Introduction

Artificial intelligence (AI) is rapidly entering clinical workflows, promising to transform the management of complex chronic conditions. Multiple sclerosis (MS) serves as a critical model for this transition: as a paradigm of data-driven care, AI is expected to influence imaging, risk stratification, prognosis, and long-term monitoring [1]. However, the clinical value of AI will translate into routine care only if patients consider these systems acceptable and trustworthy. This is particularly relevant in MS, where many AI-enabled services depend on continuous patient-generated data (e.g., digital biomarkers, patient-reported outcomes) and are intended to support shared decision-making rather than replace clinician judgement [2–5].

People with MS (pwMS) represent a population with comparatively high digital engagement, which may facilitate adoption of AI-supported care models [6]. Earlier work has shown frequent internet use and widespread uptake of digital communication among pwMS, and acceptance of newer communication channels was higher among those already using established technologies [4,7,8]. This baseline digital engagement provides a lever for introducing AI-based tools tailored to this population. Given the demographic profile of MS, which disproportionately affects females and typically begins in early adulthood, this population may further support digital readiness and openness toward AI-driven health innovations [9,10].

While digital literacy is high, acceptance of automated and AI-based systems in healthcare is not universal [4,11,12]. Beyond individual familiarity, acceptance is also shaped by trust in the health system itself [13]. This creates an opportunity to introduce AI applications that build on existing digital behaviours. Yet, digital access does not automatically imply readiness to rely on algorithmic decision support. Evidence from other healthcare settings suggests that acceptance of automated or AI-based systems is heterogeneous and shaped by sociodemographic context, technology familiarity, and perceived risk [4,11,14]. For example, studies in primary care and hospital populations indicate that trust in AI varies across groups and tends to be higher among individuals with greater AI knowledge and more frequent technology use, while some patients—particularly those who feel vulnerable—express reservations about automation in high-stakes decisions [4,11]. In chronic disease contexts, patients often prefer AI as an assistive “support tool,” valuing convenience and monitoring while maintaining a strong preference for human accountability in critical decisions [5,15,16].

In the context of MS, recent perspectives emphasize that successful AI implementation requires patient input and co-design to address concerns around transparency, privacy, and responsibility [17,18]. Despite this, empirical data describing how pwMS perceive AI, both in general and across distinct clinical use cases, remain scarce. Given the heterogeneity of the MS population (age, disability, socioeconomic background) and the long-term nature of care, understanding where pwMS are comfortable with AI and which factors predict acceptance is essential for responsible and effective implementation [18,19].

Therefore, this study investigated attitudes toward AI among pwMS in Germany, distinguishing general acceptance from comfort with specific clinical domains and examining how demographic, clinical, and AI-exposure variables relate to these perceptions.

Results

Study population and clinical characteristics

A total of 241 participants were included. Participant characteristics are shown in Table 1. Most respondents were female (73.5%), had relapsing–remitting MS (61.4%), and reported low-to-moderate disability (PDDS median 2.0 [IQR 1.0–4.0]). Mean age was 50 ± 11 years and mean disease duration 13.5 ± 9.8 years. Age correlated positively with disability (r_s = 0.434, P < 0.001) and disease duration (r_s = 0.418, P < 0.001); disease duration was also positively associated with PDDS (r_s = 0.359, P < 0.001).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Baseline demographic and clinical characteristics (N = 241).

https://doi.org/10.1371/journal.pdig.0001236.t001

AI literacy, exposure, and attitudes across domains

Nearly half of respondents rated their AI knowledge as good/expert (47.7%). General AI use ≥ 1 × /week was reported by 30.7%, whereas health-related AI use ≥ 1 × /week was less common (10.8%; S3 Fig).

Acceptance differed significantly across the eight assessed MS-care domains (Cochran’s Q(7)=98.89, P < 0.001; Fig 1), suggesting a “responsibility gradient.” Respondents were largely optimistic about AI’s potential (60.5% somewhat/very optimistic) and supported increasing AI use in MS care (56.9% agree/strongly agree). Comfort was highest for supportive applications, including chronic management (54.4%) and symptom screening (50.2%), but decreased for high-stakes use cases, notably treatment selection (38.6%) and diagnosis (35.3%).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Attitudes toward AI in MS care (Likert distributions, % of valid responses).

Caption: Five-point Likert distributions shown as percent of valid responses. Valid denominators were n = 241 for panels a and c–h, and n = 232 for panel b (“AI should increasingly be used in MS care”). Panels: (a) perceived potential; (b) agreement with increased use; (c-f) comfort across use-cases (symptom detection, diagnosis, treatment selection, chronic management); (g) preference for AI-enabled centres; (h) desire for an MS-specific tool. Responses skewed toward optimism and comfort for management/screening, with more neutral or mixed views for diagnosis/treatment. AI: Artificial Intelligence. MS: Multiple Sclerosis.

https://doi.org/10.1371/journal.pdig.0001236.g001

Regarding adoption preferences, 48.1% agreed they would prefer treatment at an AI-enabled MS centre, and 60.2% expressed interest in an MS-specific AI tool to be used alongside their clinician. When asked about diagnostic decisions assuming equal accuracy, the majority preferred joint AI–clinician decision-making with clinician final responsibility (78.8%); 7.9% preferred clinician-only decisions and 0.8% preferred AI as the final decision-maker.

Overall acceptance (AI attitudes composite) and subgroup differences

The AI attitudes composite showed excellent internal consistency (Cronbach’s α = 0.90). Overall acceptance was moderate (mean 3.39 ± 0.78; median 3.50 [IQR 3.00–3.88]). In unadjusted analyses, higher composite scores were observed in men (P = 0.024, Fig 2), younger participants (P = 0.008), those living in Western vs Eastern Germany (P < 0.001), participants with lower disability (PDDS 0–2 vs ≥ 3; P = 0.035), those with higher AI knowledge (P < 0.001), and those reporting frequent general AI use (≥1 × /week; P < 0.001). Differences by education, disease duration, and health-related AI use were not statistically significant (P = 0.213, 0.272, and 0.059, respectively; S4 Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Between-group differences in AI acceptance (composite).

Caption: Forest plot of Mann–Whitney effect sizes (r) for between-group differences in the AI-attitudes composite score across predefined subgroups. Points represent effect sizes; horizontal lines indicate 95% confidence intervals. Larger absolute values indicate greater between-group differences. The sign of r reflects direction: positive values indicate higher AI-attitudes in the first group named on the y-axis. Blue points denote P < 0.05 (two-tailed); grey points denote non-significant comparisons. Annotated labels show exact p-values. Subgroup-defining valid denominators were n = 241 for education, PDDS, AI knowledge, general AI use, and health-related AI use; n = 239 for age and disease duration; n = 238 for gender; and n = 207 for region. AI: artificial intelligence; DD: disease duration; PDDS: Patient-Determined Disease Steps.

https://doi.org/10.1371/journal.pdig.0001236.g002

Independent predictors of AI acceptance

In the multivariable linear regression model for the AI attitudes composite (Table 2), the model explained 26% of variance (adj. R² = 0.226; F(9,194)=7.57; P < 0.001; VIF ≤ 1.71). Frequent general AI use emerged as the strongest independent predictor (B = 0.58, 95% CI 0.34–0.82; P < 0.001), followed by residence in Western Germany (B = 0.21, 95% CI 0.03–0.39; P = 0.025). Age, education, gender, disease duration, PDDS, AI knowledge, and health-related AI use were not independently associated with the composite (S2 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Predictors of AI acceptance (AI attitudes composite), n = 204.

https://doi.org/10.1371/journal.pdig.0001236.t002

Across proportional-odds models for individual AI-preference outcomes, frequent general AI use was the most consistent predictor of greater acceptance, including optimism about AI, support for increased deployment, comfort with AI for symptom detection, treatment selection, and chronic disease management, preference for AI-enabled centres, and interest in an MS-specific AI tool. Western German residence was associated with higher acceptance for perceived potential for AI-supported chronic management, preference for AI-enabled centres, and desire for an MS-specific tool. Higher AI knowledge was associated with greater comfort for screening only, and older age was associated with lower comfort for AI-supported chronic management. Notably, comfort with AI for diagnosis showed a regional association (lower in the East), whereas general AI use was not a significant predictor in that domain (Fig 3). Sensitivity analyses using the full 5-level ordinal outcomes yielded materially similar conclusions (S5 Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Predictors of artificial intelligence acceptance across eight clinical domains among people with MS: impact of frequent general AI use, region, and age.

Caption: Forest plots show adjusted proportional odds ratios (POR; log scale) with 95% confidence intervals for (a) frequent general artificial-intelligence use (≥1x/week vs < 1x/week), (b) region (Western vs Eastern Germany), and (c) age (per +10 years), across the eight secondary outcomes. POR > 1 indicates higher odds of endorsing a higher (more accepting) response category. Estimates come from proportional-odds (ordinal logistic, logit link) models with outcomes collapsed to three levels (negative: 1–2; neutral: 3; positive: 4–5) to improve model stability and interpretability in the presence of sparse observations. All effects are adjusted for further covariates (demographics, disease characteristics, and artificial-intelligence literacy groupings). Outcome-specific denominators were n = 241 for all modeled items except “AI should increasingly be used in MS care” (n = 232), while models including region were additionally restricted to participants with valid region data (n = 207). AI: Artificial Intelligence; MS: Multiple Sclerosis.

https://doi.org/10.1371/journal.pdig.0001236.g003

Discussion

In this cross-sectional survey of pwMS, we found moderate overall acceptance of AI in MS care, but acceptance was highly use-case dependent. Two patterns stood out: a pronounced “responsibility gradient,” with greater comfort for supportive applications (e.g., screening, chronic management) and lower comfort for high-stakes tasks (diagnosis, treatment selection), and experiential familiarity—operationalized as frequent general AI use—was the strongest and most consistent correlate of acceptance, whereas clinical indicators such as disability and disease duration were not independently associated.

Interestingly, our results suggest that patient acceptance may be more closely related to digital inclusion and everyday exposure to AI than to disease burden. While a recent larger multinational study of hospitalized patients found that poorer health status was a significant predictor of lower AI acceptance, we observed no such association with disability in our MS cohort [5].

The strong association between frequent general AI use and both the composite score and multiple single-item outcomes is consistent with the Technology Acceptance Model. This framework proposes that perceived usefulness and perceived ease of use are the primary drivers of adoption, both of which are likely enhanced by the patients’ prior experience and familiarity with the technology [20]. Importantly, prior work suggests pwMS often show strong health information engagement and limited information avoidance [21], indicating that reluctance toward AI may not stem from disinterest or low competence, but from uncertainty about how AI works, what it can and cannot do, and how it will be governed in clinical settings. This may be clinically actionable: unlike disease severity, familiarity is potentially modifiable through guided, low-risk exposure and patient education, although our cross-sectional design does not allow us to determine whether greater familiarity increases acceptance or whether more accepting individuals are more likely to use AI [13,22].

Acceptance was not uniform across clinical domains. PwMS were more comfortable with AI in supportive roles but more cautious when AI was framed as influencing high-stakes medical decisions. This aligns with evidence from other patient populations showing stronger acceptance of AI as an assistive tool than as an autonomous decision-maker [5,23,24]. Moreover, the observed ‘responsibility gradient’ aligns with recent global findings suggesting that patient caution correlates directly with the personal stakes of the clinical application [5]. Importantly, most respondents preferred a collaborative model where clinicians retain final responsibility. This preference should be treated as a design requirement rather than a barrier: AI tools in MS care are more likely to gain trust when implemented as clinician-support systems embedded in shared decision-making and accompanied by clear accountability structures [17,25].

Regional and age-related differences further suggest that acceptance may be shaped by context. Residency in Western Germany was associated with greater acceptance across several domains. However, this finding should be interpreted cautiously. Because recruitment pathway and recruiting centre were not captured at participant level, we were unable to assess whether the observed association was influenced by differences in recruitment composition or centre association. In addition, region was derived from self-entered postal-code prefixes. The observed West/East difference may therefore reflect unmeasured recruitment-related factors, although broader contextual differences in digital health literacy and engagement with health technologies in Germany may also contribute [26,27]. It may also reflect recruitment-related factors that could not be directly modelled. Age showed a negative association, particularly for AI-supported chronic management, consistent with previous work indicating that older adults’ concerns often center on operational errors, loss of personal interaction, privacy/security, and unequal access to care [28–30]. Together, these findings highlight a potential equity issue: groups that may benefit from monitoring and supportive AI tools could also be those with lower baseline comfort, underscoring the need for age-sensitive and region-sensitive implementation approaches.

Taken together, our findings argue against a “one-size-fits-all” rollout. A pragmatic approach would be to start with high-acceptance use cases (monitoring, screening, administrative support), where benefits are tangible and perceived risk is lower, and then iteratively expand toward more decision-critical applications as familiarity and trust grow. To support patient buy-in, implementation strategies should include transparent explanation of the tool’s role, explicit preservation of clinician accountability, clear communication of what the tool can and cannot do, and early deployment in lower-risk contexts where practical benefit is readily understandable.

Because familiarity appears to be associated with greater acceptance, implementation strategies could also include low-stakes onboarding (e.g., scheduling/administrative chatbots), transparent communication about data use and limitations, and patient-facing explanations that emphasize augmentation rather than replacement of clinician judgement [5,17,25,31]. This staged strategy may be particularly relevant for older pwMS and for settings with lower baseline acceptance [32,33]. Given the growing evidence for the potential of assistive AI systems in the care of MS, this is an urgent matter that requires both sound clinical judgement as well as roll-out strategies that are coordinated with pwMS.

This study provides, to our knowledge, one of the first empirical baselines of AI acceptance in pwMS across multiple clinical domains, using both a reliable composite measure and domain-specific outcomes. However, several limitations should be considered. The cross-sectional design precludes causal inference; in particular, plausible reverse causality exists for the association between frequent AI use and AI acceptance, and participation itself may have favoured respondents who were more open to digital tools. Moreover, recruitment through academic centres and digital dissemination channels likely introduced selection bias toward digitally engaged and highly educated participants. Because recruitment pathway and recruiting centre were not captured at participant level, we were unable to determine whether observed regional differences partly reflected differences in recruitment mix rather than underlying contextual differences. Additionally, the attitudinal items were primarily framed around acceptance, comfort, and perceived potential. While items allowed respondents to express caution or reluctance, the questionnaire did not include dedicated items specifically probing perceived harms, risk, or safety concerns. Considering sparse data in some response categories, collapsing Likert responses improved model stability but reduced response granularity. Although MS serves as a valuable model for data-driven chronic care, these findings may have limited generalizability to broader MS populations, particularly people with lower educational attainment, different disease phenotypes, or less access to specialized care. Additionally, MS diagnosis and disability were self-reported and not clinically verified. Finally, attitudes were assessed during a period of intense public discourse on AI (including large language models), which may have influenced responses and may not fully reflect stable long-term views.

Ultimately, our results suggest that the successful integration of AI into MS care is not merely a technological challenge, but a sociotechnical one. While the potential for algorithmic precision and newer models is clear, our findings suggest that perceived clinical utility is closely linked to clear communication about the tool’s role and to users’ prior familiarity with AI-enabled technologies, adding granularity by showing that caution is not a blanket rejection. Moving forward, the priority must shift from simply validating model performance to designing ‘human-in-the-loop’ workflows that respect the patient’s clear preference for augmented rather than automated care. By treating patient familiarity as a modifiable factor—cultivated through active education and continuous, low-risk engagement—clinicians and developers can bridge the current trust gap, ensuring that AI evolves from a theoretical asset into a clinically integrated partner in the long-term management of multiple sclerosis [13].

Conclusion

PwMS appear open to AI especially for supportive functions while clearly preferring human accountability for high-stakes decisions. Our results highlight that acceptance was more strongly associated with experiential familiarity than with clinical severity, a pattern that may also be relevant to other data-rich chronic diseases. This suggests that careful, transparent, and staged implementation focused on building digital trust may be central to successful integration of AI into routine clinical practice. Our approach of patient-centred Living Labs may provide the necessary real-world framework for careful, transparent, and staged AI implementation [34,35].

Materials and methods

Supporting information

Acknowledgments

The authors would like to thank all people with MS who participated in this study for their time and valuable insights. Special thanks to aMStart and the MS Perspektive podcast for their assistance with study recruitment and dissemination.