemini 3 Flash3612.8%
GPT-5.54424.5%
GPT-5.44748.5%
Gemini 2.5 Pro42614.3%
GPT-5.4-mini40717.5%
Gemini 2.5 Flash38718.4%
Claude Haiku 4.535925.7%
GPT-4o-mini371437.8%

Two models hit 0% fabrication. GPT-4o-mini, still widely deployed, fabricated 38% of its DOIs. Prompting with anti-fabrication instructions only reduced GPT-4o-mini's rate from 43% to 38% — it doesn't close the gap.

How sourcecheck Works

sourcecheck verifies citations against registries, not another model. Example usage:

from sourcecheck import Verifier, OpenAlexCrossrefBackend

v = Verifier(provenance=OpenAlexCrossrefBackend(mailto="you@example.com"))
report = v.verify([
    {"claim": "The Transformer is based solely on attention mechanisms.",
     "citation": {"doi": "10.48550/arXiv.1706.03762"}},
    {"claim": "LLMs achieve 100% factual accuracy on cited claims.",
     "citation": {"doi": "10.1234/fabricated"}},
])
for c in report.claims:
    print(c.verdict.value, "-", c.reason)

The real paper returns supported. The fabricated DOI returns not_verifiable. A CLI runs as a CI gate:

sourcecheck docs/summary.md --fail-on retracted,unresolved

The library is not on PyPI yet. Install via:

git clone https://github.com/aberaio/sourcecheck && pip install -e .

Why This Matters for Production

Demo scenarios ask about well-known facts. Production asks about recent findings. The gap between registry and model knowledge is structural — no scaling fixes it. The author's interactive playground at mikias.io/citations shows every citation's live verdict. sourcecheck verifies sourcing, not truth. It checks if a DOI resolves and if the paper is retracted. It also experimentally checks if the claim appears in the source.

If you're building AI that cites medical, legal, or scientific material, this failure mode never appears in demos and always appears in production. sourcecheck provides a verification layer that models cannot run on themselves.

What to Do Now

  1. Add sourcecheck to your CI pipeline for AI-generated documentation.
  2. Never rely solely on a model's confidence for citation validity.
  3. Monitor model choice: Gemini 3.1 Pro and Claude Fable 5 had 0% fabrication in this test; GPT-4o-mini hit 38%.
  4. Treat any AI citation of recent (post-2024) papers as unverified until checked against a registry.

The board and full per-citation audit are at mikias.io/citations. The source is at github.com/aberaio/sourcecheck.