Master10
Artificial Intelligence Module

AI Ethics, Bias & Global Governance

The deployment of artificial intelligence raises critical ethical, legal, and safety considerations regarding algorithmic bias, surveillance, data privacy, and accountability. Algorithmic bias occurs when training datasets contain historical disparities, leading models to perpetuate demographic inequities in hiring, criminal justice, or lending. Explainable AI (XAI) develops interpretable methods (such as SHAP and LIME) to audit 'black box' machine decisions. In 2024, the European Union approved the EU Artificial Intelligence Act, the world's first comprehensive risk-based horizontal regulation for AI systems, categorizing models into unacceptable, high, and minimal risk tiers. India engages through the Global Partnership on Artificial Intelligence (GPAI) and NITI Aayog's National Strategy for AI.

Key Concepts & Examination Highlights

  • The European Union adopted the EU AI Act in 2024, establishing the world's first comprehensive risk-tiered regulatory framework for artificial intelligence.
  • The Global Partnership on Artificial Intelligence (GPAI) is an international multi-stakeholder initiative formed in 2020 to guide responsible AI development.
  • Explainable AI (XAI) frameworks like SHAP (Shapley Additive Explanations) provide mathematical interpretability for complex machine learning predictions.
  • The EU AI Act adopts a risk-based classification framework categorizing AI applications into four tiers: Unacceptable Risk (banned), High Risk (strictly regulated), Limited Risk, and Minimal Risk.
  • Deepfakes are synthetic media generated using deep generative models (such as autoencoders and GANs) to manipulate or replace faces, voices, and actions of individuals realistically.
  • The Bletchley Declaration, signed by 28 countries including India, the US, China, and the EU at the UK AI Safety Summit in November 2023, established global cooperation on frontier AI safety risks.
  • India's NITI Aayog released the 'National Strategy for Artificial Intelligence (#AIforAll)' in 2018, emphasizing ethical AI deployment in healthcare, agriculture, education, smart cities, and mobility.
  • Differential privacy is a mathematical framework for quantifying and limiting the privacy risk to individual data records when sharing aggregate machine learning model statistics.
  • Algorithmic transparency and the right to explanation are incorporated into data protection frameworks like the General Data Protection Regulation (GDPR) to protect individuals from automated decision-making.
  • Data drift and concept drift occur when the statistical properties of production input data or target variables change over time, leading to silent model performance degradation.
  • The IndiaAI Mission, approved by the Union Cabinet in March 2024 with a budget of over ₹10,300 crore, establishes public supercomputing AI compute infrastructure and the IndiaAI Safe & Trusted AI pillar.
  • Algorithmic bias occurs when machine learning models produce systematically prejudiced results against certain demographics due to historical biases embedded in training datasets or flawed sampling methods.
  • The principle of fairness in AI governance includes disparate impact metrics, equalized odds, and demographic parity to ensure equitable outcomes across diverse demographic groups.
  • Explainable AI (XAI) tools like LIME (Local Interpretable Model-agnostic Explanations) approximate complex black-box models locally around a specific prediction to explain why a decision was made.
  • UNESCO adopted the 'Recommendation on the Ethics of Artificial Intelligence' in November 2021, establishing the first global normative standard on AI human rights, dignity, and environmental protection.
  • The White House Office of Science and Technology Policy (OSTP) published the 'Blueprint for an AI Bill of Rights' in 2022, outlining five core protections against automated system harms.
  • Algorithmic red-teaming involves structured adversarial testing by independent security researchers to uncover model vulnerabilities, security bypasses, prompt injection exploits, and toxic outputs.
  • Model watermarking embeds invisible statistical signatures or cryptographic watermarks into AI-generated text, images, and audio to trace origin and authenticate provenance.
  • The G7 Hiroshima AI Process, initiated in 2023, established international guiding principles and a voluntary Code of Conduct for organizations developing advanced AI systems.
  • Data provenance and data lineage in AI track the chronological origin, transformations, and licensing rights of datasets used to train commercial foundation models.
  • Automated facial recognition technology in public surveillance has faced regulatory moratoria and strict legal bans in several jurisdictions due to high error rates on minority populations and civil liberty risks.
  • Intellectual property disputes in Generative AI focus on whether scraping copyrighted books, art, and code for training LLMs constitutes legal 'fair use' or copyright infringement.
  • The Indian AI Safety Institute (AISI) framework is being developed under the IndiaAI Mission to formulate testing benchmarks, vulnerability assessments, and safety standards for foundation models.
  • Environmental sustainability of AI models focuses on the massive carbon footprint and water consumption associated with training and running inference on gigawatt-scale data centers.
  • Dual-use risks of foundation models involve the potential misuse of advanced AI capabilities to assist in synthesizing chemical, biological, radiological, or nuclear (CBRN) weapons or automated cyberattacks.
  • The EU AI Act officially classifies biometric categorization systems and social scoring mechanisms based on social behavior or personality traits under the 'Unacceptable Risk' prohibited tier.
  • High-Risk AI systems under the EU AI Act (such as AI in critical infrastructure, medical devices, and recruitment) must undergo mandatory conformity assessments and continuous risk management.
  • The European AI Office, established under the European Commission, oversees compliance, develops standards for general-purpose AI (GPAI) models, and enforces EU AI Act penalties.
  • The UNESCO Recommendation on the Ethics of AI emphasizes four core values: respect for human rights, living in peaceful societies, ensuring diversity and inclusiveness, and flourishing of the environment.
  • The OECD AI Principles, adopted in 2019 by OECD member countries, set international standards for trustworthy AI focusing on inclusive growth, human-centered values, transparency, and accountability.
  • The NIST AI Risk Management Framework (AI RMF 1.0), published by the US National Institute of Standards and Technology in 2023, provides voluntary guidance to manage AI risks across Map, Measure, Manage, and Govern functions.
  • The US Executive Order 14110 on Safe, Secure, and Trustworthy Artificial Intelligence, issued in October 2023, mandates safety test reporting and red-team evaluations for dual-use foundation models.
  • The G7 Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems sets voluntary commitments for watermarking AI content and disclosing vulnerabilities.
  • India's Digital Personal Data Protection Act (DPDPA) 2023 mandates that data fiduciaries implement technical safeguards and obtain verifiable parental consent when processing personal data of minors for AI systems.
  • The IndiaAI Independent Business Division (IBD), under the Digital India Corporation, acts as the implementation agency for the IndiaAI Mission's supercomputing compute and application pillars.
  • Algorithmic impact assessments (AIAs) are systematic evaluation frameworks used by government agencies to assess public risks, civil liberty impacts, and societal harms before deploying algorithmic decision systems.
  • Disparate impact occurs when a facially neutral algorithm disproportionately disadvantages members of a protected class (such as race, gender, or age) without a legitimate business necessity.
  • Demographic parity (statistical parity) requires that the decision rate (e.g., loan approval rate) of an algorithm is identical across all demographic groups regardless of group membership.
  • Equalized odds requires that an algorithm has equal True Positive Rates and equal False Positive Rates across all demographic groups, ensuring error rates are distributed fairly.
  • Model interpretability refers to the degree to which a human observer can understand the internal mathematical mechanics and feature contributions leading to a machine learning output.
  • SHAP (Shapley Additive Explanations) uses cooperative game theory concepts to calculate the marginal contribution of each input feature to a model's final prediction.
  • LIME (Local Interpretable Model-agnostic Explanations) perturbs input data points and builds local linear surrogate models to explain specific individual black-box predictions.
  • Saliency maps in computer vision visualize gradient-weighted class activation mapping (Grad-CAM), highlighting the spatial image regions that most heavily influenced a convolutional network's decision.
  • Prompt injection attacks in LLMs occur when malicious user inputs manipulate system prompts to override safety instructions, leak confidential system data, or execute unauthorized tool calls.
  • Data poisoning attacks in machine learning involve maliciously manipulating or inserting tainted samples into training datasets to introduce deliberate backdoors or degrade model accuracy.
  • Model inversion attacks reconstruct private training data samples (such as patient medical images) by repeatedly querying a trained model's prediction probabilities.
  • Membership inference attacks determine whether a specific individual's data record was included in the private training dataset of a machine learning model.
  • AI watermarking embeds subtle statistical perturbations or cryptographic tokens into generated text, image, and audio files to enable automated provenance detection and verify authenticity.
  • The Coalition for Content Provenance and Authenticity (C2PA) standard establishes open technical specifications for embedding cryptographically verifiable metadata (content credentials) in digital media.
  • Autonomous Weapons Systems (AWS), often termed 'killer robots', are military weapons that select and engage targets without human intervention, raising international humanitarian law concerns at the UN CCW.
Curriculum & Reference Sources: European Parliament (EU AI Act), NITI Aayog (National Strategy for AI #AIforAll), Global Partnership on Artificial Intelligence (GPAI)

Sample Solved Questions & Concept Explanations

8 Verified Concept Questions
Q1.MEDIUM

The 'European Union AI Act' passed in 2024 classifies AI systems based on what governance framework?

Q2.HARD

What is 'Direct Preference Optimization' (DPO, Rafailov et al. 2023) in LLM post-training alignment?

Q3.HARD

In reinforcement learning from human feedback (RLHF), what is the primary role of the 'Kullback-Leibler (KL) Divergence Penalty' added to the PPO objective?

Q4.HARD

In mechanistic interpretability, what is 'Superposition' in neural network representations (Elhage et al., Anthropic 2022)?

Q5.EASY

What is "Algorithmic Bias" in artificial intelligence and machine learning systems?

Q6.EASY

What is a "Deepfake" in artificial intelligence media and cybersecurity?

Q7.EASY

What is the comprehensive European Union regulation enacted in 2024 to regulate artificial intelligence based on risk tiers (Unacceptable, High, and Low risk)?

Q8.EASY

What is the concept of "AI Explainability" (XAI / Explainable AI) in machine learning ethics?