How the AI Audit Was Conducted
This page explains the methods used to examine the records and identify inconsistencies discussed in this project.
This project uses modern language-model tools to analyze the records examined here and highlight where explanations diverge from the written record.
The goal of the analysis is not to generate new evidence or opinions, but to examine whether the reasoning contained in administrative decisions aligns with the documents those decisions describe.
Document Review
The analysis in this project is based on records related to a workers’ compensation claim for a workplace injury.
These records include several types of documents that are part of the workers’ compensation process.
Requests for Authorization (RFAs)
These are treatment requests written by my treating physician — the doctor responsible for my medical care.
The doctor submits these requests asking the workers’ compensation system to approve treatment for the injury.
Utilization Review (UR) decisions
These are written responses to those treatment requests.
They are prepared by doctors or nurses retained by Apple and its claims administrator, Sedgwick, to review the request and decide whether the treatment will be approved.
Claim correspondence
Emails, letters, and other communications related to the claim, including exchanges with claims adjusters, their supervisors, defense counsel, and other representatives involved in the administration of the claim.
Medical records
Clinical notes and treatment documentation describing the injury and the treatments that have been attempted.
Legal and regulatory references
Relevant provisions from the California Labor Code and other statutes referenced in the claim, including laws governing treatment authorization and related federal statutes.
These materials make it possible to compare the medical history, treatment requests, and the explanations used to approve or deny treatment.
Analytical Approach
The analysis focuses on a straightforward comparison:
what the documents say
and
what decision-makers said the documents meant.
Large language models are particularly effective at examining written explanations within large collections of text.
In this project, tools including ChatGPT, Claude, Grok, and Gemini were used to assist in reviewing the records examined here and identifying where explanations in administrative decisions diverge from the underlying documents.
The systems were used to help:
• compare denial explanations with the treatment history described in physician requests
• compare utilization review reasoning with the documents reviewed by those decisions
• review regulatory language referenced in treatment decisions
• identify statements that conflict with other portions of the written record
The models assist with document comparison and pattern recognition across the materials reviewed.
Scope of the Analysis
The analysis presented in this project reflects the records examined here.
The purpose is to highlight where explanations contained in claim decisions diverge from the written record available in those materials.
Readers are encouraged to review the documents and draw their own conclusions.