Exam Notes Assistance

I will ask some topics which i need for my exam and this is the paper i was told to refer. Can you
help me so that i get notes for my university exam

Yes! I can help you prepare notes for your university exam based on the document you uploaded. Let
me know the specific topics you need, and I'll summarize them for you.

https://www.aporia.com/learn/fairness-metrics-in-machine-learning/?authuser=0
https://medium.com/@james.irving.phd/blog-post-series-ai-fairness-360-mitigating-bias-in-
machine-learning-models-2268e01584bd

These are the other two sources as well

Ensuring fairness in machine learning models is essential to prevent biased outcomes that can adversely
affect individuals or groups. The resources you've provided offer comprehensive insights into fairness
metrics and tools for bias mitigation. Here's a consolidated overview:​
Aporia+2

Common Fairness Metrics in Machine Learning:

1. Statistical Parity Difference (SPD): Also known as disparate impact, SPD measures the difference
in positive outcome rates between privileged and disadvantaged groups. A value close to zero
indicates fairness. ​Aporia

2. Equal Opportunity: This metric evaluates whether individuals from different groups have equal
chances of receiving a positive outcome, focusing on true positive rates across groups.​Aporia
3. Demographic Parity: Ensures that the likelihood of a positive outcome is the same across all
demographic groups, irrespective of their protected attributes.​Aporia

AI Fairness 360 Toolkit:

Developed by IBM, AI Fairness 360 (AIF360) is an open-source toolkit designed to detect and mitigate
bias in machine learning models. It offers:​ Medium+6

Pre-Processing Algorithms: Techniques applied to training data to improve fairness before model
training. For instance, reweighing assigns weights to training samples to balance representation
across groups. ​Medium+1
In-Processing Algorithms: Methods that modify the learning algorithm itself to incorporate
fairness constraints during model training.​

Post-Processing Algorithms: Techniques applied after model training to adjust predictions and
enhance fairness without retraining the model.​

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 1/12
Implementing these algorithms involves creating dataset objects, computing fairness metrics, and
applying appropriate bias mitigation strategies. For detailed guidance and code examples, refer to the
blog series on AI Fairness 360. ​ Medium+3
General Practices to Ensure Fairness in Machine Learning:

Diverse and Representative Data Collection: Ensure that the training data encompasses all
relevant groups to prevent underrepresentation.​

Regular Bias Audits: Periodically evaluate models using fairness metrics to detect and address
biases promptly.​Medium
Transparency: Maintain clear documentation of data sources, model decisions, and bias mitigation
steps to foster accountability.​

By integrating these metrics and tools, you can develop machine learning models that are both accurate
and equitable, minimizing unintended biases and promoting fairness.​

Sources

Understanding Bias in AI

Thought for 4 seconds

Bias in artificial intelligence (AI) refers to systematic errors or prejudices in an AI system’s decision‐
making that lead to unfair outcomes. These biases can originate from multiple sources and have
significant societal impacts. Below is a structured overview to help you understand this complex topic:

1. Definition of Bias in AI

Concept:
Bias in AI is when a model’s outputs consistently deviate from fair, unbiased results, often
reflecting historical or systemic prejudices. While “bias” in statistics can simply mean a systematic
deviation, in AI it implies unfair treatment of certain individuals or groups.

2. Sources of Bias

Bias in AI systems can emerge from several areas during the development lifecycle:

Data Bias:
Occurs when the training data is unrepresentative or contains historical prejudices. For
example, facial recognition systems often perform poorly on darker-skinned individuals if the
training data mostly features light-skinned faces.
Metrics like statistical parity and demographic parity are used to assess how balanced the
data is.
sap.com

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 2/12
 Algorithmic Bias:

Arises from the design and optimization of the AI model itself. The chosen model might
inadvertently prioritize certain features or correlations that reflect societal stereotypes.
mdpi.com
Human Decision Bias:

Introduced during data labeling, feature selection, or model evaluation, reflecting the
conscious or unconscious biases of developers and stakeholders.

Emergent Bias:
Develops over time when AI systems operate in new contexts or when feedback loops
reinforce discriminatory outcomes (e.g., predictive policing algorithms that concentrate more
police presence in already over-policed communities).
accuray.com

3. Impacts of Bias in AI

Individual and Societal Effects:

Discrimination: AI bias can reinforce and amplify existing inequalities in areas such as
healthcare, employment, criminal justice, and credit scoring. For instance, biased hiring tools
might undervalue resumes from women or minority groups.
zendata.dev
Economic and Social Inequities:

Biased systems may restrict access to opportunities and services, further entrenching
disparities.

Trust and Accountability:

Unchecked bias undermines public trust in AI systems and can lead to legal challenges
and reputational harm for organizations.

4. Mitigation Strategies

Researchers and practitioners have developed several approaches to detect and reduce bias:

Data Preprocessing:

Techniques like reweighing, dataset augmentation, or transforming data into bias-free

representations aim to balance the training data before model training.
en.wikipedia.org
In-Processing Methods:

These involve modifying the learning algorithm during training to impose fairness constraints
(e.g., adversarial debiasing, where the model is trained to both predict outcomes and
“unlearn” sensitive attributes).
sap.com

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 3/12
 Post-Processing Techniques:

Adjusting the model’s outputs (e.g., modifying decision thresholds) to ensure equitable
outcomes across different groups.
en.wikipedia.org
Ongoing Audits and Transparency:

Continuous monitoring using fairness metrics (such as statistical parity, equal opportunity,
and calibration) helps identify biases as models evolve. Tools and frameworks like AI Fairness
360 provide practical guidance and code examples.
holisticai.com

5. Fairness Metrics

Statistical Parity/Demographic Parity:

Measures if the probability of a positive outcome is the same for different groups.
Equal Opportunity:

Ensures that true positive rates (or recall) are equal across groups.
Predictive Parity:
Checks if the precision (positive predictive value) is balanced.

Calibration:
Validates that individuals with the same predicted probability score have similar actual
outcome rates regardless of group membership.
en.wikipedia.org

Conclusion

Understanding bias in AI involves recognizing that it is not simply a technical flaw but a reflection of
deeper societal issues. Addressing bias requires a holistic approach—starting from diverse,
representative data; implementing fair algorithmic design; and maintaining transparency and
continuous evaluation. These strategies are essential to ensure AI systems benefit all users equitably
and support ethical, trustworthy decision-making.
By integrating these concepts, you can build a comprehensive set of notes for your exam on bias in AI.
Let me know if you need further details on any specific aspect!

Types of Bias(Algorithmic, Data, Societal)

Reasoned for 8 seconds

There are three broad categories of bias in AI, each stemming from different stages of the system's
lifecycle:

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 4/12
 Algorithmic Bias:
This bias arises from the design, assumptions, or implementation choices in the AI model. It can
result from how the model is structured, which loss functions are chosen, or how the optimization
process is set up. These factors can unintentionally favor certain outcomes over others,
independent of the training data.

Data Bias:
Data bias occurs when the training data is unrepresentative, incomplete, or reflects historical
prejudices. For instance, if a dataset over-represents one demographic or contains measurement
errors, the model will likely learn and propagate these biases. Examples include sampling bias
(when the data doesn’t cover all groups) and label bias (when the outcomes in the data are already
skewed).
Societal Bias:
Societal bias reflects the broader social and cultural prejudices that exist in society. These biases
seep into AI systems through the data they are trained on and even through the perspectives of
the people designing them. As a result, AI systems can inadvertently reinforce and amplify existing
social inequities, such as discrimination based on race, gender, or socioeconomic status.

Together, these biases interact and compound one another, making it crucial to address each layer—
from ensuring fair algorithms and representative data to understanding and mitigating the impact of
broader societal inequalities—in order to create AI systems that are both accurate and equitable.

Fairness metrics

Reasoned about fairness metrics for 10 seconds

Fairness metrics are quantitative measures designed to assess whether an AI system produces equitable
outcomes across different groups. They help identify if the system's decisions disproportionately favor
or disadvantage particular groups. Here are key categories and examples:

1. Group Fairness Metrics

These metrics evaluate disparities between defined demographic groups (e.g., based on gender, race,
age):

Statistical/Demographic Parity:
Checks if the probability of a positive outcome is the same across groups. For example, if 60% of
one group receives a positive decision, the same should hold for others.
Equal Opportunity:
Focuses on equal true positive rates (TPR) across groups, ensuring that qualified individuals are
equally likely to receive a positive outcome regardless of their group membership.
Predictive Parity:
Requires that the positive predictive value (PPV) is balanced across groups. In other words, when
the model predicts a positive outcome, the accuracy of that prediction should be similar for all
groups.

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 5/12
 Equalized Odds:
A stricter criterion combining equal TPR and equal false positive rates (FPR) across groups. This
means that both the sensitivity and the error rate are balanced.

2. Individual Fairness Metrics

Fairness Through Awareness:

Proposes that similar individuals should receive similar outcomes. This requires defining a task-
specific similarity metric and ensuring the decision function is Lipschitz continuous with respect to
that metric.

3. Causal Fairness Metrics

Counterfactual Fairness:
This metric checks whether an individual’s outcome would remain unchanged in a counterfactual
scenario where their sensitive attribute (like race or gender) is different. It relies on causal
modeling to isolate the effect of the sensitive attribute on the outcome.

Tools and Practical Applications

Frameworks:
Toolkits such as IBM’s AI Fairness 360 provide implementations of many of these metrics, along
with methods for mitigating bias.
Trade-Offs:
Often, optimizing for one fairness metric may lead to compromises in another or in overall
accuracy. It’s important to select metrics that align with the ethical and practical goals of the
application.

Understanding and selecting the right fairness metrics is crucial for auditing AI systems and ensuring
they operate ethically and equitably.
en.wikipedia.org

Statistical parity

Thought for a second

Statistical parity (also known as demographic parity) is a fairness metric that requires the probability of a
positive outcome to be equal across all groups defined by a sensitive attribute. In mathematical terms,
for a binary classifier, it means that:

P (R = 1∣A = a) = P (R = 1∣A = b) ∀a, b ∈ A

where R is the model’s predicted outcome (with "1" representing a positive decision), and A is the
sensitive attribute (such as gender or race).

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 6/12
If the probabilities differ, the difference—called the Statistical Parity Difference (SPD)—indicates the
extent of imbalance between groups. An SPD of zero means the model treats all groups equally in terms
of the positive outcome rate. However, while this metric is easy to compute and interpret, it doesn’t
account for underlying differences in the true outcome distributions (i.e., base rates) among groups,
which may be justified in certain contexts.
fairlearn.org dataplatform.cloud.ibm.com

Equal opportunity

Reasoned about equal opportunity fairness for 8 seconds

Equal opportunity is a fairness metric that focuses on ensuring that among the individuals who truly
qualify (i.e., those with a positive true label Y = 1), the probability of receiving a positive prediction is
the same across all groups. In other words, a classifier satisfies equal opportunity if:

P (R = 1 ∣ Y = 1, A = a) = P (R = 1 ∣ Y = 1, A = b) for all groups a, b ∈ A.

This metric is often used in scenarios like hiring or loan approvals, where it’s crucial that all qualified
candidates have an equal chance of being selected, regardless of their group membership (such as
gender or race). It is essentially a relaxation of equalized odds—it only requires equal true positive rates
across groups without necessarily equalizing false positive rates.
By focusing solely on the true positive rate, equal opportunity aims to mitigate the risk that a qualified
individual is unfairly denied a favorable outcome due to biases in the model.
en.wikipedia.org superwise.ai

Disparate Impact Mitigation Strategies - (Pre-Processing, In-Processing, Post-Processing)

Reasoned for 12 seconds

Disparate impact mitigation strategies aim to reduce the unequal outcomes produced by a model across
different demographic groups. These strategies generally fall into three categories:

1. Pre-Processing

Pre-processing methods modify the training data before model training. The goal is to reduce bias in the
dataset so that the model does not learn unfair associations. Common techniques include:

Reweighing: Assign weights to data points so that underrepresented (or disadvantaged) groups
contribute more equally during training.
Data Transformation/Editing: Remove or alter features that are strongly correlated with the
sensitive attribute, or generate synthetic data to balance the distribution across groups.
Disparate Impact Removal: Techniques that adjust the data distribution to eliminate differences
in selection rates between groups.

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 7/12
2. In-Processing

In-processing strategies integrate fairness objectives directly into the model training process. This can
be achieved by modifying the learning algorithm to balance predictive accuracy with fairness
constraints. Techniques include:

Fairness-Constrained Optimization: Incorporate fairness constraints (like equalized odds or

equal opportunity) into the loss function during training.

Adversarial Debiasing: Train a model to both predict the target variable and to be “blind” to the
sensitive attribute by introducing an adversary that tries to predict the sensitive attribute from the
model’s outputs, while the predictor is trained to minimize both its prediction error and the
adversary’s accuracy.

3. Post-Processing

Post-processing methods adjust the model’s outputs after training to achieve fairer outcomes. This is
done without altering the underlying model and can include:

Threshold Adjustment: Modify decision thresholds for different groups so that the rate of positive
outcomes is balanced across groups.
Reject Option Classification: For uncertain cases (those near the decision boundary), assign the
favorable outcome to the unprivileged group to counteract bias.
Calibrated Equalization: Alter the final predictions to satisfy specific fairness metrics, ensuring
that error rates (such as false positives or false negatives) are balanced between groups.

Each category offers trade-offs. Pre-processing focuses on data quality and representation, in-
processing embeds fairness into the model’s training objective, and post-processing adjusts outputs to
align with fairness goals. Organizations often choose a combination of these approaches to mitigate
disparate impact while maintaining overall predictive performance.
dataplatform.cloud.ibm.com superwise.ai

Ethical AI Design Principles (Centered design, ethical by design)

Reasoned about ethical AI design principles for 15 seconds

Ethical AI design principles are guidelines that help ensure AI systems are built in a way that prioritizes
human values and minimizes harm. Often described as “ethical by design” or “centered design,” these
principles emphasize the need to incorporate ethical considerations throughout the entire development
lifecycle, rather than as an afterthought. Key principles include:

Human-Centered Design: The system should be designed around human needs, ensuring that
technology complements human decision-making and safeguards human dignity.

Transparency and Explainability: Developers should ensure that AI systems are interpretable so
that decisions can be understood, audited, and trusted.

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 8/12
 Fairness and Non-Discrimination: The design must actively prevent biases by ensuring equitable
outcomes across different demographic groups.

Privacy and Data Protection: Systems should safeguard personal data and comply with privacy
standards.
Accountability: Clear lines of responsibility must be established so that actions and decisions
made by AI can be traced back to human oversight.

Robustness and Safety: AI should be resilient to errors and adversarial attacks, ensuring reliable
operation in diverse environments.

Inclusive and Participatory Development: Engage diverse stakeholders and incorporate

interdisciplinary insights to capture a wide range of ethical and cultural perspectives.

These principles serve as a blueprint for developing AI that not only performs well but also contributes
positively to society.
superwise.ai dataplatform.cloud.ibm.com

Deployment challenges

Reasoned about AI deployment challenges for 11 seconds

Deployment of ethical AI systems faces a range of challenges that go beyond model development. Some
key challenges include:

Integration and Scalability:

Transitioning from a controlled development environment to a production system often means
integrating with legacy systems and scaling up while maintaining performance and fairness.

Data Drift and Continuous Monitoring:

Once deployed, models encounter real-world data that can change over time (data drift).
Continuous monitoring is essential to detect and mitigate emerging biases or performance
degradation.
Maintaining Fairness and Transparency in Production:
Ethical safeguards implemented during development (e.g., fairness constraints) may weaken over
time. Ensuring the model’s decisions remain explainable and fair requires robust auditing and
transparency mechanisms.

Regulatory Compliance and Security:

Deployed AI systems must comply with legal and regulatory standards, including data protection
and non-discrimination laws, while ensuring cybersecurity and privacy.
Operational Complexity:
Managing resource allocation, ensuring system reliability under varying loads, and providing
timely feedback for model updates are additional operational hurdles.

Addressing these challenges requires a combination of technical strategies (such as continuous

integration pipelines for fairness monitoring), robust governance frameworks, and collaboration

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 9/12
between technical teams and stakeholders to ensure that ethical principles are sustained throughout
the model’s lifecycle.
dataplatform.cloud.ibm.com superwise.ai

Real world implementation

Reasoned about real-world AI implementation challenges for 4 seconds

Real world implementation of ethical AI involves integrating ethical design principles and fairness
metrics into operational systems while addressing practical challenges like data drift, regulatory
compliance, and stakeholder engagement. In practice, this means:

• Integrating Ethical Frameworks:

Organizations embed ethical guidelines—such as human-centered design, transparency, fairness, and
accountability—into their AI development pipelines. For instance, companies like IBM, Microsoft, and
Google have developed toolkits (e.g., AI Fairness 360, Fairlearn, Model Card Toolkit) that guide teams
through ethical audits and bias mitigation during development and after deployment.
• Continuous Monitoring and Feedback:
Once deployed, AI systems face real-world variability. Firms use continuous monitoring systems to
detect shifts in data (drift), emerging biases, or performance drops. Automated fairness assessments
and periodic audits ensure the system continues to meet ethical standards even as the underlying data
or environment changes.

• Stakeholder Engagement and Regulatory Alignment:

Implementing ethical AI in the real world also requires engaging a diverse range of stakeholders—
including affected communities, regulatory bodies, and interdisciplinary experts—to shape guidelines
that are culturally sensitive and legally compliant. Some organizations have even established ethics
boards or advisory councils to oversee these processes.

• Case Studies and Pilot Programs:

Examples from sectors like finance, healthcare, and criminal justice show how pilot programs and case
studies are used to refine ethical AI approaches before wider rollouts. For example, banks are
experimenting with fairness-aware credit-scoring models that adjust for historical biases in loan
approvals, while healthcare providers deploy diagnostic tools with built-in safeguards against bias.

These real-world implementations underscore that ethical AI isn’t just about technical fixes—it’s about
creating a robust ecosystem of policies, continuous evaluation, and multi-stakeholder collaboration.
dataplatform.cloud.ibm.com superwise.ai

Monitoring and Feedback Loops

Reasoned about monitoring and feedback loops for 12 seconds

Monitoring and feedback loops are critical for maintaining the ethical and fair operation of AI systems
once they are deployed. They enable continuous assessment and timely intervention to address
performance degradation, data drift, and emerging biases. Key aspects include:
Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 10/12
 Continuous Monitoring:
Implement systems that track not only standard performance metrics (accuracy, precision, recall)
but also fairness metrics (e.g., statistical parity difference, equal opportunity) over time. This
ongoing surveillance helps detect when the model’s behavior starts to diverge from its intended
ethical standards.

Feedback Loops:
Establish processes for collecting feedback from end users, domain experts, and other
stakeholders. This feedback is integrated into the model update cycle so that any identified issues
(such as new forms of bias) can trigger a retraining or recalibration of the system.

Automated Alerts and Retraining Pipelines:

Use automated systems that generate alerts when key metrics exceed pre-defined thresholds.
These alerts can prompt investigations and initiate retraining pipelines to adjust the model based
on recent data, ensuring it remains robust and fair.
Governance and Documentation:
Create clear governance frameworks to document monitoring practices, update logs, and feedback
responses. This transparency supports accountability and helps organizations meet regulatory and
ethical standards.

Overall, effective monitoring and feedback loops are essential for adapting to real-world changes and
ensuring that ethical AI systems continue to operate as intended throughout their lifecycle.
dataplatform.cloud.ibm.com superwise.ai

Responsibe AI Framework Guidelines and best practises for ethical deployment

Reasoned about responsible AI deployment guidelines for 4 seconds

Responsible AI framework guidelines and best practices focus on embedding ethical considerations
throughout the AI system's lifecycle—from development through deployment and beyond. Key
guidelines and practices include:

Ethical Design and Human-Centered Principles:

Design systems with human values in mind. This means incorporating user feedback, ensuring that
systems complement human decision-making, and addressing potential harm proactively.
Transparency and Explainability:
Ensure that AI models are interpretable and their decision-making processes are well-documented.
Tools like model cards, explainability frameworks (e.g., LIME or SHAP), and thorough
documentation help stakeholders understand how decisions are made.

Fairness and Non-Discrimination:

Implement fairness metrics (such as statistical parity, equal opportunity, and equalized odds)
during development and monitor them continuously in production. Mitigation strategies—pre-, in-,
and post-processing techniques—should be used to address biases in data and algorithms.

Privacy and Data Protection:

Adhere to data protection standards (such as GDPR or CCPA), using privacy-preserving techniques

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 11/12
 (e.g., differential privacy or anonymization) to safeguard personal data throughout the model
lifecycle.
Robustness, Safety, and Security:
Build models that are resilient to adversarial attacks, data drift, and unexpected changes in the
environment. Regular testing, validation, and updates are crucial for long-term reliability.
Accountability and Governance:
Establish clear lines of accountability with governance frameworks that define roles and
responsibilities. This includes regular audits, oversight committees, and stakeholder involvement
to ensure ethical performance.
Stakeholder Engagement and Inclusivity:
Involve diverse stakeholders (including marginalized groups) during the design and deployment
phases. This helps ensure that the system addresses a wide range of needs and mitigates
unintentional harms.
Continuous Monitoring and Feedback Loops:
Deploy monitoring systems that track both performance and fairness metrics. Set up automated
alerts and feedback loops to trigger retraining or model adjustments when issues are detected.

These best practices are reflected in frameworks and guidelines provided by industry leaders and
regulatory bodies (e.g., IBM’s AI Fairness 360, Microsoft’s Fairlearn, Google’s Model Card Toolkit, and the
EU’s AI Act). By following these guidelines, organizations can deploy AI ethically and responsibly,
ensuring that systems are fair, transparent, and aligned with societal values.
superwise.ai dataplatform.cloud.ibm.com

Printed using ChatGPT to PDF, powered by PDFCrowd HTML to PDF API. 12/12