Academic Databases
Explore Our Semantic Databases
STEM & Tech
Math, Physics, CS & Engineering...
Economics
Macro, Micro, Econometrics...
Business
Management, Marketing, Finance...
Health & Life Sciences
Biology, Medicine, Neuroscience...
Humanities & Social Sciences
Psychology, Law, Philosophy, History...
General Studies
General & interdisciplinary studies...
Explore Topics
Features

Algorithmic Bias in AI: 5 Types and Fairness Solutions

Added: 2026-08-06

Deep Dive

Basic understanding of the machine learning pipeline, specifically how models are trained on historical datasets to make predictions.

Class 5 Machine learning fundamentals part 2

Class 5 Machine learning fundamentals part 2

1.4K

The machine learning pipeline consists of several interconnected steps: (1) Data preparation and splitting into train/test sets, (2) Model selection (parametric or non-parametric), (3) Training the model by estimating parameters using optimization algorithms, (4) Making predictions on training, validation, and test sets, (5) Evaluating performance using appropriate metrics (RSS, accuracy, etc.), and (6) Hyperparameter tuning using cross-validation. The key insight is that the model doesn't learn anything until after parameter estimation occurs. Before training, the model makes arbitrary predictions based on initial parameter guesses. The learning happens during the optimization phase when the model adjusts its parameters to minimize the loss function. This distinction between model definition and model training is fundamental to understanding how machine learning systems actually work.

Click to watch →

Prediction: Categorical Variables | Full Course On Statistics for Data Science with Python.

Prediction: Categorical Variables | Full Course On Statistics for Data Science with Python.

345

The complete machine learning pipeline includes: (1) Test-train splitting (80-85% training, 20-15% testing) to prevent overfitting, (2) Model initialization specifying algorithm type and parameters, (3) Training using fit() on training data to learn patterns, and (4) Evaluation using test data. During training, the model learns input-output relationships from examples. Testing evaluates generalization to unseen data, providing unbiased performance assessment.

Click to watch →

Machine Learning Algorithms | Machine Learning Tutorial | Data Science Training | Edureka

Machine Learning Algorithms | Machine Learning Tutorial | Data Science Training | Edureka

374.4K

This extensive demonstration covers the complete machine learning workflow using R programming. The process begins with data preparation, where datasets are split into training and testing sets (typically 70/30 ratio) to prevent overfitting and enable unbiased evaluation. Using the mtcars dataset, the demonstration shows how to implement logistic regression for binary classification (predicting V-engine vs straight-engine based on weight and displacement). The workflow includes: loading data, splitting using sample() function, training the model with glm() and binomial family, making predictions with predict() function, and validating against actual values. The demonstration reveals how logistic regression produces probabilities thresholded at 0.5 for classification decisions, showing correct predictions for test cases. This end-to-end example illustrates the complete machine learning pipeline: data preparation, model training, prediction, and validation.

Click to watch →

Alteryx para Todos: Fundamentos, Aplicaciones y Machine Learning

Alteryx para Todos: Fundamentos, Aplicaciones y Machine Learning

821

Training a machine learning model involves: loading training data with historical examples, selecting the target variable to predict, selecting predictor variables (features), choosing a model type (like Random Forest), configuring parameters (defaults often work well), training the model on the data, and executing predictions on new data. The model learns patterns from training data and applies them to make predictions on new, unseen data. Predictions can be written to external storage like Google Drive for use in other tools.

Click to watch →

Building Machine Learning Pipeline using Scikit-Learn

Building Machine Learning Pipeline using Scikit-Learn

1.9K

A machine learning pipeline is a structured sequence of data preprocessing steps that automates the workflow of preparing data, training models, and making predictions. In scikit-learn, pipelines combine multiple preprocessing steps (such as imputation, feature scaling, dimensionality reduction, and encoding) with a final estimator into a single object that can be fitted and used for predictions. The pipeline handles both numerical data (using SimpleImputer with mean strategy, MinMaxScaler, and PCA) and categorical data (using SimpleImputer with constant strategy and OneHotEncoder) through ColumnTransformer, which applies different preprocessing steps to different column types. This approach ensures that all preprocessing steps are applied consistently during both training and prediction, preventing data leakage and simplifying the machine learning workflow.

Click to watch →

Deep Dive

The concept of training data representation and how data collection methods can introduce skew or underrepresentation.

Introduction - Machine Learning # 1

Introduction - Machine Learning # 1

83.2K

Training data must represent the cases you want to generalize to. If data is non-representative, the model performs poorly on new cases. The 1936 Literary Digest poll predicted Landon would win 57% of votes but Roosevelt won 62%, because their sampling method favored wealthier Republican voters (sampling bias). Similarly, excluding Switzerland from the GDP-life satisfaction analysis would produce a misleading regression line.

Click to watch →

Aprende a implementar herramientas para garantizar el uso ético de IA en sus políticas públicas

Aprende a implementar herramientas para garantizar el uso ético de IA en sus políticas públicas

547

Data collection methods can introduce systematic bias: (1) Phone surveys exclude those without phones, typically lower-income populations; (2) This creates a sample biased toward higher-income groups; (3) Models trained on such data will perform poorly when predicting for excluded populations; (4) The bias is not corrected by model accuracy metrics. This illustrates why understanding data collection methods is essential - the same data can produce different results depending on how it was collected, and models trained on biased data will perpetuate those biases.

Click to watch →

Machine Learning from a statistical perspective

Machine Learning from a statistical perspective

1.3K

Training sets must represent the target population for valid inferences. Random sampling provides good representation, but oversampling minority groups (stratification) is needed when particular subpopulations are underrepresented. Challenges arise when training data comes from subsets of subsets—for example, social media data represents only active users who post specific content. Inferences about the total population from such data require careful consideration. Practical approaches include manual checking of classifications and applying corrections for representativeness issues.

Click to watch →

Own and Claim your inspirations and dreams #iit #iitpatna

Own and Claim your inspirations and dreams #iit #iitpatna

31.3K

The speaker explains that AI models are trained on data that underrepresents certain groups, including Dalits (lower castes), women, minorities, and marginalized communities. Since these groups are underrepresented in the training data, AI models may not adequately represent their perspectives or needs. This creates a cycle where marginalized voices are further silenced because the technology that shapes society does not include them in its foundational data.

Click to watch →

IB SL Data Presentation

IB SL Data Presentation

647

Bar charts are used for discrete data with gaps between bars, while histograms are for continuous data with touching bars; skew describes whether data is bunched toward lower values (left/negatively skewed) or higher values (right/positively skewed), with the tail direction determining the skew classification.

Click to watch →

Deep Dive

Fundamental statistical concepts, such as correlation versus causation, and how historical patterns are replicated by algorithms.

Global Classroom Dr Rayenda

Global Classroom Dr Rayenda

241

Correlation does not imply causation. Classic examples include falling leaves correlating with stock market falls (neither causes the other) and ice cream sales correlating with shark attacks (both increase in summer, but neither causes the other). This distinction is foundational before learning machine learning, as it underlies the critical difference between estimation (causal inference) and prediction (probabilistic forecasting).

Click to watch →

Panel: Content oder Code? Sind Algorithmen die neuen Gatekeeper?

Panel: Content oder Code? Sind Algorithmen die neuen Gatekeeper?

565

Traditional statistical analysis sought to understand causality—why something happens. However, in the algorithmic age, there is a fundamental shift toward correlation. Algorithms no longer try to understand why events occur but instead identify patterns and configure outcomes based on observed correlations. Unlike traditional statistics that relied on data collected for specific purposes, modern algorithms use big data from diverse sources originally collected for entirely different reasons. This aggregation enables new types of analysis and decision-making that were previously impossible.

Click to watch →

Ep 13 : Quand corrélation → Causalité

Ep 13 : Quand corrélation → Causalité

37.7K

Correlation refers to a statistical relationship where two variables vary together, but this does not imply that one causes the other. Multiple explanations exist for observed correlations: one variable may cause the other, both may be caused by a third factor, or the relationship may occur by chance. This distinction is crucial because identifying causation enables intervention (prevention, treatment), while mere correlation provides no guidance for action. Francis Galton first clearly articulated this principle in the late 19th century, observing that while height correlates between parents and children, this does not mean that parental height directly causes offspring height.

Click to watch →

Data Mining in Arabic Lecture 1: Introduction to Data Mining - تنقيب البيانات بالعربي

Data Mining in Arabic Lecture 1: Introduction to Data Mining - تنقيب البيانات بالعربي

2.4K

Correlation refers to a relationship where two variables change together, while causation means one variable directly causes changes in another. The instructor emphasizes that correlation does not imply causation, using the example of ice cream sales and shark attacks both increasing in summer due to a hidden variable (summer season). Machine learning is a subset of artificial intelligence that focuses on building algorithms that can learn from data and improve performance over time. The instructor explains that machine learning enables computers to make decisions and predictions without being explicitly programmed for each specific task, using the analogy of teaching a child to recognize animals.

Click to watch →

CONSENSUS - Cryptos, Causation vs Correlation

CONSENSUS - Cryptos, Causation vs Correlation

2.5K

In crypto trading, historical patterns and narratives (like 'Consensus has happened three years in a row') often confuse correlation with causation, leading traders to incorrectly project past market conditions onto current environments; successful trading requires recognizing that market events are merely triggers or permission points within broader trends, not the actual cause of market movements, and traders must adapt their strategies to the current market environment rather than relying on repeated narratives.

Click to watch →

Deep Dive

A general awareness of how artificial intelligence is deployed in high-stakes decision-making areas like hiring, lending, and policing.

AI Fairness 360

AI Fairness 360

12.7K

AI is increasingly deployed in high-stakes decision-making contexts including credit loans, employment, college admissions, and criminal justice sentencing. Beyond accuracy, these applications require fairness, explainability, security, and accountability. Machine learning inherently involves statistical discrimination by identifying patterns to classify samples, but this becomes objectionable when applied to people and creates systemic disadvantage for certain groups. Unwanted biases in training data arise from prejudiced historical labels or disproportionate sampling of population groups, leading to biased models that perpetuate unfair outcomes.

Click to watch →

How Does Machine Bias Work? - DTNS 4367

How Does Machine Bias Work? - DTNS 4367

468

AI models are not perfect and will make errors. In high-stakes applications like parole decisions, even a 68% accuracy rate means significant errors occur—people may be wrongly kept in jail or wrongly released. The challenge is that when people's lives are at stake, our tolerance for error should be very low. This is why medical practitioners are more careful about blindly following algorithms—they know the consequences of errors. The problem is that many companies using AI for hiring and other decisions don't have the same awareness of the stakes involved, and the technology is often marketed as a magical solution rather than a tool that requires human oversight.

Click to watch →

Towards Trustworthy Artificial Intelligence

Towards Trustworthy Artificial Intelligence

134

Second-generation AI is pervasively deployed to make critical decisions about our lives, including employment (EEOC estimates over 90% of organizations use AI hiring tools), academic admissions, medical calculators and triage tools, policing and parole decisions, immigration and refugee adjudication, tax auditing, loans and mortgages, airport security, defense systems, government budget investment, political platforms, public health decisions, urban planning, and emergency preparedness.

Click to watch →

Learn AI or Be Replaced

Learn AI or Be Replaced

31.1K

AI is now being used in hiring decisions, with algorithms deciding whether candidates get interviews. AI is also being used in housing applications to determine apartment approvals, in welfare determinations to calculate assistance eligibility, and in schools for student surveillance. The ACLU has documented AI systems making life-altering decisions about people with zero transparency and zero recourse.

Click to watch →

[7/27 MON] 출근길 영어뉴스 I Morning Wave in Busan

[7/27 MON] 출근길 영어뉴스 I Morning Wave in Busan

1.3K

High-impact AI means AI used in important areas such as hiring, loans, healthcare, education, or public services. These systems can affect a person's job, money, health, or future. For example, if a bank rejects a loan application based on AI, the bank cannot just say 'the computer said no' - they must have stronger safety steps, manage risks, keep human control, and protect users. Similarly, if AI reads an X-ray incorrectly, it should not replace careful human review. When decisions are important, a person must check it and someone must be responsible if things go wrong.

Click to watch →

Bias Basics

Data Limits

Measurement Flaws

Feedback Loops

Fairness Checks

Forward Path

Bias Basics

0:00

Playing Section

1
Explains algorithmic bias, distinguishing it from harmful discrimination.
2
Highlights how hidden societal biases in training data skew AI outputs.

Deep Dive

Mathematical formulations of fairness metrics, such as demographic parity, equalized odds, and predictive rate parity, and the mathematical trade-offs between them.

CP2021 Tutorial "Tutorial in equity modeling"

CP2021 Tutorial "Tutorial in equity modeling"

204

Statistical fairness metrics from machine learning literature focus on bias between groups, partly due to legal requirements on fairness and affirmative action. These metrics are based on classical statistical measures of classification error. Examples include: demographic parity (same fraction of yes/no decisions across groups); equalized odds (equality of opportunity and equality of false positive rates); accuracy parity (considers correct decisions); predictive rate parity (compares selected individuals' actual outcomes across groups). These metrics substitute zero-one decisions for utility vectors, providing limited accounts of actual consequences.

Click to watch →

How to Make Machine Learning Fair and Accountable - Sray Agarwal | ODSC Europe 2019

How to Make Machine Learning Fair and Accountable - Sray Agarwal | ODSC Europe 2019

407

Multiple mathematical definitions quantify fairness: (1) Average Odds Ratio - difference between false positive rate and true positive rate should be equal across groups, (2) Equal Opportunity - true positive rate should be equal across groups, (3) Predictive Equality - false positive rate should be equal across groups, (4) Equalized Odds - both true positive and false positive rates should be equal across groups, (5) Demographic Parity - overall prediction distribution should be equal regardless of group membership. Equal Opportunity and Equalized Odds are most comprehensive, covering both false negatives and false positives. When implemented correctly, these metrics can improve while simultaneously reducing costs.

Click to watch →

Fairness Criteria, Exploring Fairness in Machine Learning

Fairness Criteria, Exploring Fairness in Machine Learning

11.9K

This video explains three key fairness criteria for machine learning models: demographic parity (requiring outcomes to be independent of protected attributes), equalized odds (matching both true positive and false positive rates across groups), and equalized opportunity (only requiring equal true positive rates for worthy individuals). The instructor demonstrates how these criteria relate to the confusion matrix components (true positives, true negatives, false positives, false negatives) and discusses their trade-offs, noting that demographic parity may inadvertently discriminate against qualified individuals while equalized odds provides stronger fairness guarantees but is harder to achieve.

Click to watch →

Machine Learning NeEDS Mathematical Optimization with Prof Daniel Kuhn

Machine Learning NeEDS Mathematical Optimization with Prof Daniel Kuhn

500

Fairness through unawareness (removing sensitive attributes) fails when features remain correlated with sensitive attributes—multiple slightly correlated features can predict sensitive attributes with high accuracy. Statistical parity requires predictions to be statistically independent of the sensitive attribute, meaning identical cumulative distribution functions across groups. Equalized odds provides an alternative by ensuring equal benefit probabilities for equally qualified individuals. Both represent group fairness criteria requiring conditional distributions to be equal, though they capture different aspects of fairness in decision-making systems.

Click to watch →

FAT* 2018: Nicole Immorlica - Decoupled Classifiers for Group-Fair and Efficient Machine Learning

FAT* 2018: Nicole Immorlica - Decoupled Classifiers for Group-Fair and Efficient Machine Learning

367

Demographic parity requires equalizing the fraction of positives across groups, expressed as |positive rate minority - positive rate majority|. Equal odds aims to equalize error rates across populations. Joint loss functions combine accuracy and fairness terms with specified trade-offs, such as 3/4 weight on total error (accuracy) and 1/4 on fairness.

Click to watch →

Deep Dive

Hands-on application of bias mitigation frameworks and toolkits, such as Fairlearn, IBM's AI Fairness 360, or Google's What-If Tool.

Bias and Fairness in AI Systems with Lead Machine Learning Developer at AltaML, Graham Erickson

Bias and Fairness in AI Systems with Lead Machine Learning Developer at AltaML, Graham Erickson

1.2K

Three major open source toolkits assist in addressing AI bias: (1) Fairlearn (Microsoft-backed) - provides easy-to-use fairness metrics and mitigation techniques; (2) AI Fairness 360 (IBM-backed) - offers comprehensive fairness metrics, tutorials, and mitigation algorithms; (3) What-If Tool (Google) - focuses on counterfactual reasoning and explainability, allowing users to observe how changes in features and demographics affect predictions.

Click to watch →

Bias and Fairness Testing in AI Models - real life examples

Bias and Fairness Testing in AI Models - real life examples

617

AI models can inherit and amplify biases present in their training data, leading to unfair decisions in real-world applications such as hiring, healthcare, and credit scoring; testers must use fairness testing tools like Fairlearn, IBM AI Fairness 360, and Google What-If Tool to detect demographic disparities and implement mitigation strategies including improving data collection, applying preprocessing/postprocessing fixes, and adjusting model algorithms to ensure ethical and equitable AI deployment.

Click to watch →

Gabriela de Queiroz: vieses em inteligência artificial e ferramentas open source #CodaBr20

Gabriela de Queiroz: vieses em inteligência artificial e ferramentas open source #CodaBr20

682

Several tools exist to help calculate and mitigate AI biases: FATE (Microsoft), What-If Tool (Google), and AI Fairness 360 (IBM, now donated to Linux Foundation). AI Fairness 360 provides metrics and algorithms to examine discrimination in data and models. Bias mitigation occurs at three stages: pre-processing (intervening in data before training), in-processing (intervening during training), and post-processing (intervening in predictions). Data scientists must look beyond numbers and consider the implications of what they create—whether their work will impact people's lives and what kind of implications that will have. When reusing code, developers must be careful because code reflects the thinking of those who created it and may perpetuate existing biases.

Click to watch →

Responsible AI in Practice with Sarah Bird - #322

Responsible AI in Practice with Sarah Bird - #322

310

The FairLearn toolkit brings together various fairness techniques published in research, enabling users to compare how well models work for different demographic groups. There are over 20 fairness metrics, and the toolkit includes common ones. Mitigation techniques like FairLearn allow users to pick a metric and groups, then the algorithm retrains the model by reweighting data to reduce disparity. This produces a series of models that can be compared to find versions that work better for desired fairness properties.

Click to watch →

AI Bias Testing and Mitigation: How to Choose the Right Open Source Tools - Melinda Thielbar

AI Bias Testing and Mitigation: How to Choose the Right Open Source Tools - Melinda Thielbar

308

Audit AI provides straightforward tests with minimal code, ideal for quick yes/no answers on specific models. AIF360 is a comprehensive toolkit designed for production environments where multiple models need automatic auditing and debiasing. The What-If Tool from Google is an interactive visualization tool that works with any model but is easiest with TensorFlow, making it suitable for exploration rather than automated production testing.

Click to watch →

Deep Dive

Explainable AI (XAI) techniques, including SHAP (SHapley Additive exPlanations) and LIME, to audit and interpret black-box model decisions.

Explicabilidad en Machine Learning - "Data & Analytics - Innovación y Tecnología"

Explicabilidad en Machine Learning - "Data & Analytics - Innovación y Tecnología"

191

LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) are two prominent techniques for explaining predictions from complex, non-interpretable models. LIME works by locally approximating complex models with simpler, interpretable models like linear regression, allowing users to understand which features contributed most to a particular prediction. SHAP is based on cooperative game theory, treating each feature as a 'player' and calculating how much each feature contributes to the final prediction by considering all possible combinations of features. Both techniques transform black-box models into explainable ones, enabling users to understand and trust model decisions.

Click to watch →

XAI Review - 2. LIME & SHAP

XAI Review - 2. LIME & SHAP

6.4K

LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) are two essential model-agnostic methods for explaining black-box machine learning predictions. LIME approximates complex models locally using linear models around specific data points, making it ideal for understanding individual predictions but potentially slower with many features. SHAP, based on game theory's Shapley values, calculates feature contributions by comparing model predictions with and without each feature, providing both local and global explanations while ensuring consistency across different feature combinations. Both methods enable users to understand which features drive predictions, enhancing model trust and interpretability across various applications from healthcare to finance.

Click to watch →

Introduction to Explainable AI (XAI) | Interpretable models, agnostic methods, counterfactuals

Introduction to Explainable AI (XAI) | Interpretable models, agnostic methods, counterfactuals

18.7K

Explainable AI (XAI), also known as interpretable machine learning (IML), is a field focused on making machine learning models transparent and understandable to humans. XAI employs two main approaches: intrinsically interpretable models (simple models like decision trees and linear regression whose parameters can be directly understood) and model agnostic methods (techniques like SHAP, LIME, PDPs, and ICE plots that can explain any blackbox model without requiring knowledge of its internal structure). Additional approaches include causal models, counterfactual explanations, adversarial examples, and model-specific methods. The field addresses the critical need for transparency in AI systems, especially when decisions affect people's lives such as mortgage approvals or medical diagnoses.

Click to watch →

Explainable AI (xAI) | Nubank DS & ML Meetup #77

Explainable AI (xAI) | Nubank DS & ML Meetup #77

1.3K

This section covers the most popular XAI techniques. SHAP (SHapley Additive exPlanations) uses game theory to assign importance values to features by considering all possible feature combinations. LIME (Local Interpretable Model-agnostic Explanations) creates local surrogate models to explain individual predictions. Both techniques produce visualizations like bar charts and scatter plots. The speaker notes that different techniques can produce different explanations for the same model and prediction, and neither is inherently correct. Alternative techniques include Partial Dependence Plots (PDP) for global feature importance, counterfactual explanations showing what changes would alter predictions, and model simplification approaches. The speaker emphasizes that multiple techniques are needed because different techniques address different aspects of model interpretability.

Click to watch →

A Guide to Explainable AI | Artificial Intelligence Masterclass for Beginners

A Guide to Explainable AI | Artificial Intelligence Masterclass for Beginners

18.3K

Three fundamental post hoc XAI techniques for interpreting complex models: (1) LIME (Local Interpretable Model-Agnostic Explainability) zooms into specific predictions by fitting a simple linear model to data points in the local neighborhood, simplifying complex non-linear models to linear approximations for that area; (2) SHAP (SHapley Additive exPlanations) uses game theory to compute feature contributions by randomly selecting feature subsets, inserting random values for remaining features, and comparing predictions to determine each feature's impact; (3) Partial Dependence Plots (PDP) study how target variables change as a single feature varies across its range while keeping other features constant, revealing feature importance and identifying thresholds where predictions become confident versus uncertain. Each technique serves different analytical purposes: LIME for local explanations, SHAP for quantified feature contributions, and PDP for understanding feature-target relationships.

Click to watch →

Deep Dive

Global AI governance, policy, and legal frameworks, such as the EU AI Act, designed to regulate algorithmic discrimination and enforce compliance.

📚 AI ACT: il nuovo REGOLAMENTO europeo sull'INTELLIGENZA ARTIFICIALE

📚 AI ACT: il nuovo REGOLAMENTO europeo sull'INTELLIGENZA ARTIFICIALE

7.6K

The AI Act (Regulation 1689/2024) is Europe's comprehensive AI regulation, created because AI, while ubiquitous in schools, hospitals, and devices, can be exploited maliciously. Unlike widely-used tools like Excel, AI can perpetuate discrimination through biased training data, enable mass surveillance through social scoring systems, and threaten fundamental rights. The EU proposed the regulation in April 2021, approved it in spring 2024, and published it in July 2024. The Act employs a risk-based approach similar to GDPR: the more a system impacts safety or fundamental rights, the stricter the rules. It categorizes AI into four tiers: prohibited practices (8 banned activities), high-risk systems (medical, employment, credit, education, immigration, justice), limited-risk systems (chatbots, recommendation engines), and minimum-risk systems (spam filters, game AI). General-Purpose AI models (GPAI) like Chat GPT have specific obligations. The Act establishes four roles: providers (developers), deployers (users), importers (non-EU systems), and distributors (retailers), each with tailored obligations. Italy implemented the Act through Law 132/2025, establishing governance through existing authorities. The Act does not ban AI but establishes 'rails' to channel development while protecting fundamental rights through human-in-the-loop principles.

Click to watch →

AI Regulations Explained: EU AI Act, Colorado Law, and NIST Framework

AI Regulations Explained: EU AI Act, Colorado Law, and NIST Framework

368

Organizations can navigate the complex landscape of AI regulations by using the EU AI Act as a baseline compliance framework, then measuring the delta against other regulations like Colorado's algorithmic discrimination law and the NIST voluntary framework, implementing automated discovery tools like AI Control Tower to maintain visibility across all AI systems, and establishing governance policies with clear ownership across product, compliance, legal, and engineering teams to proactively address compliance requirements before enforcement deadlines such as August 2026 for the EU AI Act and June 2026 for Colorado.

Click to watch →

【陽明交大科法教授學術論壇】NYCU School of Law 公司作為人工智慧治理代言人：論歐盟人工智慧規則草案 (2022/10/25)

【陽明交大科法教授學術論壇】NYCU School of Law 公司作為人工智慧治理代言人：論歐盟人工智慧規則草案 (2022/10/25)

119

The EU AI Act proposes to be the first binding international legal instrument regulating AI, establishing a model for global governance through the 'Brussels Effect.' The EU operates as an international organization based on treaties between 27 member states, transferring specific competencies to achieve common objectives like creating an internal market. The Court of Justice interprets treaties and decides what constitutes obstacles to free movement, enabling individuals to enforce EU rights directly in national courts. The AI Act pursues multiple objectives under Article 114 TFEU: facilitating single market development while ensuring safety, fundamental rights protection, and compliance with Union law. The goal is 'trustworthy AI'—clear, predictable rules enabling innovation while protecting citizens.

Click to watch →

AI Policy Insights with Dan Nechita

AI Policy Insights with Dan Nechita

264

The EU AI Act represents the world's first binding regional AI regulation, implementing a risk-based framework that categorizes AI systems into prohibited (mass surveillance, social scoring), high-risk (law enforcement, workplace monitoring, education), and transparent (chatbots, manipulated media) categories, with compliance timelines ranging from 6 to 24 months depending on risk level; the Act balances innovation support through sandboxes, startup exemptions, and the AI Pact initiative with stringent safeguards for powerful AI models, while pursuing international alignment through OECD and multilateral forums to establish global AI governance standards.

Click to watch →

Is Europe Really Losing the AI Race Because of Regulation? InnovEU S5E4

Is Europe Really Losing the AI Race Because of Regulation? InnovEU S5E4

100

The EU AI Act, co-reported by Brando Benifei, establishes Europe's first comprehensive legal framework for artificial intelligence. Its primary purpose is to guarantee protections where risks exist while enabling safe adoption of AI systems. The Act balances innovation with fundamental rights protection, ensuring AI remains human-centric and prevents health and safety risks. Globally, multiple countries including South Korea, US states, Canada, Japan, Brazil, and African nations are developing similar AI regulatory frameworks, indicating a shared recognition that AI requires governance.

Click to watch →

Bias Basics

Data Limits

Measurement Flaws

Feedback Loops

Fairness Checks

Forward Path

Bias Basics

0:00

Playing Section

1
Explains algorithmic bias, distinguishing it from harmful discrimination.
2
Highlights how hidden societal biases in training data skew AI outputs.

231.6K views3.9K11:20@crashcourseOriginal Release: 2019-12-13

Algorithmic bias occurs when AI systems replicate or amplify existing human biases through five main mechanisms: (1) training data reflecting hidden societal biases, such as associating certain professions with specific genders; (2) unbalanced classes in training data leading to poor performance on underrepresented groups; (3) inability to properly quantify complex human characteristics like relationships or creativity; (4) feedback loops where algorithm predictions influence future data collection, creating self-reinforcing cycles of discrimination; and (5) malicious manipulation of training data by humans. These biases can result in unfair treatment in applications like hiring, law enforcement, and automated decision-making, making it essential to critically evaluate AI recommendations and advocate for transparency and fairness in algorithmic systems.

[00:00:00]Hi, I’m Jabril and welcome back to CrashCourse AI.

[00:00:03]Algorithms are just math and code, but algorithms are created by people and use our data, so biases that exist in the real world are mimicked or even exaggerated by AI systems.

[00:00:14]This idea is called algorithmic bias.

[00:00:16]Bias isn’t inherently a terrible thing.

[00:00:19]Our brains try to take shortcuts by finding patterns in data.

[00:00:23]So if you’ve only seen small, tiny dogs, you might see a Great Dane and be like “Whoa that dog is unnatural” This doesn’t become a problem unless we don’t acknowledge exceptions to patterns or unless we start treating certain groups of people unfairly.

[00:00:37]As a society, we have laws to prevent discrimination based on certain “protected classes” (like gender, race, or age) for things like employment or housing.

[00:00:46]So it’s important to be aware of the difference between bias, which we all have, and discrimination, which we can prevent.

[00:00:52]And knowing about algorithmic bias can help us steer clear of a future where AI are used in harmful, discriminatory ways.

[00:00:59]INTRO There are at least 5 types of algorithmic bias we should pay attention to.

[00:01:12]First, training data can reflect hidden biases in society.

[00:01:15]For example, if an AI was trained on recent news articles or books, the word “nurse” is more likely to refer to a “woman,” while the word “programmer” is more likely to refer to a “man.” And you can see this happening with a Google image search: “nurse” shows mostly women, while “programmer” mostly shows mostly men.

[00:01:32]We can see how hidden biases in the data gets embedded in search engine AI.

[00:01:36]Of course, we know there are male nurses and female programmers and non-binary people doing both of these jobs!

[00:01:42]For example, an image search for “programmer 1960” shows a LOT more women.

[00:01:47]But AI algorithms aren’t very good at recognizing cultural biases that might change over time, and they could even be spreading hidden biases to more human brains.

[00:01:55]t’s also tempting to think that if we just don’t collect or use training data that categorizes protected classes like race or gender, then our algorithms can’t possibly discriminate.

[00:02:04]But, protected classes may emerge as correlated features, which are features that aren’t explicitly in data but may be unintentionally correlated to a specific prediction.

[00:02:14]For example, because many places in the US are still extremely segregated, zip code can be strongly correlated to race.

[00:02:20]A record of purchases can be strongly correlated to gender.

[00:02:24]And a controversial 2017 paper showed that sexual orientation is strongly correlated with characteristics of a social media profile photo.

[00:02:32]Second, the training data may not have enough examples of each class, which can affect the accuracy of predictions.

[00:02:38]For example, many facial recognition AI algorithms are trained on data that includes way more examples of white peoples’ faces than other races.

[00:02:46]One story that made the news a few years ago is a passport photo checker with an AI system to warn if the person in the photo had blinked.

[00:02:54]But the system had a lot of trouble with photos of people of Asian descent.

[00:02:58]Being asked to take a photo again and again would be really frustrating if you’re just trying to renew your passport, which is already sort of a pain!

[00:03:05]Or, let’s say, you got a cool gig programming a drone for IBM… but it has trouble recognizing your face because your skin’s too dark… for example.

[00:03:15]Third, it’s hard to quantify certain features in training data.

[00:03:19]There are lots of things that are tough to describe with numbers.

[00:03:22]Like can you really rate a sibling relationship with a number?

[00:03:25]It’s complicated!

[00:03:26]You love them, but you hate how messy they are, but you like cooking together, but you hate how your parents compare you... It’s so hard to quantify all that!

[00:03:35]In many cases, we try to build AI to evaluate complicated qualities of data, but sometimes we have to settle for easily measurable shortcuts.

[00:03:43]One recent example is trying to use AI to grade writing on standardized tests like SATs and GREs with the goal to save human graders time.

[00:03:51]Good writing involves complex elements like clarity, structure, and creativity, but most of these qualities are hard to measure.

[00:03:58]So, instead, these AI focused on easier-to-measure elements like sentence length, vocabulary, and grammar, which don’t fully represent good writing… and made these AIs easier to fool.

[00:04:09]Some students from MIT built a natural language program to create essays that made NO sense, but were rated highly by these grading algorithms.

[00:04:16]These AIs could also potentially be fooled by memorizing portions of “template” essays to influence the score, rather than actually writing a response to the prompt, all because of the training data that was used for these scoring AI.

[00:04:28]Fourth, the algorithm could influence the data that it gets, creating a positive feedback loop.

[00:04:33]A positive feedback loop basically means “amplifying what happened in the past”… whether or not this amplification is good.

[00:04:39]An example is PredPol’s drug crime prediction algorithm, which has been in use since 2012 in many large cities including LA and Chicago.

[00:04:47]PredPol was trained on data that was heavily biased by past housing segregation and past cases of police bias.

[00:04:53]So, it would more frequently send police to certain neighborhoods where a lot of racial minority folks lived.

[00:04:59]Arrests in those neighborhoods increased, that arrest data was fed back into the algorithm, and the AI would predict more future drug arrests in those neighborhoods and send the police there again.

[00:05:08]Even though there might be crime in neighborhoods where police weren’t being sent by this AI, because there weren't any arrests in those neighborhoods, data about them wasn’t fed back into the algorithm.

[00:05:18]While algorithms like PredPol are still in use, to try and manage these feedback effects, there is currently more effort to monitor and adjust how they process data.

[00:05:27]So basically, this would be like a new principal who was hired to improve the average grades of a school, but he doesn’t really care about the students who already have good grades.

[00:05:36]He creates a watchlist of students who have really bad grades and checks up on them every week, and he ignores the students who keep up with good grades.

[00:05:44]If any of the students on his watchlist don’t do their homework that week, they get punished.

[00:05:50]But all of the students NOT on his watchlist can slack on their homework, and get away with it based on “what happened in the past.” This is essentially what’s happening with PredPol, and you can be the judge if you believe it’s fair or not.

[00:06:02]Finally, a group of people may mess with training data on purpose.

[00:06:06]For example, in 2014, Microsoft released a chatbot named Xiaoice in China.

[00:06:11]People could chat with Xiaoice so it would learn how to speak naturally on a variety of topics from these conversations.

[00:06:17]It worked great, and Xiaoice had over 40 million conversations with no incidents.

[00:06:21]In 2016, Microsoft tried the same thing in the U.S. by releasing the Twitterbot Tay.

[00:06:26]Tay trained on direct conversation threads on Twitter, and by playing games with users where they could get it to repeat what they were saying.

[00:06:33]In 12 hours after its release, after a “coordinated attack by a subset of people” who biased its data set, Tay started posting violent, sexist, anti-semitic, and racist Tweets.

[00:06:43]This kind of manipulation is usually framed as “joking” or “trolling,” but the fact that AI can be manipulated means we should take algorithmic predictions with a grain of salt.

[00:06:52]This is why I don’t leave John-Green-Bot alone online… The common theme of algorithmic bias is that AI systems are trying to make good predictions, but they make mistakes.

[00:07:02]Some of these mistakes may be harmless or mildly inconvenient, but others may have significant consequences.

[00:07:08]To understand the key limitations of AI in our current society, let’s go to the Thought Bubble.

[00:07:13]Let’s say there’s an AI system called HireMe! that gives hiring recommendations to companies.

[00:07:18]HireMe is being used by Robots Weekly, a magazine where John-Green-bot applied for an editorial job.

[00:07:24]Just by chance, the last two people named “John” got fired from Robots Weekly and another three “Johns” didn’t make it through the hiring process.

[00:07:31]So, when John-Green-Bot applies for the job, HireMe! predicts that he’s only 24% likely to be employed by the company in 3 years.

[00:07:39]Seeing this prediction, the hiring manager at Robots Weekly rejects John-Green-bot, and this data gets added to the HireMe!

[00:07:46]AI system.

[00:07:47]John-Green-Bot is just another “John” that got rejected, even though he may have been the perfect robot for the job!

[00:07:52]Now, future “Johns” have an even lower chance to be hired.

[00:07:56]It’s a positive feedback loop, with some pretty negative consequences for John-Green-Bot.

[00:08:01]Of course, being named “John” isn’t a protected class, but this could apply to other groups of people.

[00:08:07]Plus, even though algorithms like HireMe!

[00:08:09]Are great at establishing a link between two kinds of data, they can’t always clarify why they’re making predictions.

[00:08:16]For example, HireMe! may find that higher age is associated with lower knowledge of digital technologies, so the AI suggests hiring younger applicants.

[00:08:25]Not only is this illegally discriminating against the protected class of “age,” but the implied link also might not be true.

[00:08:32]John-Green-bot may be almost 40, but he runs a robot blog and is active in online communities like Nerdfighteria!

[00:08:39]So it’s up to humans interacting with AI systems like HireMe! to pay attention to recommendations and make sure they’re fair, or adjust the algorithms if not.

[00:08:48]Thanks, Thought Bubble!

[00:08:50]Monitoring AI for bias and discrimination sounds like a huge responsibility, so how can we do it?

[00:08:56]The first step is just understanding that algorithms will be biased.

[00:09:00]It’s important to be critical about AI recommendations, instead of just accepting that “the computer said so.” This is why transparency in algorithms is so important, which is the ability to examine inputs and outputs to understand why an algorithm is giving certain recommendations.

[00:09:15]But that's easier said than done when it comes to certain algorithms, like deep learning methods.

[00:09:20]Hidden layers can be tricky to interpret.

[00:09:23]Second, if we want to have less biased algorithms, we may need more training data on protected classes like race, gender, or age.

[00:09:30]Looking at an algorithm’s recommendations for protected classes may be a good way to check it for discrimination.

[00:09:36]This is kind of a double-edged sword, though.

[00:09:39]People who are part of protected classes may (understandably) be worried about handing over personal information.

[00:09:44]It may feel like a violation of privacy, or they might worry that algorithms will be misused to target rather than protect them.

[00:09:52]Even if you aren’t actively working on AI systems, knowing about these algorithms and staying informed about artificial intelligence are really important as we shape the future of this field.

[00:10:02]Anyone, including you, can advocate for more careful, critical interpretation of algorithmic outputs to help protect human rights.

[00:10:09]Some people are even advocating that algorithms should be clinically tested and scrutinized in the same way that medicines are.

[00:10:15]According to these opinions, we should know if there are “side effects” before integrating AI in our daily lives.

[00:10:21]There’s nothing like that in the works yet.

[00:10:24]But it took over 2400 years for the Hippocratic Oath to transform into current medical ethics guidelines.

[00:10:30]So it may take some time for us to come up with the right set of practices.

[00:10:34]Next time, we have a lab and I’ll demonstrate how there are biases in even simple things like trying to adopt a cat or a dog.

[00:10:41]I’ll see ya then.

[00:10:42]Speaking of understanding how bias and misinformation spread, you should check out this video on Deep Fakes I did with Above the Noise -- another PBSDS channel that gets into the research behind controversial issues.

[00:10:53]Head over to the video in the description to find out how detect deep fakes.

[00:10:57]Tell them Jabril sent you!

[00:10:59]Crash Course AI is produced in association with PBS Digital Studios!

[00:11:03]If you want to help keep all Crash Course free for everybody, forever, you can join our community on Patreon.

[00:11:08]And if you want to learn more about prejudice and discrimination in humans, you can check out this episode of Crash Course Sociology.

Up Next

Meaningful Human Control of AI: A Framework for Responsible Innovation

@xomnia-BV

136 views•2021-01-27

Building Real-Time ML Pipelines with Feature Stores and MLOps Frameworks

@ODSCAI

5.1K views•2022-02-20

Bypassing Tor Censorship: Bridges and Pluggable Transport Guide

@Coding_ForEveryone

397 views•2024-06-11

Neural Networks Explained: Math, Layers, and Learning Fundamentals

@3blue1brown

21.9M views•2017-10-05

Related Study Plans & Knowledge Roadmaps

Structured learning paths in Artificial Intelligence

View All Artificial Intelligence Topics

artificial-intelligence

Building Recommenders: Two-Tower & Vector DBs

01Foundations of Python & Machine Learning Basics

02Introduction to Recommender Systems

03Deep Learning & Representation Learning

+ 3 more learning modules

Explore Roadmap

artificial-intelligence

Latent Diffusion: PyTorch, UNet & CLIP

01Deep Learning & PyTorch Foundations

02Convolutional Networks & Autoencoders

03The Math and Mechanics of Diffusion Models

+ 3 more learning modules

Explore Roadmap

artificial-intelligence

NeRFs: 3D Reconstruction & Instant-NGP

01Math & Python Foundations for 3D Vision

02Photogrammetry & Camera Pose Estimation

03Neural Radiance Fields (NeRF) Core Concepts

+ 2 more learning modules

Explore Roadmap

artificial-intelligence

TFT Forecasting: Energy Demand & Deep Learning

01Programming & Data Science Foundations

02Introduction to Machine Learning & Time Series

03Deep Learning & Sequential Models

+ 3 more learning modules

Explore Roadmap