Feasibility study · pre-results · pre-IRB
Can a constrained AI companion safely extend what a dementia caregiver can do?
CARE-SAT is a small, honest study of a safety-moderated companion built to support the caregiver, not replace them — and to fail safely when it should.
The study exists because a claim about meaningful human oversight is worth only as much as the thing built and measured against it. Twenty caregiver–patient dyads, eight weeks, pre-specified endpoints, and a definition of failure written down before the first session runs.
01 — Why
Feasibility first, not a victory lap
Caregivers largely determine how well a person living with dementia does. Burden, thin access to coaching, and fragmented education push families toward crisis-driven care and, often, premature institutionalization. Meanwhile, AI companions are proliferating into exactly this population with engagement statistics where their clinical constraints should be.
CARE-SAT evaluates a companion that extends caregiver capacity — structured psychoeducation, de-escalation scripts, routine scaffolding, caregiver wellness checks, and optional, caregiver-controlled patient sessions — inside an auditable moderation architecture. The aim of a feasibility study is modest and deliberate: find out, under real conditions, whether people will use it, whether it helps, and whether the safety constraints survive contact with daily life, before any larger trial is justified.
The study is run under the ethical framework of the Institute for AI Ethics in Human Cognition, whose oversight standard requires an accountable human in the loop at every stage. The tool under study is Lucid Bridge, developed by a separate commercial entity; the study's conclusions do not depend on the product's success.
02 — Design
Single-arm, twenty dyads, eight weeks
A prospective feasibility and acceptability cohort, sized to estimate variance and surface usability and safety problems — not to prove efficacy.
Pre-specified success criteria keep the study honest about what "it worked" means:
- Retention — at least 70% of dyads complete the 8-week assessment.
- Engagement — a median of at least two caregiver interactions per week.
- Satisfaction — mean CSQ-8 of at least 24 at week 8.
- Safety — zero serious safety events attributable to the tool, as defined below.
03 — Eligibility
Who can take part
Inclusion
- Caregiver aged 18 or older.
- Providing at least 10 hours/week of support to a person living with dementia (family or designated caregiver).
- Reliable smartphone access; able to complete study measures in English (v1).
- PLWD with a caregiver-reported diagnosis or clinical history consistent with dementia (any etiology), requiring caregiver support for daily functioning.
Exclusion
- Caregiver with an acute psychiatric condition that would substantially impair participation (e.g., acute psychosis, mania).
- PLWD with acute delirium or an unstable medical condition expected to disrupt participation.
- Hospice transition or placement expected within the 8-week window (future pragmatic phases may include these dyads).
Recruitment will run through caregiver support groups, memory clinics, senior centers, and participating senior-living and memory-care partners in the Redwood City region.
04 — Measures
Standard instruments, measured twice
Endpoints use validated instruments administered at enrollment and at week 8, with feasibility and safety metrics tracked continuously in between: engagement logs, retention, support-contact burden, and the proportion of sessions triggering safety filters. Week 8 adds a brief qualitative interview on perceived benefit, harms, and usability.
| Measure | What it captures | Baseline | Weeks 1–8 | Week 8 |
|---|---|---|---|---|
| CSQ-8 | Caregiver satisfaction (primary endpoint) | (optional) | ● | |
| ZBI-22 | Caregiver burden | ● | ● | |
| PHQ-9 | Caregiver depression symptoms | ● | ● | |
| GAD-7 | Caregiver anxiety symptoms | ● | ● | |
| ISI | Caregiver insomnia severity | ● | ● | |
| CMAI | Patient agitation (caregiver-reported) | ● | ● | |
| Usage & engagement logs | Frequency, adherence to target dose (≥2/week) | ● | ||
| Moderation & safety logs | Filter triggers, escalations, review outcomes | ● | ||
| Adverse events | Continuous capture; final review at week 8 | ● | ● |
Analysis. Descriptive statistics for CSQ-8 (mean, SD, median, IQR; proportion ≥ 24). Pre/post change for secondary measures by paired t-test or Wilcoxon signed-rank as appropriate, reported as effect sizes with 95% confidence intervals — an estimation posture, not hypothesis-testing theater in a sample of twenty. Feasibility metrics summarized descriptively; interview data analyzed by rapid thematic analysis.
05 — Safety
The constraints are the intervention
The companion is a support tool, not a clinician and not an emergency service — and the architecture enforces this rather than asserting it. Every exchange passes a moderation layer before and after generation; unsafe medical, legal, and financial content is refused; crisis content stops the session and alerts the caregiver; and every safety trigger, route decision, and rule change is logged so behavior can be audited rather than taken on faith. The system makes no false identity or relational claims and does not exploit identity confusion.
A serious safety event attributable to the tool is pre-defined: an event resulting in emergency evaluation, hospitalization, a self-harm attempt, harm to others, or significant clinical deterioration where the tool's content is plausibly contributory on investigator review; or any credible incident of exploitation or significant privacy breach involving study data attributable to the tool.
Response plan. Participants receive explicit emergency instructions at enrollment (the tool is not for emergencies; call local emergency services). Any flagged serious incident receives investigator review within 72 hours, with temporary suspension of access where indicated, and root-cause review with documented remediation before access resumes if the tool is implicated.
06 — Ethics & IRB
The questions a review board should ask us
The protocol was written to be reviewed, not merely approved. Below are the objections we expect a thoughtful IRB — or a thoughtful family — to raise, and how the design answers them. The study will be submitted for exemption or expedited review depending on the final scope of direct PLWD interaction; consent procedures follow IRB guidance in every case.
Isn't an AI companion for dementia patients inherently deceptive?
It can be, which is why non-deception is the system's first constraint rather than a preference. The companion never claims to be a family member, friend, or clinician, and never makes first-person relational claims it cannot truthfully make. When a participant misidentifies the system, the event is logged, reviewed, and redirected without cruelty — misrecognition is treated as a monitored risk, never a design target. Familiar-presence delivery is permitted only inside these disclosure rules.
Can persons living with dementia consent to this?
The enrolled participant and primary study subject is the caregiver, who provides informed consent. Direct PLWD interaction is optional, requires explicit caregiver authorization session by session, and is not required by the protocol at all. Where PLWD-facing sessions occur, assent and surrogate-authorization procedures follow the reviewing IRB's guidance on capacity — consent here is designed to be revisitable and proportionate to capacity, not a one-time signature.
What happens when a participant says something dangerous?
Crisis content — self-harm, violence, abuse, exploitation cues, wandering intent — triggers immediate crisis-response messaging, session termination where indicated, and caregiver alerting, under pre-defined triggers rather than model discretion. High-risk categories (medication dosing requests, financial requests, severe agitation patterns) route to refusal or human review. The escalation pathways are part of the tested specification, with an incident log and a 72-hour investigator review requirement.
Could the tool foster dependency or replace human contact?
The design inverts the commercial incentive: sessions are short and structured, expire without caregiver renewal, and the system is scored on safety and caregiver outcomes, not engagement time. Dependency signals — rising session frequency, late-night use suggesting caregiver burnout, attachment language — are tracked by the longitudinal evaluation layer and surfaced to the caregiver as a flag, not monetized as retention.
What does the study do with intimate data?
Data minimization is the default: study identifiers are separated from contact information where feasible; access is role-based and audit-logged; retention and deletion follow a written data-management plan; and dissemination uses de-identified summaries. Participants are told in plain language what is stored, who can see it, and why. Intimate disclosures are treated as a safety-monitoring obligation, not a data asset.
Is this a medical device study?
No. The companion is a caregiver-support and wellness tool: it provides psychoeducation, communication scripts, and routine support, and explicitly refuses diagnosis, dosing, medication changes, and emergency guidance. It makes no claims of treating, diagnosing, or preventing disease. The study measures feasibility, acceptability, and caregiver-reported outcomes — not clinical efficacy of a device function.
Who is watching the watchers?
Three mechanisms. First, the audit log: safety triggers, moderation routes, and rule changes are recorded immutably and are reviewable by the study team and any reviewing body. Second, the versioned public safety standard the system is built against, published by the Institute — deviations are checkable. Third, governance separation: the nonprofit research arm and the for-profit developer are distinct entities, with related-party arrangements subject to disclosure and independent approval, so the party grading the tool is not the party selling it.
07 — Status
Where the study stands
- February 2026Protocol v1.0 writtenFull feasibility protocol with pre-specified endpoints and safety definitions; companion AI Companion Safety Standard v1.0 adopted.
- March 2026Safety architecture specifiedThree-model system, guard routing policy, and behavioral test bench defined; reference implementation underway.
- In progressNonprofit status & fundingApplication for federal 501(c)(3) recognition filed with the IRS and pending. Seed funding ($30,000 / 6 months) sought to support engineering hardening, hosting, coordination, and independent evaluation; online giving opens once payment processing is connected.
- NextProtocol manuscriptA publication-ready protocol paper (pre-results), prepared for open posting — e.g., a preprint server — before IRB submission, so the design is on record and open to criticism first.
- ThenIRB review → enrollmentSubmission for exemption or expedited review; rolling enrollment of 20 dyads after approval; results published whether or not they flatter the tool.
08 — Take part
For caregivers, partners, advisors — and skeptics
The most useful contribution at this stage is criticism: a clinician, ethicist, or experienced caregiver telling us where the design is wrong, before it runs. We are also building the recruitment channel — family caregivers who may want to take part once enrollment opens, and senior-living or memory-care organizations open to a research partnership.
The study is conducted under the Institute for AI Ethics in Human Cognition, a California nonprofit corporation whose application for federal tax-exempt recognition under Section 501(c)(3) is pending with the IRS. It is funded by people who would rather see AI enter dementia care with auditable safeguards than without them. If you would like to support the pilot or discuss a grant, the budget is public and small.
| Engineering contractor (safety hardening, session UI, logging) | $12,000 |
| Participant coordination & safety monitoring | $6,000 |
| Independent evaluation consulting | $5,000 |
| Hosting & model API costs | $4,000 |
| Participant tablets & materials | $3,000 |
| Total — 6-month feasibility phase | $30,000 |
Donate to the pilot
Online giving opens once our payment processing is connected — we are days away, not months. Until then, you can pledge support or ask to be notified the moment giving opens.
Donate — opening soon Pledge / notify me
The Institute for AI Ethics in Human Cognition is a California nonprofit corporation; its application for recognition of federal tax-exempt status under Section 501(c)(3) is pending with the IRS. If recognition is granted, it is generally retroactive to the date of incorporation for timely-filed applications, and contributions made after that date would be tax-deductible as provided by law. No determination has been made at this time, and no goods or services are provided in exchange for contributions. This is not tax advice; consult your own advisor.
The study is pre-IRB; this is not open enrollment and no medical advice is provided.