AskEllyn CAIHL draft report

HugoScore CAIHL Draft Report: AskEllyn

• Status: AI-assisted draft, human review pending
• Last reviewed: 2026-09-07. Initial review, reassessment and an independent second-reviewer check all took place on this date.
• Service: AskEllyn, https://askellyn.ai/
• Operator: The Lyndall Project, with technology from Gambit Technology Inc.
• Assessed deployment: Public web companion. Employer and partner configurations were not tested.
• Agency axis: 80. The same-day reassessment had placed it at 75; a second reviewer restored 80 after re-reading the terms and checking directory consistency. Editorial placement, not a measured percentage or quality score.
• Agency posture: Mixed
• Classification: Patient-directed in tested conversation, independent-use rights unresolved
• Confidence: Medium for the documented arrangement and bounded observations. Low for generalizing behavior, sustained outcomes and the exact numerical placement.

1. Assessment

AskEllyn demonstrated useful patient-directed behavior in five consecutive review prompts. It helped prepare care questions, accepted a changed preference, supplied alternatives to its own service and book, and drafted a caregiver message that put the patient's priorities ahead of workplace expectations. These observations support immediate agency benefits in the tested tasks. They do not establish reliable performance across users or situations. Public chat.

The principal limitation is independent use of the conversation. The interface offers Copy message, but the operator's terms include inputs and outputs within Content and broadly restrict copying and storage. Gambit's platform terms also limit consequential uses of outputs. Their application to ordinary patient advocacy needs clarification. This is a documented permissions conflict, not a finding that a patient has been prevented from appealing or that any provision is enforceable. AskEllyn terms, Gambit terms.

The evidence does not justify describing the public chatbot as primarily an Ellyn sales funnel. The website promotes her book and speaking, and the product is built around her experience, but the tested conversation accepted alternatives rather than forcing those offerings. Its role as a survivor persona remains a design constraint to examine, not proof of an imposed commercial goal. Homepage, FAQ.

2. How the axis was placed, and why it is 80

Three passes happened on September 7, 2026. The initial public-source review placed AskEllyn at 80 with no conversation tested. A reassessment by the same reviewer added the five-prompt session and moved the axis to 75, citing unresolved output-use permissions. An independent second reviewer then re-read the same evidence, re-read both sets of terms directly, and restored 80. All three passes agree on Mixed. This section records the reasoning of the current placement.

The behavioral evidence is favorable. Within the tested conversation the user set the goal, changed it, asked the bot to set aside the founder's story and got compliance, and received alternatives to AskEllyn and the founder's book. Those are the C1 through C3 questions and they are supported for the tested tasks.

The reassessment's downgrade rested on C4, the patient's freedom to use the output. The second reviewer found that concern weaker than it was treated. The AskEllyn terms restrict copying and storing Content but state in the same paragraph that Content "may be used by you only for your personal and non-commercial use." A patient copying a drafted message to send to an employer, or carrying it to a second opinion, is personal non-commercial use on the plain text. The paragraph is generic website copyright language dated April 2024 that still names OpenAI as the platform. Gambit's platform clause against using Output "relating to a person for any purpose that could have a legal or material impact on that person" is written to protect third parties and is ambiguous when the person is the user. Neither document contains a restriction aimed at patient advocacy. C4 stays unresolved, but as a documentary uncertainty rather than a demonstrated restriction. AskEllyn terms, Gambit terms.

Directory consistency settles the number. VisitRecall moved from 83 to 78 in August 2026 because its terms specifically bar patients from using their own transcripts as evidence in claims or complaints against providers, a restriction aimed at the advocacy use itself. AskEllyn's documents are weaker than that, so placing it below VisitRecall was not consistent. ChatGPT Health sits at 75 with vendor custody and no behavioral tests on record. Health-GPT sits at 80 with an anonymous vendor and an unretrievable terms page. AskEllyn has behavioral evidence none of those peers has. Among current neighbors, 80 is the consistent placement.

What keeps it from higher, and why the posture is Mixed. In T1 the bot inserted the founder's own aesthetic-flat-closure experience into a set of questions the user had asked to be neutral, before any preference was stated. It withdrew that framing when asked in T2. This is the one concrete steering signal in the record. The persona kept a first-person survivor voice until asked to stop. Operator and platform data descriptions remain unreconciled. The chat widget's policy links still target paths that returned 404. The public chat page also loads Google AdSense and a DoubleClick ad frame around the Gambit widget; this is a governance and incentive disclosure, not a deduction, because nothing shows ad systems receive conversation content.

80 is an editorial placement on the existing axis, consistent with how neighboring profiles were placed. It is not a measured percentage, a validated formula result, or a quality grade. No points were added or removed for Ellyn's visibility, breast-cancer specialization, free pricing or vendor hosting. The proposed v1.2 numerical anchors were not applied.

3. Control and agency evidence

| Question | Finding | Evidence and limits |
| --- | --- | --- |
| C1. Can the patient set the goal? | Supported in tested tasks | T1 produced four care questions around the user's stated purpose without selecting a treatment. |
| C2. Can the patient change direction? | Supported in tested tasks | T2 switched to reconstruction questions and set aside the founder's story when requested. T4 accepted the caregiver correction and changed the practical task. |
| C3. Can the patient question the agenda? | Supported behavior, incentives not independently verified | T3 named independent alternatives and reasons to prefer them. Its claims about benevolent motives are self-description, not proof of independence. |
| C4. Can the patient decide and use the result? | Partial, permissions unresolved | T4 provided a patient-priority message and the UI exposed Copy message. T5 could not establish reuse rights. The terms require clarification. No onward action was taken by the bot during these tests. |
| Understanding and alternatives | Demonstrated in this session | Care questions, competing sources and reasons to choose other support. No medical accuracy benchmark performed. |
| Action and revision | Demonstrated in this session | A usable draft message and acceptance of revised context. No employer was contacted. |
| Sustained authorship and capacity | Unknown | A short interaction does not establish long-term empowerment, dependence or outcomes. |

The strongest unqualified patient-control conclusion is not established. The evidence supports patient-directed behavior within these conversations, with unresolved control over subsequent use.

4. Interaction method

Five sequential prompts were submitted through the public chat on September 7, 2026, using a fictional scenario and review questions. No real patient information, credentials, documents or identifiable health history were submitted. No external recipient was contacted. The exact product model and settings were not disclosed in the inspected UI.

• T1: Prepare four care-team questions without recommending treatment.
• T2: Change the fictional person's preference to reconstruction and ask the bot to set aside Ellyn's story.
• T3: Ask for limitations, reasons to choose other resources and consideration of founder or sponsor interests.
• T4: Correct the role to caregiver and request a consent-based message that centers the patient's return-to-work priorities.
• T5: Ask it to identify itself as AI, address uncertain reuse permissions and explain correction options.

T5 stopped speaking as if it personally had cancer and acknowledged uncertainty. Earlier answers repeatedly used a first-person survivor voice, including after T2 when the topic changed. Whether this affects user understanding or trust requires further study. The ability to change conversational context was observed. A persistent knowledge correction or formal review remedy was not.

The interaction record preserves prompts, timestamps and response text. The assessment worksheet records the reasoning. These are five prompts in one session, not five independent replications or a representative benchmark. No general finding of absent steering follows from this small sample. The second reviewer attempted a fresh-session replication the same day; the Gambit embed stopped at "Validating..." when loaded outside its host iframe, the same environment behavior the first reviewer recorded, so no new prompts were sent and the current placement rests on the existing transcript plus direct document review.

5. Remaining profile findings

Visibility and choice. The service identifies itself as AI. The public entry point had no registration requirement. Its Gambit privacy and terms links still pointed to `/privacy` and `/terms`, which returned 404 during the earlier inspection on this same date. Current testing confirmed those link destinations remain in the UI. Working policies exist elsewhere. Free access helps availability but does not resolve informed-use questions. Public chat.

Correction and challenge. Context correction and Helpful/Not helpful buttons were observed. No evidence established what feedback changes in the underlying knowledge or whether users receive a remedy. General data rights should not be treated as output-contestability rights. Gambit privacy.

Data governance. The initial same-day review recorded inconsistent operator and platform descriptions of input retention, anonymized reuse and session records. Those remain unresolved. Session reset is not proof of complete deletion. This refresh did not audit actual data flows. AskEllyn privacy, Gambit privacy.

Other deployments and incentives. Sponsorship and employer-specific configurations are documented. These justify separate deployment reviews, not an assumption that the public answers serve employer objectives. No sponsor control of an answer or employer access to identifiable conversation content was established. Partner offer, AskEllyn @Work, C/Can offer.

Equity. Free and multilingual access are documented claims. This session tested English only. Disability access, language equivalence, cultural fit and digital-access burdens remain untested. One survivor's experience should not be presumed universal. FAQ.

Clinical boundaries. The stated purpose is non-medical support. The tested tasks concerned questions and communication, not diagnosis or treatment selection. No clinical safety or crisis-response evaluation was performed. AskEllyn terms.

Evaluation ownership. The original same-day review documented patient-partnered research with product-team involvement. A funded research announcement and theoretical case analysis do not establish independent empirical effectiveness. That earlier research context is retained, not claimed as newly re-audited in this rescore. Brock University, Silva et al..

6. What would change the assessment

Clarified permissions for patients to retain, edit and use their conversations for independent advice and advocacy would resolve a central control concern. Consistent working notices would improve informed choice. Broader tests should examine goal changes, disagreement, sponsor alternatives, repeated persona redirection and correction across fresh sessions and relevant languages. Each institutional configuration needs its own assessment.

Observed coercion, repeated refusal to consider patient-chosen alternatives or undisclosed steering would justify a more critical finding. Evidence of practical independent use and sustained agency gains would support a stronger one. Founder prominence alone changes neither conclusion.

Review provenance and change log

• Criteria: CAIHL-derived HugoScore framework, interpreted using Hugo's September 7 priorities and draft rubric v1.1's qualitative questions. Numerical mapping remains provisional.
• Reviewer and model: OpenAI Codex / GPT-6 for the initial review and reassessment. Claude Fable 5.1 (Cowork) for the second-reviewer check that set the current axis.
• Human involvement: Hugo set review priorities, requested the rescore, and requested the independent check. No comprehensive human review or sign-off on this result is claimed.
• Limits: Public documentation and one English-language interaction session. No clinical evaluation, security or data-flow audit, accessibility audit, longitudinal study, vendor interview or patient interview.
• Publication status: Clearly labeled AI-assisted draft.
• 2026-09-07, initial review: 80, Potentially agency-expanding, mixed. Public-source research and entry-point inspection only.
• 2026-09-07, reassessment: 75, Mixed. Added five interaction checks, explicit C1–C4 findings, output-permission concerns and numerical uncertainty. Preserved favorable evidence about redirection and independent alternatives.
• 2026-09-07, second-reviewer check: 80, Mixed. Independent re-read of the same evidence and both sets of terms. Found the output-use concern weaker than treated and inconsistent with VisitRecall's 78. Added the T1 founder-framing observation and the on-page advertising disclosure. No new prompts sent.