HugoScore hugoscore.org

Full review

CareClinic Health Tracker + connected AI assistant CAIHL draft report

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

Agency checklist · October 30, 2026

Leans expands 9 of 12 yes

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

Assessed function: CareClinic MCP connector with ChatGPT as host

Astra review conflicted: OpenAI model in role record. November 10 follow-up prepared by an Anthropic model. Anthropic competes with OpenAI, the maker of the ChatGPT host.

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.
  • The independent review screened rules and selected evidence; it did not replicate all twelve answers for this tool.
  • November 10, 2026 accepted follow-up: Q7 changed from yes to no, giving 9/12, Leans expands. Q2 and Q10 stay yes with quotations added. Seven other yes answers gained their stored quotation. Prepared by Claude Fable 5.2 (Anthropic) from public documents and accepted by Hugo Campos. No second coder.
See all twelve answers and their evidence
QuestionAnswerEvidence / reviewSource and reasoning
1. Told it is AIYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
Choose the connection in a supported AI assistant or MCP app and review the requested access. Source

Connector page and privacy policy describe connected AI services.

2. Operator and model maker namedYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
CareClinic Software Inc., 43-45 Karachi Drive, Markham, Ontario, Canada, L3S 0B6, serves as the data controller Source
What can I ask ChatGPT about my health? Source
our language models that power ChatGPT Source

Yes retained. CareClinic names itself as the operator and names ChatGPT as the host. CareClinic's own pages do not name OpenAI. The model maker is named in the host's policy. No specific model is named.

3. Patient chooses to use itYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
CareClinic does not activate this connection automatically. You review the requested permissions and choose whether to connect. Source

Connection is never activated automatically; user chooses.

4. Patient sets the taskYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
Ask about schedules, symptoms, moods, medication entries, or patterns from what you recorded. Source

Patient asks about their own logs.

5. Patient can decline or stopYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
You can deny a connection or revoke it later. Source

Deny or revoke the connection.

6. Patient sees the AI outputYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
Review the answer. If you ask to log a symptom or mood, confirm or cancel before anything is saved. Source

Answers shown; check-in previews before saving.

7. Patient can correct or challengeNonot-documented
Changed to no; accepted by editor after November 10 document follow-up; independent confirmation incomplete
Fix anything that is wrong Source

Changed from yes to no on November 10, 2026. The documented fix covers a pending check-in before it is saved, shown as an example on a product page. No documented route lets the patient correct or contest the host AI's answers or summaries from the connector. The host rule lends host terms to Q8 through Q12 only, so the host's own correction route is not counted here. Whether that rule should reach Q7 is an open method question. Not documented counts as no. This is not a finding that corrections are refused.

8. Patient can get a copyYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
Access their data and obtain a copy Source

Access and obtain a copy; portability.

9. Patient can deleteYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
You may delete your account at any time through User Settings or by contacting [email protected]. Source

Delete account through settings or support.

10. Training or sale does not override patient choiceYesdocumented
Yes retained; quotation added in November 10 document follow-up; independent confirmation incomplete
CareClinic does not sell health information or use it for advertising or AI model training. Source
how you can opt out of our use of your Content to train our models Source

Yes retained. CareClinic says it does not sell health information or use it for advertising or model training. The ChatGPT host may use content to train models and documents an opt-out, which Q10 accepts. The host side is an opt-out, not off by default.

11. Patient can run it or take it elsewhereNonot-documented
Not independently replicated
Request portability of their data Source

CareClinic is vendor-hosted. The MCP connector works with several supported AI clients, but the underlying service stays at CareClinic. Portability is a stated right, with no self-hosting and no way to run the same function elsewhere.

12. No advertising or data businessNodocumented
Not independently replicated
For Free and Go users, to personalize the ads you see on our Services (subject to your settings) Source

The no comes from the ChatGPT host. OpenAI documents advertising to Free and Go users and de-identified sharing. CareClinic's own privacy policy says it does not sell health information or use it for advertising or AI model training. CareClinic's MCP page also says the AI host may retain results under its own privacy policy.

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

CareClinic Health Tracker + connected AI assistant: CAIHL review

Method v1.3 review, October 30, 2026

Agency finding: Unclear. Reason: benefit documented, constraint search incomplete.

Control: Who directs: 60 or 80 or 90, pending C3, C4. Toward the patient end, with the remaining conditions open. The patient chooses to connect, chooses ChatGPT as host, approves the OAuth scope, and asks their own questions about their own logs in a general-purpose assistant; CareClinic's tools are read-only and return deterministic data, and the patient can revoke access (E1, E2, E3, E7). No document shows CareClinic or OpenAI fixing an objective for the conversation. C3 is U: no document shows a block on questioning, and the connector's server instructions were not readable. C4 is U: the ChatGPT Terms say users 'must not use any Output relating to a person' for 'medical ... decisions about them' (E10), and whether that covers a patient's decisions about their own care is unresolved. R is N: OpenAI controls the model and its operating rules, and CareClinic controls the connector. Candidate set 60, 80 or 90 pending C3 and C4. Documentary only.

Disclosure (transparency gate): Service operator named: yes, in public documents. Model developer named: yes, in public documents. AI names its operator when asked: not tested (documentary review).

Evidence: Documentary review. Public sources only, no hands-on testing by HugoScore.

Review status: AI draft. No person has settled the findings. Before publication, Hugo Campos confirmed what was reviewed and the disclosure results, October 30, 2026. Not verification of the service's claims.

Coded by: Claude Opus 5.5 (Anthropic), the one eligible coder.

What it helps a patient do

  • Ask about your own logs in plain language (Understand, documented, not tested). The patient can ask a general-purpose assistant about their own logged symptoms, moods, medication schedule and adherence entries over a period they choose, and get counts, averages and changes without cause attribution (E1, E2, E6). Comparator: reading the same entries in the CareClinic app's lists and charts, where free questions combining entries are not documented. Documented potential; not observed.

What limits the patient

  • None supported in this review.

Strongest evidence against these findings

Against the candidate set: OpenAI's Terms clause on using Output for 'medical ... decisions about them' (E10) would put C4 = N and drop 90 if read to cover the patient's own decisions, and the connector's server instructions, which ChatGPT loads into the model (E7), were not read, so a steering instruction cannot be ruled out. Against the configuration itself: the only readable statement that Pro can add custom MCP servers is a two-month-old archive (E9), and if the live plan list now excludes individual plans, the patient path would run under different OpenAI terms and the configuration would need to be re-frozen. Against the Unclear posture: CareClinic's documents are specific, consistent and against its own interest on retention after revocation (E3), so with the host clause resolved and server instructions read, Potentially agency-expanding would follow.

Open questions

  • Current ChatGPT plans that may add custom MCP servers (live help page inaccessible).
  • CareClinic MCP server instructions and whether tool descriptions differ from the Smithery mirror.
  • Scope of the OpenAI Terms clause on medical decisions about a person, as applied to a patient's own care.
  • Whether ChatGPT shows CareClinic's confirmation card or any write tool in Pro developer mode.
  • Default ChatGPT model at the time of use and any health-specific behavior.
  • Claude-host configuration: needs a non-Anthropic coder.
  • Announced CareClinic proactive features: refresh trigger if deployed.

What changed from the earlier review

Freezing a named host (F23, F24) brought the host's binding terms into the review: OpenAI's clause on using Output for medical decisions about a person is a candidate restriction of unresolved scope, and the connector's server instructions could not be read, so the constraint search is incomplete and v1.3 gives Unclear.

Sources

Earlier review, September 28, 2026

The review below was prepared on September 28, 2026 with OpenAI Codex / GPT-6 (exact variant not recorded), before method v1.3. It is kept unchanged as history. Where it differs from the v1.3 review above, the v1.3 review is current.

Reviewed September 28, 2026. AI-assisted draft. Potentially agency-expanding; unscored.

Executive summary

CareClinic can supply authorized records to a separately governed AI assistant. Scoped access, confirmation before supported writes and revocation may expand agency; host interpretation and retained copies remain outside those controls. Unscored.

CAIHL question: who does this AI serve?

Patient-authorized record access with interpretation by a separately governed AI host. Primary user: Individuals using their CareClinic records through an AI assistant they connect. CareClinic hosts records and connector; the selected AI host processes returned context under its own terms. Consumer choice supports a patient-directed pathway, but does not by itself establish patient alignment. Vendor interests in subscriptions and continued use coexist with the patient's purposes.

What the service does and the scope of this review

Consumer CareClinic MCP connection plus the chosen AI host. The deterministic tracker alone and built-in image AI are not the evaluated AI deployment.

Patient-impact pathway

An AI assistant interprets authorized health records and can prepare supported check-ins, affecting understanding and the patient-held narrative.

Mixed HugoScore profile

Who sets and changes the goal?

Partial. Users authorize a host connection and choose questions; authority over the host interpretation was not tested. S1 S2

Can patients tell AI is involved?

Yes. The connector workflow explicitly involves an AI assistant chosen by the user. S1

Can patients meaningfully choose?

Partial. Scoped authorization and revocation are described. Revocation stops future access, not host-held copies. S1 S3

Can patients correct or challenge the output?

Partial. Supported symptom/mood check-ins use preview and confirmation; medication writes are disabled. These controls do not verify correction of AI conclusions. S2

Does it help patients understand or act?

Potentially. Authorized record context may support chosen questions and patterns; causal or treatment conclusions are not established. S1 S2

Can patients evaluate the conditions of use?

Partial. Documentation distinguishes connector and host responsibilities; users must assess both policies and actual granted scope. S1 S3

What happens to patient data?

Two governance layers. CareClinic describes minimized connector returns and separate host terms. Account-linked MCP usage events have no automatic expiry; deletion reconciliation is distinct from revocation. Built-in AI training commitments do not govern an independent host. [S3]

Who is left out or burdened?

Host-dependent. A CareClinic account is required; Premium is not required to connect. Premium adds deterministic trend detail. Host costs and accessible connector support vary. [S2]

Are the clinical boundaries clear?

Documented; behavior untested. MCP help describes consumer informational use, excluding diagnosis and HIPAA covered-entity/clinical use. Broad security language does not override this scope. [S2, S4, S5]

Who defined what good looks like?

Tracker evidence is not AI validation. The cited Skeens study evaluates marketplace apps using MARS. It does not establish clinical benefit or validate MCP-mediated AI interpretation. Patient influence on AI outcome selection is not established. [S6, S7]

Evidence boundaries

CareClinic’s clinical-evidence page frames tracker research as evidence for its app. The Skeens paper is a systematic marketplace search and app-quality assessment, not a trial of this connector or its AI host. Its ratings cannot establish patient health outcomes or AI accuracy. The review does not equate tracker engagement ratings with patient agency. The paper’s publisher page was inaccessible to this pass; the indexed abstract and PubMed record support the methodological distinction. Vendor evidence page · Study record

Key unknowns

  • Which host/model and privacy configuration each person uses
  • Actual permission and preview-confirmation enforcement across supported hosts
  • Correction and deletion of host-derived summaries or inferences
  • Independent evaluation of the combined AI workflow
  • Accessible use, costs and subgroup outcomes

Risks to investigate

These are plausible failure modes, not observed harms.

  • Treating connection revocation as erasure from every recipient
  • Mistaking descriptive record patterns for causes or treatment guidance
  • Extending tracker quality evidence or built-in AI privacy claims to the connected host
  • Unsupported generalization of vendor security claims to clinical deployment

Patient agency interpretation

Could support patient-selected questions about records and deliberate authorship of supported check-ins. Revocation cannot retract information already copied by the host; host behavior and downstream correction are untested. Critical reflection requires inspectable evidence; strategic action requires the ability to use the result for the patient's own purposes; neither is proved by engagement or vendor assurances alone. Documented access and write controls support patient choice. Agency still depends on the host, actual enforcement and whether interpretations can be challenged.

Publication recommendation

Publish as a source-backed AI-assisted draft. Retain a null numerical axis position: documented architecture is insufficient for a calibrated comparative number. Confidence: Medium for connector documentation, low for host behavior and outcomes (AI-assisted draft).

Evidence ledger

All sources accessed September 28, 2026. Vendor disclosures establish what is promised, not that controls work.

  • S1: CareClinic MCP — CareClinic. User-authorized records and assistant workflow.
  • S2: MCP help — CareClinic. September 3, 2026; previews, limits and consumer scope.
  • S3: Privacy policy — CareClinic Software Inc.. September 5, 2026; built-in AI versus MCP hosts, event retention and deletion.
  • S4: Terms of use — CareClinic Software Inc.. Consumer self-management, unverified records and account/subscription distinctions.
  • S5: Data security — CareClinic. February 12, 2026; controls and audits in progress.

Review provenance

criteriaSource: CAIHL-derived HugoScore framework and current qualitative review method

reviewer: AI-assisted public-source review prepared in OpenAI Codex

model: OpenAI Codex / GPT-6 (exact variant not recorded)

humanReview: Hugo Campos authorized review and publication. No comprehensive human verification is claimed.

reviewDate: 2026-09-28

limitations: Public-source documentary review only. No account creation, health-data upload, interaction tests, security audit, clinical evaluation or vendor interview. Host-specific and provider-specific behavior is unverified.