This study aimed to design and evaluate a personalized AIGC-based chatbot to support older individuals with cognitive impairment in memoir writing and emotional support. The system dynamically adapts its interaction mode based on users’ MMSE scores, offering keyword prompts for those with mild impairment and active narrative generation for those with severe cognitive decline. The chatbot identifies and responds to emotional needs such as loneliness and the desire for achievement through KE principles. Following system design, a user evaluation study assessed effectiveness.

This study recruited 20 participants aged 67‐80 years. Baseline MMSE scores spanned the range of mild to moderate cognitive decline, and there was individual variation in baseline memoir quality (as assessed by a psychiatrist) as well as in attitudes toward and proficiency in technology use. Participants had 3‐12 years of education, reflecting the characteristics of an older population with diverse educational backgrounds. Detailed participant information is provided in Multimedia Appendix 5. There were no participant dropouts throughout the entire experimental process.

Baseline comparability between the experimental (n=10) and control (n=10) groups was evaluated using paired t tests and standardized mean differences (SMDs). No statistically significant differences were found on any baseline variable (all P>.05), and all SMDs were negligible (|SMD|≤0.20), confirming successful group equivalence prior to the intervention (Table 1).

As shown in Table 2, normality of difference scores was assessed using the Shapiro-Wilk test, separately for within-group pre- and postdifferences and for the matched-pair between-group differences that form the basis of the primary analysis. Within-group differences were normally distributed only for PA in the experimental group (W=0.866; P=.09); all remaining within-group differences violated the normality assumption (Ps<.05). Among the matched-pair between-group differences, PA (W=0.909; P=.28) was normally distributed and was analyzed with a paired t test; NA (W=0.843; P=.048) and psychiatrist-rated quality (W=0.833; P=.04) violated normality and were analyzed with the Wilcoxon signed rank test.

Within-group changes were examined separately for each group; all tests were 2-tailed (α=.05). In the experimental group, PA increased significantly from pretest (mean 23.50, SD 4.58) to posttest (mean 25.30, SD 4.74; Δ=1.80, 95% CI 0.55-3.05; t₉=3.25; P=.01; Cohen d=1.03); the control group showed no significant change (Δ=−0.20; W=10.0; P=.54; r=−0.29). NA decreased significantly in the experimental group from pretest (mean 22.10, SD 5.22) to posttest (mean 18.40, SD 3.69; Δ=−3.70, 95% CI −7.19 to −0.21; W=3.0; P=.01; r=−0.89); the control group showed no significant change (Δ=0.50; W=0.0; P=.09; r=1.00). Psychiatrist-rated memoir quality, assessed from participant-generated content only, improved significantly in the experimental group from pretest (mean 2.775, SD 0.63) to posttest (mean 2.95, SD 0.73; Δ=0.175, 95% CI 0.054-0.296; W=0.0; P=.03; r=1.00); the control group showed no significant change (Δ=−0.05; W=2.5; P=.42; r=−0.50). Full within-group results are presented in Table 3. Additionally, Multimedia Appendix 6 presents separate ratings for AIGC.

Between-group comparisons were conducted on matched-pair differences rather than independent-samples tests, because participants had been allocated within matched pairs, and each pair therefore constitutes the appropriate unit of comparison. Paired t tests were used for PA and psychiatrist-rated memoir quality, and the Wilcoxon signed rank test was used for NA, based on the normality of the matched-pair differences (Table 2). The experimental group showed significantly larger improvements than the control group across all 3 outcomes: PA (matched-pair Δ=2.00, 95% CI 0.31-3.69; t₉=2.68; P=.03; Cohen d=0.85), NA (matched-pair Δ=−4.20, 95% CI −7.59 to −0.81; W=2.0; P=.008; rank-biserial r=−0.93), and psychiatrist-rated memoir quality, assessed from participant-generated content only (matched-pair Δ=0.225, 95% CI 0.093-0.357; W=0; P=.02; rank-biserial r=1.00; Table 4).

To account for multiple comparisons across the 3 prespecified coprimary outcomes (PA, NA, and psychiatrist-rated memoir quality), we applied Benjamini-Hochberg correction (Q=0.05) to the 6 hypothesis tests involving the experimental condition (3 within-group and 3 between-group matched-pair comparisons). Baseline equivalence tests in Table 1 were excluded, as these test for nondifference and are not part of the confirmatory hypothesis family. All 6 tests remained significant after correction (q≤0.031; Table 5). The 3 within-group control-condition comparisons, reported descriptively, remained nonsignificant (P≥.089).

System usability was assessed using the simplified SUS [30], in which odd-numbered items are positively worded and even-numbered items are negatively worded; even-numbered items were reverse-scored prior to analysis. The mean SUS total score among intervention group participants was 92.2 (SD 5.1; range 82.5‐97.5), which exceeds the widely used acceptability threshold of 68 and falls within the “excellent” range (>85) for dementia-assistive technologies [31,32]. Individual scores ranged from 82.5 (Participant 10) to 97.5 (Participant 6 and Participant 7), with no participant scoring below the acceptable threshold. Subscale analysis following Orgeta et al [22] yielded a mean effectiveness score (items 1, 3, 5, 7, and 9) of 4.80 (SD 0.19) and a mean simplicity score (items 2, 6, 8, and 10) of 4.47 (SD 0.40), both on a 1‐5 scale, indicating consistently high perceived effectiveness and adequate simplicity across participants.

The experimental group showed greater improvement than the active control group across PA, NA, and psychiatrist-rated memoir quality. Because the control condition was matched to the intervention condition in session duration, number of prompting turns, frequency of audio feedback, device exposure, and researcher behavior, the observed differences are less likely to be explained solely by interaction time or device novelty. However, the fidelity of the manually delivered control condition was not independently assessed, and the study was not designed to determine which specific component of the AIGC intervention—personalized prompting, memory-anchor retrieval, narrative generation, or their combination—accounted for the observed differences. The findings should therefore be interpreted as preliminary evidence concerning the integrated intervention rather than as proof of the effectiveness of any single system component.

Subgroup patterns offer additional insight. In the present sample, 12 of 20 participants (experimental: n=6; control: n=6) had MMSE scores below 20 (range 11‐18), triggering the closed-ended, high-proactivity interaction mode; the remaining 8 (experimental: n=4; control: n=4) had MMSE scores of 20‐23, triggering the open-ended, keyword-based mode. Participants in the lower MMSE tier appeared to benefit more from AI-driven narrative generation, while those in the higher tier responded better to keyword-based prompting. This is consistent with cognitive load theory [33]: when residual capacity is limited, proactive narrative scaffolding reduces the compositional burden and allows users to engage with their memories without being overwhelmed. Users with greater preserved capacity may find fully automated narrative generation less engaging, though the mechanism remains speculative. These patterns tentatively suggest that cognitive severity may shape the appropriate level of AI proactivity, and that a uniform interaction mode would be inadequate for this population. Given the small n within each tier (6 and 4 per group), these observations are descriptive and should not be interpreted as subgroup-level statistical conclusions.

First, most participants reported some initial trial and error during the first 1 or 2 conversational turns within the session, but that they became more familiar with the system as the number of turns increased. (Throughout this section, references to “uses” or “times” refer to individual conversational turns or exchanges within the single, approximately 10-minute session, not separate sessions on different occasions.) This pattern of initial difficulty followed by increasing fluency recurred throughout the interviews, particularly with regard to the timing and triggering of voice commands. Many participants mentioned that during their first few turns they were unsure “when they should speak and when they should wait for the system to finish speaking,” and this uncertainty clearly diminished as the session progressed.

Second, in terms of emotional experience, participants’ descriptions showed a fairly consistent positive tendency. Most described the act of “telling their story” itself as a process of being taken seriously and being listened to. Some participants compared this with their everyday communication with family and friends, feeling that the system “does not interrupt” and “does not get impatient,” which they saw as setting it apart from interpersonal communication.

Next is the issue of autonomy. This was the theme discussed in the greatest depth in the interviews, and the one on which participants’ attitudes diverged most clearly. Most participants felt that even though the system organized and rewrote their spoken content, the resulting story still “belonged to them,” because the core facts and emotions came from themselves and the system only played a supporting role in “polishing” or “recording” it.

However, a small number of participants (such as Participant 6 and Participant 8) expressed more nuanced reservations, noting that the wording the system produced after rewriting sometimes differed from their usual way of expressing themselves. They hoped to have the opportunity to check or revise the rewritten portions rather than simply accepting the final text passively.

Regarding the system’s design of adjusting its prompting style according to cognitive state (open-ended prompts vs stepwise closed-ended prompts), participant feedback showed a divergent trend related to their self-reported cognitive burden. Among participants with lower cognitive burden (corresponding to the MMSE stratum of 20 to 23), some felt that stepwise prompting was “a bit wordy” and preferred to speak directly and freely, feeling that their train of thought was interrupted.

Participants with higher cognitive burden (corresponding to an MMSE score below 20), by contrast, were more inclined to report that stepwise prompting “was a huge help,” reducing the pressure of organizing their language and easing the anxiety of “not knowing where to start.”

This divergence corroborates the original intent behind this study’s MMSE stratified design, namely, that greater cognitive burden calls for more structured guidance. It also suggests that future versions could allow users to actively switch the density of prompting during use rather than having it entirely preset by the system, so as to accommodate the desire for autonomous expression among participants with milder impairment. It should be noted that because this study is only a preliminary formative study with a small sample size, it is difficult to support adequate subgroup analysis, and the description above should be treated only as an indication of a trend.

With regard to willingness to continue using the system, all participants expressed a positive attitude, but this positive attitude often came with specific conditions. For example, whether the system could remember what had previously been said (to avoid the fatigue of repeating the same content) and whether interest would wane once the novelty faded.

One participant, Participant 8, linked the willingness to continue using the system to a deeper motivation, that of “leaving something behind for later generations,” which goes beyond a purely instrumental evaluation of the tool and positions the system as a medium for recording and passing things on rather than simply a companion or entertainment tool.

Improvements in PA and NA, as measured by PANAS, suggest that the system may have supported emotional well-being during the memoir-writing session. This aligns with the KE component of the design framework, which translated affective needs (eg, the desire for recognition and anxiety about forgetting) into features such as empathetic audio feedback and sensory memory cues. Pappadà et al [7] noted that dementia-assistive technologies tend to prioritize cognitive function while underserving emotional needs; these results suggest that emotional engineering principles may contribute meaningful affective benefits.

The empirical findings, together with the design process, suggest 3 implications for future dementia-assistive AIGC systems. First, AI proactivity should adapt to cognitive capacity: more severe impairment calls for greater AI narrative initiative, while milder impairment is better served by user-driven prompting that preserves authorial agency, or the user’s sense that the memoir remains their own story. Second, memoir tools should support affective safety alongside memory retrieval. Third, long-term personalization requires a structured life-event memory system with caregiver verification mechanisms and ethical safeguards to ensure that AIGC remains grounded in the user’s actual experience. The multimethod design framework (KE, QFD, AD, and TOPSIS) remains relatively underexplored in dementia-assistive technology [21] and may offer a reproducible starting point for similar work.

This study extends prior work on AI-mediated reminiscence and narrative support for older adults [5,12,27]. More broadly, multimodal AI-driven platforms have also shown promise in supporting expressive and creative interventions for home-based older adult care, such as interactive art therapy systems [34], reinforcing the potential of AI-based creative engagement to benefit older populations beyond memoir writing specifically. Wang et al [12] demonstrated that AI-driven multimodal photo albums support personalized reminiscence in cognitively intact older adults; our study extends this to individuals with diagnosed cognitive impairment, using MMSE scores to drive adaptive interaction control. Feng et al [27] found that contextual and multimodal stimuli enhanced engagement in people with dementia during human-robot interaction; our results suggest that similar multimodal principles may be effective in chatbot-mediated memoir writing. Chang et al [35] further demonstrated that reinforcement learning-based AI can deliver personalized adaptive support in dementia care contexts, reinforcing the value of dynamic AI adaptation for this population.

The present findings also converge with recent longitudinal evidence. Kang et al (2025) [36] found that regular AI care calls over 7 months significantly reduced depression scores and improved memory function in 80 community-dwelling individuals with dementia. That study found that sex and education level moderated intervention outcomes, paralleling our subgroup finding that cognitive severity shapes response to AI interaction and reinforcing the view that individual characteristics meaningfully affect how people with dementia benefit from AI-based support.

The findings also speak to HCI discussions of asymmetric human-AI collaboration [10,37], in which AI systems dynamically compensate for differential user capabilities. Here, narrative initiative is inversely calibrated to residual cognitive capacity, so the AI assumes greater responsibility where the user needs it most. This contrasts with many prior dementia-assistive systems that rely on relatively fixed interaction modes [7]. Whether similar mechanisms generalize to other cognitive assessment instruments remains a question for future work.

However, this study also has some limitations. First, the sample size (n=20; 10 per arm) is small. Effect sizes derived from small samples are inherently unstable, and the Benjamini-Hochberg correction applied to the 6 confirmatory hypothesis tests should be read with caution: because within- and between-group tests were derived from the same 10 participants per arm rather than independent samples, the positive regression–dependence assumption underlying the procedure was plausible but not formally verified, and several outcomes assessed with Wilcoxon signed rank tests at n=9‐10 yield P and q values that take on a limited set of discrete values. The correction should be read as a conservative response to multiplicity concerns rather than a substitute for replication in a larger, adequately powered, multisite trial.

Second, this study used a nonrandomized design: group allocation used a minimization procedure rather than simple or block randomization. This method was selected specifically to ensure equivalence on 4 baseline variables, given the small overall sample size, and is a recognized alternative to simple randomization in small trials. However, it differs from simple randomization in that assignment for later-entered pairs is partly determined by the accumulated composition of previously assigned groups rather than by chance alone. As a result, unmeasured or unbalanced covariates could, in principle, be associated with allocation in ways that pure randomization is designed to prevent, and a person with knowledge of the algorithm and running group totals could in principle anticipate upcoming assignments. Larger-scale replications should use simple or stratified block randomization with formal allocation concealment to strengthen causal inference.

Third, because participants were allocated within matched pairs, all confirmatory hypothesis tests were analyzed using paired methods (paired t tests or Wilcoxon signed rank tests on matched-pair differences) rather than independent-samples tests, consistent with the matched design. This is a more conservative and design-appropriate approach than treating change scores as independent observations; all 6 confirmatory comparisons remained statistically significant under this analysis and after Benjamini-Hochberg correction, with moderate-to-large effect sizes (Cohen d=0.85‐1.03; rank-biserial r=−0.93 to 1.00).

Fourth, this was a one-time experimental evaluation without follow-up assessment. This has 2 consequences. It precludes conclusions about the durability of affective and narrative gains; whether benefits persist, accumulate with repeated use, or diminish with habituation is unknown. It also means the high SUS scores (mean 92.2, SD 5.1) should be interpreted cautiously: single exposure to a novel system can temporarily inflate satisfaction ratings, and older adults with cognitive impairment are additionally prone to acquiescence bias, meaning the tendency to agree with survey items regardless of content, particularly under limited cognitive resources [38]. Holden [30] noted similar interpretive challenges applying the SUS in this population. Longitudinal, multisession usability assessment is needed to obtain a more reliable estimate of system acceptability.

Fifth, participants were recruited exclusively from 4 Chinese provinces, with a mean education level of approximately 6 years and corresponding technology-use attitudes. System performance may differ in populations with higher digital literacy, different cultural or linguistic backgrounds, or distinct dementia etiologies, and these findings should not be generalized beyond this context without further testing.

Sixth, the system’s core AIGC engine relies on GPT-3.5. GPT-3.5 represented a relatively advanced, readily available commercial large language model at the time this system was developed, and its selection was reasonable within that context. However, given the pace at which generative AI models have since evolved, findings obtained with this specific model may not generalize to systems built on newer architectures. Replication with current-generation models is needed, alongside systematic safety evaluation under diverse real-world conditions, which has not yet been conducted for either version.

Seventh, the control condition requires clarification and cautious interpretation. As described in the Procedures section, the control group received an active control matched to the intervention group on session duration, number of prompting turns, frequency of audio feedback, device novelty, and researcher behavior, with only the AI-generated personalization and memory anchor–driven narrative content removed. This design was intended to isolate the specific contribution of AI-generated scaffolding rather than general attention or novelty effects. However, no formal fidelity check (eg, independent rating of researcher adherence to the matched protocol) was conducted to confirm that the manually delivered control condition in fact replicated the intervention’s interaction structure and pacing. Residual, unmeasured differences in perceived responsiveness or engagement between human-delivered and AI-delivered support therefore cannot be fully excluded, and this should be addressed in future trials through structured fidelity monitoring or a non-AI digital comparator.

Eighth, psychiatrist-rated memoir quality was assessed from transcripts with AI-paraphrased content and AIGC excluded, so that scores reflect participant-generated narrative content specifically rather than the system’s language capabilities.

AI-assisted memoir construction for people with reduced decision-making capacity raises important questions about authorship, narrative authenticity, and data governance. As described in the Ethical Considerations section, this study addressed these concerns through explicit consent for AIGC, caregiver verification of generated narratives, anonymized API data transmission, and postsession transparency about system outputs. Future deployments should extend these safeguards to longitudinal use by formalizing policies for data retention, access, revision, and deletion [17,23].

Future studies should evaluate the system across multiple sessions and with stronger comparison conditions, such as structured human prompting or non-AI digital memoir tools, to better isolate the contribution of AI-generated scaffolding. A larger and more diverse randomized controlled trial would also enable subgroup analyses by dementia etiology, severity, education, and cultural context. Finally, because the current MMSE-based adaptation logic remains relatively coarse, future versions could incorporate real-time behavioral signals, such as response latency, lexical complexity, and speech-detected emotional tone, to support more fine-grained within-session adaptation [39]. Together, these steps would help move the system from a promising single-session prototype toward a clinically useful and ethically accountable long-term reminiscence support tool.