When AI Says "I have been in similar situations": Synthetic Lived Experience in Peer-like Caregiver Support

The work names and characterizes a synthetic lived experience paradox: the same experiential language that may make AI support feel warm, relatable, and peer-like can also falsely position the system as someone with lived experience. Prior discussion of AI support systems often centers on immediacy, privacy, and nonjudgmental availability; this work isolates experiential language itself as a design tension to be examined. The paradox is proposed and argued within the ADRD caregiver context, drawing on comparison between community exchanges and model responses, making it a conceptual contribution.

Psycholinguistic analysis shows that peer responses used significantly more first-person and past-focused language than peer-like AI responses. It provides quantifiable language-level evidence for the intuition that human peer support relies on personal narratives, directly contrasted with AI responses. Evidence comes from comparing caregiver support exchanges from online communities with responses generated by three models (LLaMA, GPT-4o-mini, MedGemma) under peer-like prompting, with the abstract reporting the difference as significant.

Qualitative analysis identifies seven types of personal narratives in human peer support and shows that AI often captures their emotional work but can fabricate experiential grounding. It breaks the broad notion of personal narrative into distinguishable types and locates AI's closeness at the emotional level and its divergence at the level of experiential authenticity. Evidence comes from qualitative coding and comparison of real caregiver support exchanges and model responses, an interpretive rather than causal-experimental analysis.

The work proposes a narrative authenticity gap and argues that caregiver-support AI systems need mechanisms to distinguish supportive peer-like framing from fabricated lived experience. It translates the analysis into a design orientation: preserve warmth and validation while avoiding falsely positioning the model as an experiential peer. This is a design argument and research direction grounded in the reported findings, not a validated intervention effect.

The work addresses the online peer-support setting for family caregivers of people living with ADRD, analyzing caregiver support exchanges from online communities alongside responses from LLaMA, GPT-4o-mini, and MedGemma under peer-like prompting; its conclusions apply to understanding how peer-like AI positions itself linguistically in emotional support and inform designs that offer warmth and validation without fabricating experience.

The abstract does not report sample sizes, statistics, prompt design, or coding agreement, so the effect size behind the reported significant difference and the boundaries of the seven narrative types would need the full text; whether the behavior of these three models under peer-like prompting represents broader model behavior, and how well the framework applies to other caregiving or health settings, remain open questions.