JMIR Publications

ABSTRACT

Background:

Loneliness in community-dwelling older adults is associated with adverse health outcomes. Routine social contact can help, but effective methods are difficult to deliver at scale. Calls from trained callers are effective but resource-intensive, and app-based tools ask older adults to adopt unfamiliar technology. Artificial intelligence (AI) voice agents can place calls at scale, but current systems use synthetic voices unfamiliar to the recipient. Whether a weekly call in a voice cloned from a participant’s own family member or friend can be sustained is not known.

Objective:

We assessed the feasibility of a weekly AI wellness telephone call using a participant-nominated cloned familiar voice in adults aged 65 years and older, and characterized the participant experience, including any association with the relationship to the voice donor.

Methods:

This was a single-arm, single-site, 8-week feasibility pilot in which an AI agent placed weekly check-in calls by telephone. Of 49 enrolled adults, 36 began the protocol. Seven progression criteria and a zero adverse-event safety floor anchored the feasibility judgment. We characterized the experience through participant surveys, behavioral engagement, and the conversational topics participants raised, using pre-specified mixed-effects models and thematic analysis.

Results:

The enrollment-to-start rate was 94.7% (36 of 38). Five of the seven progression criteria and the zero adverse-event floor were met in the green zone, and two fell into the amber zone. All 36 starters were retained through week 8 and completed 284 of the 288 scheduled calls (98.6%). After the paid protocol ended, 8 completers (22.2%) placed 10 further calls on their own in weeks 9 to 11. Self-reported isolation and loneliness moved favorably without reaching significance (isolation item odds ratio 0.35, 95% CI 0.11 to 1.07). The experience varied with the participant’s relationship to the voice donor, with sibling and adult-child voices the most favorable. Small cells (sibling n=4)and a false discovery rate q of .27 make this pattern hypothesis-generating only (Kruskal-Wallis (P=.04).

Conclusions:

A weekly AI wellness call in a cloned familiar voice was feasible to deliver and sustain in community-dwelling older adults. Self-reported loneliness did not change significantly, and without a control group no change can be attributed to the intervention. Participants reported friction with the identity of the voice rather than with the agent’s persona or memory. A powered trial should test the relationship to the voice donor as a planned factor and separate the acoustic and linguistic dimensions of voice identity into distinct arms. The main barrier to scale was donor recruitment during onboarding.