Module:6 Trust and
Acceptance
Metrics to evaluate XAI, Trustworthy Explainability Acceptance, Power
Quality Disturbance (PQD) classification, Methods for measuring
human intelligence. Evaluating AI system.
Dr Vinothini A
Trust
Trust in XAI measures the degree of confidence and reliance that a
user places in an AI system’s predictions, especially after viewing the
explanations.
Acceptance
Acceptance in XAI refers to the degree to which users are willing to use
and depend on an AI system after understanding its explanations and
perceiving its usefulness and trustworthiness.
Trust
• Trust is a tuple of impression and confidence.
• Trust = (Impression, Confidence)
• This means trust in an AI system is composed of two interrelated
dimensions:
✓Impression (Cognitive Dimension) – What users think about the
system.
✓Confidence (Affective Dimension) – What users feel about the
system.
• Trust (T) is often represented as a function of Impression (m) and
Confidence (c),
T(P,Q)=m(P,Q)×c(P,Q)
Metric Example Meaning
Impression m 0.8125 Average trust
Confidence c 0.915 Trust consistency
Trust t T=m * c=0.74 Reliability of model
Impression
• Impression is defined as the level of trust one entity has towards another
entity.
• This means Impression represents how much one agent (P) believes in or
relies on another agent (Q) based on past interactions, performance, and
observable behavior.
• In XAI, this is often the human user’s mental evaluation of the AI system’s
reliability and transparency.
• It is the comprehensive summary of all the measurements between two
entities (P and Q) taken over time.
• It is the result of cumulative experiences, such as: Accuracy of predictions,
Consistency of explanations, Transparency of decision logic,
Responsiveness to user feedback etc.
• Suppose an engineer (P) interacts with an XAI-based Power Quality
Disturbance classifier (Q) four times and rates the model A after each
use. Calculate Impression based on given data.
Interactio 𝑃:𝑄
Measurement 𝑟𝑖 Interpretation
n (i)
1 0.70 Model output understandable
2 0.85 High accuracy and good explanation
3 0.80 Consistent reasoning
4 0.90 Clear and trustworthy decision
N=4
1
𝐼𝑚𝑝𝑟𝑒𝑠𝑠𝑖𝑜𝑛𝑃→𝑄 = 0.70 + 0.85 + 0.80 + 0.90
4
3.25
= = 0.8125
4
The Impression value (0.81) shows that entity P has developed a strong positive trust toward Q.
Confidence
• Confidence ( c P:Q ) measures the certainty or reliability of the
Impression value.
• It reflects how sure entity P is about its impression of entity Q.
• Even if P has a positive impression about Q, confidence tells us how
strongly P believes that impression is accurate.
Confidence
• Suppose an engineer (P) interacts with an XAI-based Power Quality
Disturbance classifier (Q) four times and rates the model A after each
use. Calculate confidence based on given data.
The Confidence (0.915) indicates that the engineer’s trust in
the model’s performance is stable and consistent.
Trustworthy Explainability Acceptance (TEA)
Trustworthy Explainability Acceptance (TEA)
• refers to the extent to which users trust, understand, and willingly
accept the decisions or explanations provided by an XAI system.
• It combines three core pillars:
Focus Question it answers
Reliability & consistency of the “Can I rely on this system’s
Trust
model output?”
Clarity and interpretability of “Do I understand why it made this
Explainability
decisions decision?”
Human willingness to adopt or act “Will I use and accept what the
Acceptance
on the AI’s recommendations model says?”
Metrics to evaluate XAI
Category Focus Evaluated By
How accurately explanations
1. Model-Centric Metrics represent the model’s internal Model-level analysis
behavior.
How clear, simple, and human-
2. Explanation-Centric Metrics Structural or cognitive analysis
understandable explanations are.
How users perceive, trust, and
3. Human-Centric Metrics Human studies or surveys
accept the explanations.
Model Centric XAI Evaluation metrics
Metric Meaning
Fidelity Match with model output
Infidelity Error in explanation response
Sensitivity Change under small input perturbation
Completeness Attribution sum equals output difference
Monotonicity Logical feature-behavior relation
Sparsity Focused explanation
Stability Similar inputs → similar explanations
Models with similar outputs → similar
Consistency
explanations
Robustness Resistance to noise
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Model Centric XAI Evaluation metrics
Explanation-Centric XAI Evaluation Metrics
• 1. Fidelity: 1 - (1/N) Σ |f(xᵢ) - g(xᵢ)|
• 2. Stability: 1 - ||E(x₁) - E(x₂)|| / ||x₁ - x₂||
• 3. Complexity: |E| / d
• 4. Faithfulness: corr(Eᵢ, Δfᵢ)
• 5. Localization: |M ∩ G| / |M ∪ G|
• 6. Sufficiency: f(x_S) - f(x)
• 7. Comprehensiveness: f(x) - f(x \ S)
• 8. Soundness: |E ∩ K| / |E|
• 9. Sparsity: 1 - ||E(x)||₀ / d
• 10. Human Agreement Score (HAS): |E ∩ H| / |E ∪ H|
Human-Centric XAI Evaluation Metrics
• 1. Trust: T^{P:Q} = (m^{P:Q}, c^{P:Q})
• 2. Impression: m^{P:Q} = (1/N) Σ rᵢ^{P:Q}
• 3. Confidence: c^{P:Q} = 1 - √( (1/(N-1)) Σ (rᵢ^{P:Q} - m^{P:Q})² )
• 4. Understanding: ρ = corr(U, M)
• 5. Satisfaction: S = (1/N) Σ sᵢ
• 6. Utility: (P_XAI - P_base) / P_base
• 7. Decision Performance: ΔAcc = Acc_XAI - Acc_baseline
• 8. Cognitive Load: CL = 1 - Interpretability
• 9. Fairness: F = 1 - |E_g - E_h|
• 10. Trust Propagation: T_{A→C} = T_{A→B} × T_{B→C}
Power Quality Disturbance (PQD)
Evaluating AI Systems
↓
XAI Metrics
↓
Applied Case: PQD Classification
↓
Outcome: Trustworthy Explainability Acceptance
Power Quality Disturbance (PQD)
• A PQD classifier is an AI model that learns from historical waveform
data to automatically identify the type of disturbance present in new
or unseen power signals.
Input:
Raw or preprocessed electrical signals:
• Voltage/current waveform (1D time series)
• Spectrogram or scalogram (2D time–frequency representation)
Output:
• Predicted PQD class label (e.g., “Voltage Sag,” “Harmonic,”
“Transient,” or “Normal”).
XAI in PQD classifiers
Category Technique How It Works in PQD Output Type
Quantifies the importance
of signal features (e.g.,
Model-Agnostic (Post- SHAP, LIME, Permutation Feature importance bars
RMS voltage, frequency
hoc) Importance or heatmaps
bands) in determining
PQD type
Highlights time or
frequency regions of the
Model-Specific (Deep Grad-CAM, LRP, Visual heatmap over time-
waveform/spectrogram
Models) Integrated Gradients frequency plot
influencing the model’s
decision
Builds interpretable
Surrogate or Simplified Decision Trees, Rule Human-readable rules: “If
models mimicking CNN
Models Extraction RMS drop > 20%, → Sag”
outputs
Concept Activation Identifies representative
Shows typical examples
Prototype-based Vectors, Case-based PQD samples that the
influencing classification
Reasoning model uses as references
Evaluation Metrics
Metric Meaning in PQD Context
If you remove a highlighted frequency/time
Faithfulness component, model confidence should drop ,proving
the explanation is real.
Stability Similar PQD signals should yield similar explanations.
The explanation should visually or textually highlight
Clarity
the disturbance region clearly.
Explanation aligns with known physical cause (e.g.,
Causality
sudden RMS dip → voltage sag).
Comprehensibility Human expert easily understands the reasoning.
Human intelligence
• Human intelligence is the ability of the human mind to learn from
experience, reason, solve problems, adapt to new situations, and use
knowledge to manipulate the environment effectively.
•
Intelligence = Learning + Reasoning + Adaptation + Problem Solving
Methods for measuring human intelligence in XAI
There are two perspectives:
• Traditional Cognitive Measurement: How we quantify human
intelligence to serve as a benchmark for AI systems.
• Human-Centric XAI Evaluation: How we assess whether AI
explanations align with or enhance human cognitive understanding.
Human Intelligence Dimension AI Equivalent XAI Metric or Method
Analytical Intelligence Model accuracy & logical reasoning Fidelity, Consistency
Generating new solutions or
Creative Intelligence Novelty, Counterfactual reasoning
features
Practical Intelligence Contextual adaptation Robustness, Generalizability
Trust, Human Acceptability,
Emotional/Social Intelligence User-centered explanation
Usability
Metacognition (thinking about Self-Explainability, Transparency
AI self-explanation
thinking) Score