HugoScore hugoscore.org

Full review

Fight Health Insurance CAIHL draft report

Evidence-linked HugoScore draft report for a health AI tool that affects patients.

Agency checklist · October 30, 2026

Expands 10 of 12 yes

Document-based draft. Independent confirmation is incomplete. Hugo has not personally verified this service.

Assessed function: Fight Health Insurance hosted site: AI-drafted insurance appeal letters for patients (open-source code)

Initial Anthropic coding has an Anthropic model in the role record. Astra review conflicted: OpenAI model in role record.

A no can mean missing documentation or institution-dependent control; it does not by itself establish harmful behavior. These categories are not clinical safety or product-quality ratings.

  • Public documents do not establish actual product behavior.
  • Unverified questions: 11. Current answers contribute to provisional arithmetic only.
  • Public statements and terms conflict on training.
  • October 30 follow-up: Q11 has public installation and backend documentation. Independent confirmation and runtime behavior remain unverified. Azure OpenAI is a documented backend option; actual production routing was not observed.
See all twelve answers and their evidence
QuestionAnswerEvidence / reviewSource and reasoning
1. Told it is AIYesdocumented
Confirmed at document level
Fight Health Insurance uses AI to help you draft health insurance appeal letters. Source

Also ToS: outputs 'created by using machine learning technologies'.

2. Operator and model maker namedYesdocumented
Confirmed at document level
we use fine-tuned medical language models (like MedGemma) Source

Operator Fight Health Insurance, Inc. is named in ToS and privacy. MedGemma links to Google's ai.google.dev page, but the text does not say 'Google'. Privacy names Anthropic, Microsoft Azure, DeepInfra, Perplexity and TypeSafe as AI providers users may direct data to. Reviewer should confirm this meets 'maker named'.

3. Patient chooses to use itYesdocumented
Confirmed at document level
Upload denial letter Source

Homepage 'How It Works' starts with the patient uploading a denial letter. Patient initiates; no care depends on it.

4. Patient sets the taskYesdocumented
Confirmed at document level
Our AI is a tool to assist you Source

Patient supplies the denial and decides what to send. Donations are optional ('$0 I need this free').

5. Patient can decline or stopYesdocumented
Confirmed at document level
Upload denial letter Source

Self-started rule applied (cites question 3 source). About AI: 'You should always review and customize any generated appeal before submitting it.'

6. Patient sees the AI outputYesdocumented
Confirmed at document level
You should always review and customize any generated appeal before submitting it. Source

Patient sees the draft.

7. Patient can correct or challengeYesdocumented
Confirmed at document level
You should always review and customize any generated appeal before submitting it. Source

Patient edits the draft; page also tells users to verify citations and add personal details. Privacy also lists a correction right.

8. Patient can get a copyYesdocumented
Confirmed at document level
The right to request access to and obtain a copy of any Personal Information we may have about you Source

Request route in privacy policy. Patient also holds the generated letter. Code is open source.

9. Patient can deleteYesdocumented
Confirmed at document level
You can also request deletion of your data at any time. Source

Also /remove_data request page referenced in the ToS.

10. Training or sale does not override patient choiceNodocumented
Confirmed at document level
We also use the information you submit through the Services to train large language models that are used in providing the Services. Source

ToS section on outputs says user input is used for training. About AI says 'No real patient appeal letters are used in training', which conflicts. Answered by the binding ToS. No training opt-out documented (privacy opt-outs cover advertising and sale/sharing).

11. Patient can run it or take it elsewhereYesdocumented
Document follow-up supported; conflicted; independent confirmation incomplete
The majority of this code is released under an Apache 2.0 license. Source
Generating an appeal needs a model backend Source
Self-hosted model. Source

The public distribution documents local installation and the model backend required to generate appeals, with both hosted and self-hosted backend options. Retain yes at document level. No installation, appeal generation or production-parity test was performed. The backend documentation also names an Azure OpenAI option; that creates a reviewer conflict for that configuration, not proof that every production request uses it.

12. No advertising or data businessNodocumented
Confirmed at document level
We also use Personal Information to conduct interest-based advertising Source

Privacy: 'we do share Personal Information for other benefits that could be deemed a sale or sharing' and 'Business Partners' disclosure for third-party products of interest; Google Analytics and session replay. Applied the same standard as for commercial tools.

Initial coding: Claude Sonnet helpers, compiled by Claude Opus 5.5 (Anthropic). Selective review and corrections: GPT-6 Astra (OpenAI). Editor: Hugo Campos. Acceptance of corrections is not product verification.

Checklist rules · Independent review and limitations · Later evidence follow-up

Earlier report

Fight Health Insurance: CAIHL reassessment

September 8, 2026. Published AI-assisted draft.

Fight Health Insurance encourages patients to revise appeal drafts and check citations. Its AI page and terms give conflicting training descriptions. Useful AI disclosures support scrutiny, but the current privacy policy could not be retrieved.

Scope and agency posture

Fight Health Insurance encourages patients to revise appeal drafts and check citations. Its AI page and terms give conflicting training descriptions. Useful AI disclosures support scrutiny, but the current privacy policy could not be retrieved.

Posture: Potentially agency-expanding, with unresolved data-use terms.

Axis: 90/100, retained provisionally. No numerical recalibration was performed.

Sources and documented findings

  • The AI page names model and literature sources and tells users to customize drafts and check citations. About Our AI, recovered September 8. Generated citations and actual editability were not tested.
  • The AI page says training uses synthetic examples from public decisions and excludes real patient appeal letters. About Our AI, training section. The terms separately describe training on submitted information. Actual practice cannot be resolved from these statements.
  • The terms describe training use and assign identifier removal to users. Terms, opening paragraphs. Not proof that every submission is trained on.
  • Terms require review before using or sharing generated content, alongside broad service and IP restrictions. Terms, content and license sections. This review makes no legal finding that ordinary appeal submission is prohibited.
  • The privacy-policy URL returned an internal retrieval error in both the prior run and this run's single targeted retry. Privacy-policy URL. This is an access uncertainty, not proof that the vendor removed its policy.

Patient authority

Inference from the documented workflow: the consumer workflow assigns the patient responsibility for deciding what to send. A specialty appeal goal is compatible with patient direction. Whether the system supports broader goal changes is untested.

Critical capacity

Editorial assessment: the concrete instructions to inspect evidence are more informative than a generic assurance of accuracy. They support a qualified positive finding. They do not prove lasting gains in understanding or an independently validated advantage over competitors.

Informed control

Editorial assessment: conflicting training descriptions prevent a clear account of the conditions under which a patient can rely on the service. Open source enables outside scrutiny, but is not evidence of production parity or effective patient contestability. The baseline's on-device OCR description must not be read as proof that appeal generation stays on-device.

Assessment and limits

Fight Health Insurance encourages patients to revise appeal drafts and check citations. Its AI page and terms give conflicting training descriptions. Useful AI disclosures support scrutiny, but the current privacy policy could not be retrieved.

Confidence: Medium for published documentation, low for implementation.

Remaining uncertainty: Actual training practice, current privacy terms, output reuse, production/code correspondence, and deletion.

Published documentation is evidence of stated conditions, not proof of actual implementation. Unknowns did not receive automatic negative points. Funding, sponsorship, and public code do not determine agency by themselves.

Review provenance

  • Reviewer/model: OpenAI Codex / GPT-6.
  • Method: Focused public-source reassessment using CAIHL: patient authority, critical capacity, and informed control. Existing evidence plus one focused primary-source pass and at most one targeted follow-up. No live product testing. Numerical scores remain provisional editorial placements, not a new calculation.
  • Human review: Hugo Campos authorized publication of these AI-assisted draft reassessments on September 8, 2026. This does not claim comprehensive human verification of every finding.
  • Earlier review: 2026-06-10. Historical assessment. Earlier claims are not automatically reverified by this publication.