Stakeholders judge AI challenges they find unusually complex to be less prepared for: a one-point within-person rise in perceived complexity reduces perceived community readiness by about 0.21 points, and averaging across people can hide these challenge-specific barriers; policymakers should target support to the specific challenges that stakeholders deem unusually difficult.

Main Finding

When the same person evaluates multiple AI and robotics challenges, perceived complexity of a specific challenge predicts lower perceived community readiness for that challenge (about −0.21 readiness points per one-point higher complexity, p < 0.001). By contrast, people who tend to rate challenges as generally more complex do not systematically report lower readiness overall (between-person effect not significant, p = 0.29). Significance (perceived impact) is positively associated with readiness, but unusually high complexity weakens that alignment. These relationships vary by challenge family and by professional background. Measures such as confidence and trust (collected on "applied" cards) add substantial diagnostic and predictive information about readiness.

Key Points

• Design & scope: Repeated card-based survey; N = 982 participants; 15,200 card evaluations of 17 AI/robotics challenges spanning Biomedical (13 applied cards across families), Manufacturing, Hospitality, Cyber Security and Ethical Awareness.
• Core ratings (1–5) on every card: Significance (impact), Complexity (technical/organisational/societal difficulty), Readiness (perceived community preparedness/resources). Applied cards (13) also measured confidence, trust, familiarity, convenience and satisfaction.
• Within- vs between-person decomposition: Because each respondent rated multiple challenges, the analysis separates (a) stable respondent tendencies (between-person) from (b) challenge-specific deviations within respondents (within-person).
• Main effect sizes:
  • Within-person: a challenge rated 1 point higher than the respondent’s usual complexity → ~0.21 point lower readiness (highly significant, p < 0.001).
  • Between-person: respondents who, on average, rate challenges as more complex do not report systematically lower readiness (p = 0.29).
• Variance decomposition: Intraclass correlation (ICC) ≈ 0.325 — about 32.5% of readiness variation is between respondents; ~67.5% is within respondents (variation across challenges and residual).
• Significance-readiness gap (S − R) used descriptively to highlight where perceived impact exceeds perceived preparedness; gaps are challenge-specific and relate to complexity.
• Professional heterogeneity: Role/industry associations with adjusted readiness persist after accounting for within-person effects — e.g., technical vs non-technical differences matter for interpretation and policy.
• Predictive signals: Confidence, trust and related perceptions provide additional explanatory and predictive power for readiness on applied cards, including predicting held-out participants’ responses.
• Policy relevance: Aggregating across stakeholders can hide challenge-specific barriers; policy and adoption strategies should preserve both between-group differences and within-stakeholder, across-challenge variation.

Data & Methods

• Sample: 982 respondents (Prolific); 15,200 usable card evaluations; 864 respondents completed all 17 cards. Participant spread across age groups, education levels and professional backgrounds (technical and non-technical).
• Challenge set: 17 cards across five families — Biomedical, Manufacturing, Hospitality, Cyber Security, Ethical Awareness. Thirteen cards classified as "applied" (Biomedical, Manufacturing, Hospitality) used extra perception items.
• Measures:
  • Significance: 1–5 perceived impact of addressing the challenge.
  • Complexity: 1–5 perceived combined technical/organizational/societal difficulty.
  • Readiness: 1–5 perceived community preparedness/resources to address the challenge (explicitly not personal competence or audited organisational capacity).
  • Additional (applied cards): confidence, trust, familiarity, convenience, satisfaction.
• Analytical approach:
  • Multilevel (mixed) models with random participant intercepts to capture between-person differences and residuals capturing within-person, challenge-specific deviation.
  • Decomposition of predictors into between- and within-person components (i.e., respondent mean vs deviation for each card).
  • Descriptive use of the derived significance-readiness gap (S − R).
  • Heterogeneity analyses by challenge family and professional group; tests of predictive value of confidence/trust for held-out participants.
• Key statistics reported: Nparticipants = 982; Nobservations = 15,200; ICC ≈ 0.325; within-person complexity → readiness coefficient ≈ −0.21 (p < 0.001).

Limitations (as reported or implied): - Readiness is subjective perceived community preparedness (self-report), not an objective audit of organizational capacity. - Sample from Prolific — useful for within-person decomposition but not a population-representative sample. - Single-item measures per construct (to keep survey manageable) — trade-off between breadth and depth.

Implications for AI Economics

• Heterogeneous, task-specific constraints matter for adoption models:
  • Economic models and diffusion forecasts should account for within-agent variation by task/challenge (agents are not uniformly “ready” across applications).
  • Aggregate measures of "AI readiness" can mislead resource allocation and policy prioritization because they mask challenge-specific bottlenecks.
• Targeting investments and subsidies:
  • Policies that aim to accelerate AI adoption (training, grants, regulatory sandboxes) should be targeted to the particular challenges stakeholders find unusually complex, not only to groups that look unprepared on average.
  • Where perceived significance is high but readiness is low (large S − R gaps), prioritise capacity-building and assurance investments; where both significance and readiness are low, reassess social value or incentives.
• Design of interventions and cost–benefit analysis:
  • Incorporate measures of perceived complexity and trust/confidence into benefit–cost assessments — these perceptions influence uptake and effective use, hence realized returns.
  • Differentiate between constraints that are informational (confidence/trust), capability-based (skills, data, engineering), assurance-related (safety, verification, governance), and resource-based (funding, infrastructure) — each implies different policy instruments and expected lags to adoption.
• Measurement recommendations for AI economics research and evaluation:
  • Use repeated-measures designs (or otherwise decompose within- vs between-agent variation) to reveal challenge-level effects.
  • Include attitudinal measures (trust, confidence, familiarity) as they add predictive power for adoption/readiness beyond simple capability proxies.
  • Report both aggregate readiness and challenge-specific readiness gaps to guide prioritisation.
• Regulatory and institutional design:
  • Regulatory focus should reflect heterogeneity in perceived complexity and in who bears implementation burden; for example, audit/regulatory resources may be more urgently needed for certain high-significance yet high-complexity tasks.
  • Economic incentives (procurement rules, subsidies) can be tuned to lower the implementation/resource constraints that produce large S − R gaps.
• Practical guidance for policymakers and economists:
  • Ask not only “who is ready?” but “which challenges does each stakeholder find unusually difficult?” and “is the bottleneck capability, assurance, implementation logistics, or funding?”
  • Prioritise mixed interventions: technical assistance + governance/assurance tools + targeted training, matched to the challenge-specific diagnosis.

Summary takeaway: Perceived readiness for AI and robotics is highly task-specific. Accounting for within-agent, challenge-level variation — and augmenting readiness surveys with trust/confidence measures — yields more actionable evidence for allocation of resources, policy interventions, and economic models of AI adoption than aggregate readiness metrics alone.