HugoScore hugoscore.org

Full review

MAIA 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: Patient-provisioned MAIA private record agent. A shared demo or someone else's hosting account is a different trust arrangement.

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: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12. Current answers contribute to provisional arithmetic only.
  • October 30 accepted follow-up: Q12 changed from yes to no; 10/12 remains Expands. Personal AS deletion has public-code support. The original coder caveats are historical. All independent-confirmation flags remain because this later source pass is conflict-disclosed, not product testing.
See all twelve answers and their evidence
QuestionAnswerEvidence / reviewSource and reasoning
1. Told it is AIYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
a private AI that works only for them Source

The product is named Medical AI Assistant and describes its Private AI agent and the public AIs it can call.

2. Operator and model maker namedNonot-documented
No retained after bounded document follow-up; conflicted; independent confirmation incomplete
Quotation not recorded; evidence repair pending.

The code author (Adrian Gropper, About page) and host (DigitalOcean) are named, and a self-hosting patient is the operator. But the default private AI is named only as 'Qwen3.8-Max, an open model DigitalOcean hosts' (fallback GPT-oss-120b), with no model maker stated in the sources read. Public AI makers (Anthropic, OpenAI, DeepSeek, Gemini) are named in the design doc.

3. Patient chooses to use itYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
GET STARTED creates your MAIA and starts your private AI at once Source

The patient starts it. The README says to pay for hosting with your own credit card.

4. Patient sets the taskYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
MAIA introduces a patient-controlled Private AI agent that sits between a patient's complete health record and public AIs. Source

The patient chooses what to ask, which records to include and which AI to use. The private AI drafts the summary and rules, and the patient confirms them.

5. Patient can decline or stopYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
GET STARTED creates your MAIA and starts your private AI at once Source

Self-started tool rule: Q3 is yes and no care depends on it. Source is the Q3 source.

6. Patient sees the AI outputYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
Patients can see and edit chats, share them with clinicians, and manage current medications, a patient summary Source

Patient sees AI chats, the AI-drafted Patient Summary and the medication list.

7. Patient can correct or challengeYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
MAIA gives you the opportunity to edit and verify whatever you or others might see in the chats. Source

Patient Summary and Current Medications require patient verification, with links to the source PDF pages.

8. Patient can get a copyYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
Your folder holds your own copy of everything. Source

The MAIA folder on the patient's computer holds the summary PDFs, rules, request log and records. Chats can be saved locally.

9. Patient can deleteYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
This permanently deletes your cloud data for {userId}. Source

Destroy Account deletes Spaces files, knowledge base, agent, sessions and the user document. This text is from the account lifecycle doc for the full edition. The Personal AS edition design doc also describes account deletion.

10. Training or sale does not override patient choiceYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
DigitalOcean does not use this data to train, retrain, or fine-tune any model Source

For DigitalOcean-hosted private AI models. Public AIs are off until turned on and then follow each provider's route. DigitalOcean's page notes a 30-day mandatory retention for one named Anthropic model.

11. Patient can run it or take it elsewhereYesdocumented
Document follow-up supported for named configuration; conflicted; independent confirmation incomplete
For full control, host your own MAIA on your own account. Source

Open source (AGPL-3.0 per LICENSE) with a Deploy to DigitalOcean button. The code author needs no access to data.

12. No advertising or data businessNodocumented
Correction accepted by editor; conflicted document review; independent confirmation incomplete
This link to DigitalOcean gets up to $25 in credit to support our costs. Source
We work with third-party advertising partners to show you ads that we think may interest you. Source

Hugo accepted Q12 no on October 30 after the follow-up review. The public FAQ promotes a hosting referral that earns project credit. DigitalOcean also documents advertising to its users. This applies the current advertising rule; it is not evidence that MAIA sells health records. The FAQ covers the older full edition, so host advertising provides a separate basis for the patient-paid DigitalOcean configuration. Provider account/website marketing versus clinical-content processing remains a directory-wide scope question, not a tool-specific exception.

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

MAIA: CAIHL reassessment

September 8, 2026. Published AI-assisted draft.

MAIA describes patient-directed record interpretation with important hosting qualifications. Patient authority over questions does not establish privacy from a demo operator. Actual data handling, correction, and deployment controls remain untested.

Scope and agency posture

Patient-provisioned MAIA private record agent. A shared demo or someone else's hosting account is a different trust arrangement.

Posture: Strongly agency-expanding, with demo-host-trust, security-maturity, and technical-burden caveats.

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

Sources and documented findings

  • The live welcome and raw README fetches failed, but the targeted GitHub repository page rendered current documentation. It describes editable chats, patient-verified medications and summaries, source-page links, private instructions and optional public-model routing. https://github.com/agropper/self
  • The same README says the hosting account owner can likely control data access and recommends personally provisioned hosting. It separates code verification from operator access. These are architecture claims, not an operational audit. https://github.com/agropper/self

Patient authority

Inference from the documented workflow: User-selected documents, editable instructions and verified summaries support direction over the assistant's context. Caregiver or third-party hosting must not be mislabeled as control by the patient.

Critical capacity

Editorial assessment: Source links and required reconciliation offer concrete mechanisms for checking interpretations and preparing care discussions.

Informed control

Editorial assessment: The trust model is unusually candid. Pseudonymization does not guarantee anonymity, and choosing a public model introduces its own processing conditions.

Assessment and limits

The strongly positive design assessment remains supported for personally controlled provisioning. Openness helps inspection but supplies no automatic score credit or proof of secure operation.

Confidence: Moderate for documented design, limited for current live deployment.

Remaining uncertainty: Actual deletion, host isolation, public-model data handling and usability remain untested

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: 2026-06-26. Historical assessment. Earlier claims are not automatically reverified by this publication.