Module IV:
Ethics Washing,
Explainability,
Robustness
Understanding Ethics Washing
What is Ethics Washing?
Ethicswashing refers to the practice where organizations or institutions publicly emphasize ethical values or AI ethics initiatives to
appear responsible and trustworthy, while failing to take meaningful action to address real ethical problems in their technologies or
operations. It is similar to greenwashing in environmental contexts, where companies overstate their environmental responsibility.
Ethics Washing = Saying <We¾re ethicalî without being ethical.
Why It Happens
Organizations engage in ethics washing to:
Avoid regulation or criticism while maintaining public trust. Delay accountability by forming committees or issuing
vague ethics principles without enforcement.
1 2 3
Promote their brand image as <ethicalî or <responsible.î
Examples
AI companies creating ethics boards that have no decision-making power or are dissolved when facing
1 controversy.
Tech firms publishing <AI principlesî but continuing practices like data
2 exploitation or algorithmic bias.
Corporations claiming <fair AIî while keeping models and
3 datasets secret, preventing external audits.
Consequences
Slower progress toward genuine
2 responsible AI development.
Erosion of public trust in ethical AI
efforts.
1
3 Misleading policymakers and users
about real ethical risks.
How to Avoid Ethicswashing
Transparency: Share data sources, model details, and 1
evaluation methods.
2 Accountability: Ensure ethical principles are enforced
through audits and independent oversight.
Stakeholder inclusion: Involve affected communities, 3
ethicists, and diverse experts in decisions.
4 Action over words: Demonstrate concrete ethical
practices 4 not just policies.
Explainability from a user perspective:
Explainability problem
What is explainable AI?
Explainable artificial intelligence (XAI) is a set of processes and methods that allows human users to comprehend and trust the
results and output created by machine learning algorithms.
Explainable AI is used to describe an AI model, its expected impact and potential biases. It helps characterize model accuracy,
fairness, transparency and outcomes in AI-powered decision making. Explainable AI is crucial for an organization in building trust
and confidence when putting AI models into production. AI explainability also helps an organization adopt a responsible approach
to AI development.
What Is Explainability?
Explainability in artificial intelligence refers to the ability to describe an AI model's internal workings or outcomes in understandable
terms. It makes complex AI decisions transparent and trustworthy. In fields like healthcare or finance, where understanding why a
model made a particular decision has implications, explainability has influence. In terms of MLOps and AI security, explainability
supports accountability and helps diagnose and rectify model errors
Why Explainability Matters
Machine learning models, particularly those based on complex algorithms like neural networks, can act as black boxes, obscuring
the if/then logic behind their outputs. This opacity can lead to mistrust or skepticism among stakeholders, regulators, and
customers who need to understand the basis of decisions impacting them.
In healthcare, for example, an AI system could be employed to assist radiologists by prioritizing cases based on the urgency
detected in X-ray images. In addition to performing with high accuracy, the AI system must provide explanations for its rankings to
ensure patient safety and comply with medical regulations. In other words, it needs to be transparent enough to reveal the features
in the images that led to its conclusions, enabling medical professionals to validate the findings.
Additionally, in jurisdictions with regulations such as the EU's General Data Protection Regulation (GDPR), patients may have the
right to understand factors influencing their cases and could challenge decisions made with the aid of AI. In instances such as this,
explainability goes beyond technical performance to encompass legal and ethical considerations.
Basic Problems in Making Algorithms Explainable
Accuracy vs. Interpretability Trade-off: More complex, highly accurate models (like deep neural networks) are generally less
interpretable, while simpler, transparent models (like linear regression or decision trees) often sacrifice predictive performance
on complex data.
Model Complexity: Modern machine learning systems can have millions of parameters and intricate internal workings, making it
difficult to trace exactly how a specific input leads to a particular output, even for the developers themselves.
Lack of Standardized Evaluation Metrics: There is no universal consensus or objective metric for what constitutes a "good"
explanation. Different stakeholders (e.g., a data scientist, a regulator, an end-user) require different types and levels of
explanation, making standardized evaluation difficult.
Human Bias: Explainability methods can be affected by biases present in the training data or the design choices made by
human developers. Explanations might inadvertently reinforce or obscure these biases, leading to unfair or discriminatory
outcomes.
User Understanding: Explanations must be tailored to the target audience's expertise. A technical explanation suitable for an
ML engineer may be incomprehensible to a domain expert (e.g., a doctor or a loan officer), which can lead to over-reliance
(automation bias) or under-reliance (algorithmic aversion) on the AI system.
Computational Expense: Many post-hoc explainability methods, such as SHAP or LIME, are computationally expensive,
especially for large models or real-time applications, which can limit their practicality.
Causality vs. Correlation: Most current XAI methods highlight correlations and feature importance, but they struggle to provide
true causal explanations, which is often what humans need to make informed decisions and act effectively.
Approaches to Making Algorithms Explainable
1 2 3
Intrinsically Interpretable Models Post-Hoc Explanation Techniques Procedural and Design Approaches:
(Explainable by Design):These (for "Black-Box" Models):These
User-Centric Design: Involves
models are designed from the methods are applied after a complex
the end-users in the design
ground up to be transparent and their model (like a neural network or
process to ensure explanations
operations are directly ensemble method) has been trained
meet their specific needs,
understandable by humans. to provide insights into its behavior.
knowledge levels, and context.
Linear/Logistic Regression: The Techniques: Integrate XAI Early: Incorporate
coefficients directly indicate the Feature Relevance explainability into the AI
weight of each feature's Explanations (e.g., SHAP, development workflow from the
influence on the output. LIME): These methods assign outset, rather than as an
Decision Trees/Rule-based an importance score to each afterthought.
Systems: The logic follows input feature for a specific Human-in-the-Loop: Maintain
explicit "if-else" conditions that prediction (local explanation) meaningful human oversight so
are easy to follow and or across the entire model that a person can scrutinize the
understand. (global explanation). AI's recommendations, apply
Generalized Additive Models Visual Explanations: their own expertise, and be
(GAMs): These models allow the Techniques like saliency maps accountable for the final decision.
impact of each feature to be or Grad-CAM highlight Auditing and Regulation:
visualized individually. specific parts of the input Establish clear guidelines and
data (e.g., pixels in an image) regulatory frameworks that
that the model focused on mandate specific levels of
when making a decision. transparency and allow for
Explanations by independent audits of AI systems
Simplification (Surrogate to ensure fairness and
Models): A simpler, compliance.
interpretable model (e.g., a
small decision tree) is trained
to approximate the behavior
of the complex black-box
model, either globally or
locally around a specific
prediction.
Explanations by Example:
Involves extracting
representative examples or
prototypes from the training
data that are similar to the
instance being predicted to
help justify the outcome.
Counterfactual Explanations:
These describe the smallest
change to the input data that
would alter the model's
prediction (e.g., "If your
income was $5,000 higher,
your loan would have been
approved").
Understand the difference between explainability
and interpretability of algorithms
What are Interpretability and Explainability?
Interpretability: refers to the ability to understand the decision-making process of an AI model. An interpretable model is
transparent in its operation and provides information about the relationships between inputs and outputs. An interpretable
algorithm can be explained clearly and understandably by a human being. Interpretability is therefore important to ensure that
users can understand and trust artificial intelligence models.
Explainability: pertains to the ability to explain the decision-making process of an AI model in terms understandable to the end
user. An explainable model provides a clear and intuitive explanation of the decisions made, enabling users to understand why
the model produced a particular result. In other words, explainability focuses on why an algorithm made a specific decision and
how that decision can be justified.
Aspect Interpretability Explainability
Focus Understanding the model itself Understanding the model¾s behavior or
decisions
Model Type Usually applies to simple, transparent Can be applied to complex <black-boxî
models (e.g., linear regression, models (e.g., deep learning,
decision tree) ensembles)
Question Answered <How does the model work?î <Why did the model give this output?î
Approach Direct transparency Post-hoc (after training) explanations
Example Interpreting weights in logistic Using SHAP to explain a neural
regression network¾s prediction
AI/ML algorithmic robustness(adversarial
attacks,minimizing security risks)
Algorithmic robustness means how stable and reliable an AI/ML model is when faced with unexpected, noisy, or maliciously
altered inputs.
In simple words:
A robust model continues to perform well even when the input data slightly changes or contains attacks/errors.
Robustness ensures that the model¾s behavior remains consistent, trustworthy, and safe under real-world conditions.
þ 2. Adversarial Attacks
Adversarial attacks are intentional manipulations of input data designed to fool AI models into making wrong predictions 4 even
though changes may be imperceptible to humans.
These are small, carefully crafted perturbations that exploit the model¾s vulnerabilities.
ý Types of Adversarial Attacks
Type Description Example
Evasion Attack Modify input data at prediction time to Adding small noise to an image so that
mislead the model. a <stop signî is classified as a <speed
limit sign.î
Poisoning Attack Tamper with the training data to Injecting mislabeled samples into
corrupt the model. training data so the model learns
wrong patterns.
Model Inversion Attack Try to extract private or sensitive Guessing details of individuals in a
information from the model¾s outputs. medical dataset used to train an AI.
Membership Inference Attack Determine whether specific data Identifying if a person¾s data was part
points were used in training. of the training set, violating privacy.
ý Example (Evasion Attack on Image Model)
An attacker adds a nearly invisible noise pattern to a picture of a <stop sign.î
To humans ³ it still looks like a stop sign.
To the AI ³ it¾s now recognized as a <speed limitî sign.
7 This is dangerous for autonomous vehicles or security systems.
þ 3. Minimizing Security Risks (Improving Robustness)
To make models robust against adversarial and security threats, several strategies are used:
(a) Adversarial Training
Train the model on both normal and adversarially perturbed samples.
The model learns to recognize and resist manipulated inputs.
í Idea: <If I¾ve seen such attacks before, I won¾t be fooled again.î
(b) Defensive Distillation
Train the model to smooth its decision boundaries, making it less sensitive to small input changes.
(c) Input Preprocessing
Apply filters, noise reduction, or normalization to remove adversarial noise before feeding data to the model.
(d) Model Monitoring & Anomaly Detection
Continuously monitor model inputs and outputs for unusual patterns that might indicate an attack.
(e) Secure Data and Training Pipelines
Ensure data integrity (no tampering during collection or training).
Use encryption, access control, and secure APIs.
(f) Regular Audits and Testing
Conduct robustness testing under simulated adversarial conditions.
Evaluate model performance not only on accuracy but also resilience to perturbations.
þ 4. Why Robustness Matters
Safety: Prevents AI failures in critical systems (healthcare, self-driving cars, finance).
Security: Protects against malicious manipulation.
Trust: Builds user confidence in AI decisions.
Fairness: Prevents attackers from exploiting vulnerabilities to target specific groups.