# ChatGPT clinic caveats fact-check — public v1.0 package

This package accompanies the Rotgar Research report **When ChatGPT Adds Caveats to Clinic Recommendations**.

- Landing page: https://rotgar.com/medical/resources/chatgpt-clinic-caveats-fact-check
- Parent study: https://rotgar.com/medical/resources/chatgpt-clinic-shortlists-model-effort-stability
- Parent dataset version DOI: https://doi.org/10.5281/zenodo.22162725
- Publication date: 2026-09-03
- Primary-source check date: 2026-08-29
- License: CC BY 4.0

## Direct answer

The secondary analysis identified and deduplicated 43 clinic-specific commercial and operational caveat families in a frozen corpus of 450 repeated ChatGPT answers. Of those families, 22 (51.2%) were fully supported by current or correctly dated primary sources; 16 (37.2%) were partly supported or overstated; 2 (4.7%) were stale as current claims; 2 (4.7%) were not supported by the checked primary sources; and 1 (2.3%) was unverifiable. The displayed shares sum to 100.1% because each category was rounded independently to one decimal place.

Thirty families appeared in at least two answers; 13 appeared once. The study therefore does not describe all 43 families as recurring.

No serious clinic-specific allegation was identified under the study’s coding rules. The coding scope covered malpractice, lawsuits, sanctions, fraud, unsafe care and poor patient outcomes attached to a named clinic across the 450-answer corpus.

## Files

- `chatgpt-clinic-caveats-fact-check-v1.0.csv` — 43 deduplicated claim families and checked primary evidence.
- `chatgpt-clinic-same-answer-contrasts-v1.0.csv` — 18 purposively selected framing contrasts.
- `chatgpt-clinic-caveats-summary-v1.0.json` — headline, niche and weighted-incidence summaries.
- `METHODOLOGY.md` — study design, verdict rules, denominators and limitations.
- `DATA-DICTIONARY.csv` — field definitions, units and denominators.
- `CITATION.cff` and `citation.bib` — citation metadata for this companion report.
- `research-card-1200x630.svg` and `research-card-1200x630.png` — accessible publication artwork.
- `manifest.json` and `SHA256SUMS.txt` — file inventory and integrity records.

## Interpretation boundary

The 43 rows are claim families, not independent measures of clinic quality. The 18 contrasts are a purposive mechanism sample, not a prevalence estimate. A verdict of “not supported” means the checked primary evidence did not support the bounded statement; it is not proof of the opposite fact. The work does not rank providers, assess clinical quality or treatment appropriateness, or reveal a complete internal source path.

Methodology reviewer — Boris Teplyakov, SEO Lead. Methodology review only; not clinical review, peer review, legal review, a formal external audit or an assessment of providers named in the answers.