What do the VAERS autism reports ACTUALLY show?
Link to article
VAERS Autism Reports Analysis
- Study reexamined 1,800 VAERS autism reports (1990-2019) using LLMs
- Challenged 2006 paper’s conclusions about litigation-related reports:
- Original paper used keyword-based analysis
- Predicted increasing litigation trend after 2003
- Used report dates rather than vaccination dates
- Key discoveries:
- Litigation reports actually decreased after 2003
- Many 2002-2008 reports were delayed filings from 1990s
- 45% of litigation reports missed by keyword method
- 220 additional potential cases (“maybes”) identified
Temporal Analysis
Report Date vs. Vaccination Date Trends:
Example Case: Vaccination in 1996, report filed in 2007 due to initial discouragement from reporting and subsequent class action lawsuit participation.
Detection Methods
- LLM Analysis:
- Used OpenAI’s GPT4, Google’s Gemini Pro, Anthropic’s Claude, Twitter’s Grok
- Required 3+ LLM agreement for classification
- Extracted both litigation status and symptom onset dates
- Original Paper Method:
- Simple keyword search
- Missed 194 confirmed cases
- Missed all 220 “maybe” cases
Key Results
- Total Analysis:
- 419 confirmed litigation-related reports
- 194 cases missed by keyword method
- 220 potential additional cases
- Pattern Discovery:
- Most reports filed 2002-2008 were from 1990s vaccinations
- Reports clustered around class action lawsuit periods
- Initial discouragement led to delayed reporting
