AIRS & Appropriate Reliance Research
SKILL.md
AIRS & Appropriate Reliance Research
Domain knowledge for AI adoption measurement, psychometric instrument development, and appropriate reliance research
This skill contains knowledge about the AIRS-16 validated instrument, the proposed AIRS-18 extension with Appropriate Reliance (AR), and research methodologies for studying AI adoption and human-AI collaboration.
When to Use
- Discussing AIRS-16 or AIRS-18 instruments
- Developing or extending psychometric scales
- Analyzing AI adoption patterns
- Researching appropriate reliance / trust calibration
- Preparing academic papers or research briefs
- Meeting preparation with researchers
AIRS-16: AI Readiness Scale
Source: Correa, F. (2025). Doctoral dissertation, Touro University Worldwide.
Production: airs.correax.com | Time: 5 minutes | Built by: Alex Cognitive Architecture
Validation: N=523, CFI=.975, TLI=.960, RMSEA=.053, R²=.852
Quick Links
| Link | Purpose |
|---|---|
| Take Assessment | Start the 16-item survey |
| View History | Review past results |
| Register Org | Enterprise organization setup |
| GitHub (Platform) | AIRS Enterprise source code |
| GitHub (Research) | Validation data & analysis |
User Roles
| Role | Access |
|---|---|
| 👤 Participant | Take assessments, view personal results, download PDF reports |
| ✨ Founder | Organization creator, can be promoted to Admin |
| 🛡️ Admin | Dashboard analytics, member management, invitations |
| 👑 Super Admin | Platform-wide access, all orgs, AI prompts configuration |
8 Constructs (2 items each)
| Construct | Code | Description |
|---|---|---|
| Performance Expectancy | PE | Belief that AI will help achieve job performance gains |
| Effort Expectancy | EE | Perceived ease of use of AI systems |
| Social Influence | SI | Degree to which colleagues/leadership encourage adoption |
| Facilitating Conditions | FC | Availability of organizational resources and training |
| Hedonic Motivation | HM | Enjoyment and curiosity when exploring AI capabilities |
| Price Value | PV | Perceived benefit relative to effort invested (β=.505 — strongest predictor) |
| Habit | HB | Extent to which AI use has become automatic and routine |
| Trust in AI | TR | Confidence in AI reliability, accuracy, and data handling |
Key Finding: What Actually Predicts AI Adoption
| Predictor | β | p | Status |
|---|---|---|---|
| Price Value (PV) | .505 | <.001 | ✅ STRONGEST |
| Hedonic Motivation (HM) | .217 | .014 | ✅ Significant |
| Social Influence (SI) | .136 | .024 | ✅ Significant |
| Trust in AI (TR) | .106 | .064 | ⚠️ Marginal |
| Performance Expectancy (PE) | -.028 | .791 | ❌ Not significant |
| Effort Expectancy (EE) | -.008 | .875 | ❌ Not significant |
| Facilitating Conditions (FC) | .059 | .338 | ❌ Not significant |
| Habit (HB) | .023 | .631 | ❌ Not significant |
Insight: Traditional UTAUT2 predictors (PE, EE, FC, HB) do NOT predict AI adoption. Value perception, enjoyment, and social influence matter.
Scoring & Typology
# AIRS Score = sum of 8 construct means (range: 8-40)
AIRS = PE + EE + SI + FC + HM + PV + HB + TR
# Typology (94.5% accuracy)
if AIRS <= 20: "AI Skeptic" # 17% of sample
elif AIRS <= 30: "Moderate User" # 67% of sample
else: "AI Enthusiast" # 16% of sample
Appropriate Reliance (AR): Proposed AIRS-18 Extension
The Research Question
Is it not how much you trust AI that predicts adoption, but how well your trust is calibrated to actual AI capability?
Why AR ≠ Trust (TR)
| Dimension | Trust (TR) | Appropriate Reliance (AR) |
|---|---|---|
| Measures | Trust level | Trust calibration accuracy |
| Type | Attitude (affective state) | Metacognitive skill |
| Failure mode | Low trust → under-use | Low AR → over-reliance OR under-reliance |
| Item example | "I trust AI tools..." | "I can tell when AI is reliable..." |
Key distinction: TR asks "Do you trust AI?" — AR asks "Can you discern when trust is warranted?"
The 2×2 Independence Matrix
| Low AR (Miscalibrated) | High AR (Calibrated) | |
|---|---|---|
| High TR | ⚠️ Over-reliance → bad outcomes → abandonment | ✅ Optimal adoption |
| Low TR | ❌ Under-reliance → missed value → rejection | ✅ Calibrated skeptic → gradual adoption |
Proposed AR Items
| Item | Text | Component |
|---|---|---|
| AR1 | I can tell when AI-generated information is reliable and when it needs verification. | CAIR |
| AR2 | I know when to trust AI tools and when to rely on my own judgment instead. | CSR |
CAIR/CSR Framework (Schemmer et al., 2023)
| User Accepts | User Rejects | |
|---|---|---|
| AI Correct | CAIR ✅ (Correct AI-Reliance) | Under-reliance |
| AI Incorrect | Over-reliance | CSR ✅ (Correct Self-Reliance) |
Metric: Appropriateness of Reliance (AoR) = 1 indicates optimal calibration.
Research Hypotheses for AIRS-18 Validation
| # | Hypothesis |
|---|---|
| H1 | AR demonstrates acceptable reliability (α ≥ .70, CR ≥ .70, AVE ≥ .50) |
| H2 | AR shows discriminant validity from TR (HTMT < .85) |
| H3 | AR positively predicts BI (β > 0, p < .05) |
| H4 | AR provides incremental validity beyond AIRS-16 (ΔR² > .02) |
| H5 | AR moderates TR→BI (high AR strengthens the relationship) |
| H6 | AR mediates Experience→BI (experience → better calibration → adoption) |
Psychometric Standards
Reliability Thresholds
| Metric | Minimum | Good | Excellent |
|---|---|---|---|
| Cronbach's α | .70 | .80 | .90 |
| Composite Reliability (CR) | .70 | .80 | .90 |
| Average Variance Extracted (AVE) | .50 | .60 | .70 |
Model Fit Indices
| Index | Acceptable | Good |
|---|---|---|
| CFI | ≥ .90 | ≥ .95 |
| TLI | ≥ .90 | ≥ .95 |
| RMSEA | ≤ .08 | ≤ .06 |
| SRMR | ≤ .08 | ≤ .05 |
Discriminant Validity
| Method | Criterion |
|---|---|
| HTMT | < .85 (conservative: < .90) |
| Fornell-Larcker | √AVE > inter-construct correlations |
Intervention Strategies by Typology
| Typology | AIRS-16 Focus | + AR-Informed Focus |
|---|---|---|
| AI Skeptics (≤20) | Trust-building, low-effort demos | Calibration training: "Here's when AI excels vs. struggles" |
| Moderate Users (21-30) | Clear use cases, ROI evidence | Verification skill-building: "How to spot AI errors" |
| AI Enthusiasts (>30) | Advanced features, leadership | Reliance audits: "Are you over-relying in high-stakes areas?" |
Key References
| Reference | Contribution |
|---|---|
| Correa (2025) | AIRS-16 validation, UTAUT2 extension |
| Passi, Dhanorkar, & Vorvoreanu (2024) | AETHER synthesis on appropriate reliance |
| Schemmer et al. (2023) | CAIR/CSR framework |
| Venkatesh et al. (2012) | UTAUT2 original model |
| Lee & See (2004) | Trust calibration in human-automation interaction |
| Lin et al. (2022) | LLMs can verbalize calibrated uncertainty |
Troubleshooting
"Is AR just measuring AI experience?"
Problem: Concern that AR conflates with general AI familiarity.
Solution:
- Include experience as covariate
- Test discriminant validity (HTMT < .85)
- AR should predict beyond experience level
"Can self-reported calibration be valid?"
Problem: People may not accurately assess their own calibration ability.
Solution:
- Self-report measures perceived calibration
- Future research: correlate with behavioral CAIR/CSR in task studies
- Perceived calibration may still predict adoption intentions
"Why was Trust marginal in AIRS-16?"
Possible explanations:
- Trust level alone is insufficient — calibration matters more
- Trust may be necessary but not sufficient
- TR × AR interaction: trust only helps when calibrated
- Sample characteristics (tech-savvy population)