AI in Business and Society Midterm 2

Ai Fairness

Impartial or just treatment or behavior without favoritism or discrimination

Fair v Equal

Far does not mean equal

Fairness means that everybody in the group has an equal opportunity to benefit

Equality v Equity

Equality

• Everybody is treated the exact same

Equity

• Everybody is given the equal access to be the same

Disparate Treatment

Liability can be imposed if there is an explicit classification based on the protected attribute or if there was an intent to discriminate

Disparate Impact

Liability can be imposed even if the policy is neutral on its face, but there is a disproportionally adverse impact on minority groups

In order to build a fair ML model you must:

• define the definition a fair outcome

• who selects and defines what is a fair outcome and for whom

• what are the metrics we use

Individual Fairness

Advocates for treating similar individuals similarly

Group Fairness

Advocates to acheive the same outcomes across different demographics

Hybrid Fairness Approaches

Learning Fair Representations

• combines group and individual fairness

• examples such as learning fair representations or individual risk scores

Unaware Fairness Approach

Algorithm is blinded or unaware of any identifiable factors and prohibited attributes

Have to consider proxies for these factors that may be used instead

Can only work in cases where inequality is not an issue

Fairness through Awareness Approach

Two individuals should be treated similarly

It is entirely based on how you define similarity between applicants

Equality of False Negatives

Enforces a constant false negative rate across groups

Does not enforce any equity or false positive errors

Equal Opportunity

Individuals who qualify for the outcome should have an equal chance to be classified for the outcome

The true positive rate has to be equal across groups

Equalized Odds

Requires the same true positive and false positive rates across groups

When to use it: Strong emphasis on correctly predicting positive outcome, and strongly care about minimizing costly false positives

Accuracy v Fairness

As we try and make the model fairer, we lose accuracy of the model

Group v Individual

Increasing the between group fairness can increase the within group unfairness

Fairness Chart to remember

AI Bias

AI Bias occurs when an algorithm produces results that are systematically prejudiced against specific groups

This is important because algorithms are often used to distribute goods and services, as well as have significant impact on opportunities, health, and quality of life

Sources of AI Bias

Problems with Training Data

• Discrimination in real world is reflected in the training data

• under sampling or selection bias, causing non represented population

• Proxy Biases that stand in place for discriminatory data

Problems with testing

• One size fit all model that is used incorrectly

• Benchmark or testing bias, can give values that are biased

• Objective function bias

  • goals may unintentionally lead to disparate impacts among subpops

Identifying AI Bias

Rely on definitions of fairness and related computational measures of bias

Issues with Measuring AI Bias

Metrics of Fairness

Timing of Identifying Bias

Allocative and Representational Harms

Legal Issues

Choosing Risk Mitigation Factors

Metrics of Fairness (AI Bias)

Fairness metrics are not compatible with each other

• When you focus on one metric, you reduce the other metric

• Is each metric fair for each group?

Timing of Identifying Bias (AI Bias)

Bias risk mitigation should be considered before the model is trained

• If this is done after training, the discrimination will bleed into our interpretation of the model

Choose fairness metrics before training the model to ensure safeguards are in place

Allocative and Representational Harm (AI Bias)

Computational AI fairness does not cover all types of bias

Representational Harms

• Algorithmically filtered depictions are discriminatory

Allocative Harms

• Unfairly assigned opportunities or resources by the algorithm

Computational measures only capture allocative, humans have to identify representational

Representational Harms

Occur when systems reinforce the subordination of some groups along the line of identity

• ex) white men showing up when searching for CEO

Legal Issues (AI Bias)

Computational measures for AI bias are not compatible with existing discrimination law

• the law forbids using race and gender in decision making processes

The model may need to use these factors to ensure the model is not being bias

Choosing Risk Mitigation Strategies (AI Bias)

Bias may arise because of insufficient training data

Identifying risk mitigation strategy is a complex matter involving lots of factors

Ensure individuals with relevant experience are making these decisions

AI Bias Mitigation Strategies

Input Data

• get more data, get better proxies, examine if a one size fits all model actually works, alter weights of inputs, use bias corrected synthetic data

Decisions and Output

• Use binary choices with a threshold

Identifying Bias

• Synthetic data?

AI Explainability

The ability to understand and explain the outputs of an ML model in human terms

Explainable AI Foundations

Explain to Justify

Explain to Control

Explain to Discover

Explain to Improve

Explain to Justify

• To justify outcome

• Make sure they are not erroneous

• Make sure they are fair and ethical

• To comply with legislation

Explain to Control

• To prevent from going wrong via visibility

• Detect and correct errors

Explain to Discover

• To help humans learn

• Discover new knowledge

Explain to Improve

• To continuously improve the model

• Understanding helps us know how to improve

Who Demands an AI Explanation?

• those impacted by decisions

• employees who must trust the decisions

• regulation reasons

Why is Explainability Important?

Enables monitoring and accountability during both production and deployment

Ensures fairness and debiasing to mitigate against model drift

Lack of explanation can lead to ethical, regulatory, or legal risk

People Explanations

Why people make the decisions they did

Machine Explanations

How the model arrived at its output given the inputs

• Includes Global and Local Explanations

Global Explanations

Understanding the logic of the model as a whole

What are the main features that influence the predictions and how do these factors interact

Useful to data engineers, managers, regulators

Local Explanations

Understanding the reasons behind a specific prediction

Why did the model come to this conclusion

Useful to decision makers, end users, regulators

Machine v People Explanations

Machine explanations are complex

• A large number of variables at play, number of relationships, nature of relationships

Not understandable in the same way human explanations are

What does asking for Explainable AI really mean?

Includes: People and Machine explanations (G or L or Both)

Why not always provide machine explanations?

• Requires resources to develop

• tradeoff between accuracy and explainability

Intrinsic (Direct) Interpretability

Models are interpretable due to their simple structure

There is a tradeoff between accuracy and intrinsic interpretability

Post Hoc Interpretability

Interpreting a black box model by applying interpretability methods after training the model or while in production

Model Specific Explanation Methods

Limited to specific model or groups of models

They depend heavily on the working and capabilities of a specific model

Model-agnostic Explanation Models

Works on any model

Do not look into the black box of the algorithm but work with the input to output pairs

Permutation Models

How much does the model rely on each feature (This produces a ranking of features)

Changing input data and measuring changes in model predictions

Surrogate Models

What simple rule approximates what the model is doing (It produces an interpretable model that mimics the black box

Model is first trained to make predictions, then the surrogate model is trained using the original models’ predictions

Surrogate models are an inherently interpretable model

Importance of Machine Explanations

Shows respect, informs how to change the outcome, helps determine if rules are fair, helps monitor if the algorithm is functioning as intended, need to balance importance of explanations against accuracy and cost/resource

When Machine Explainability does not Matter

When the model does not directly deal with decisions about how anyone should be treated

When people explanations of why you want to use black box plus informed consent justifies use

When Machine Explainability does Matter

• When expressing respect is ethically required

• When people need to know how to get better results

• When people need to know how to approach and make a decision

• When outputs are strange

• When you need to justify treatment

What Makes for Good Explanations

Objective

• Demonstrate Respect, allow user to make well informed decisions, allow people to review decisions

Characteristics

• Truth

• Ease, efficiency, and effectiveness of use

• Intelligibility

Guidelines: Consult end users of software to determine whether explanation is needed and if so, what a good explanation looks like

Explainability to Build Trust

Explanations help users build the appropriate level of trust in the AI algorithm

Helps users understand when to trust the system and when to use their own judgement

How to Calibrate Trust through Explanations

Explain the system and the output

Articulate data sources

Explain what is important

Account for situational stakes

Decide if and how to show model confidence to manage AI influence

Onboarding (Calibrating Trust)

Global explanations as part of onboarding with the algorithm calibrate expectations

• What data does the algorithm use, strengths and weaknesses, what was it trained on

Data Sources (Calibrating Trust)

Feature based Explanations

• Global and Local

Example based explanations

• appropriate when it is challenging to explain the reasons behind AI predictions

Model Confidence Displays (Calibrating Trust)

Show how certain the AI is in its recommendation

First Step: Determine whether or not to show confidence levels

• It can help users calibrate trust, but is often not intuitive

• Can lead to misinterpretation or confusing data

What is Privacy?

The right to control who can collect what data about you and what they can use it for

3 Aspects of Privacy

Regulatory Compliance

Cybersecurity

Ethics

Privacy Act of 1974

Centers around consumer choice or ‘consent’

Issues with Consent:

• Impossible to enforce in many AI applications (smart driving cars)

• Onerous to properly inform oneself about privacy policies of organizations

• People not only care about how their information will be used, but also about their data being collected

• Desire to change this model by shifting the burden of protecting individual privacy from consumers to businesses

Unique AI Privacy Concerns

Data Persistence

• Data existing longer than the human subjects created it, driven by low storage costs

• Privacy preferences change over time, aggregation across sources

Data Repurposing

• Data being used beyond their originally imagined purpose

• Unanticipated correlations between features, proxies, algorithms may identify people through data

Data Spillovers

• Data collected on people who are not the target of data collection

• Being in pictures taken in public, using a person to gather data on relatives

AI Privacy Issues

The danger is not the data that the AI organizations possess, its about what they DO with the data

4 Elements of Ethical Privacy

Transparency

Data Control

Opt In or Out by Default

Full Services

Transparency

• If you are not aware of what/how the data is being used then you cannot exercise control

Data Control

• Having the choice to control the information about oneself

• Important for these reasons: may be regulated, communicates respect, gives users the ability to correct information

Opt In or Out by Default

Do users automatically opt in for collection of their information?

• Opt in puts the burden on the user to investigate how to obtain the data the org has about them

• Opt out puts the burden on the organization to demonstrate to the user that will benefit from sharing their data

Full Services

Orgs may need to adjust the level of services it provides based on amount and type of data a person shares

If the service is essential and is only provided on the condition the user provides full access to the data, and is uncomfortable, the company is exhibiting undue influence

5 Ethical Levels of Privacy

Blindfolded and Handcuffed

Handcuffed

Pressured

Slightly Curtailed

Grateful

Blindfolded and Handcuffed

People are in the dark and passive with regard to their data

Handcuffed

People are knowledgeable about their data, but they are passive with regards to what is collected and done with it

Information is readily available to users

Organization put lots of effort into making this possible

Pressured

People are knowledgeable and they have some degree of control over what data is collected

Able to opt out of certain uses of their data, but some has to be collected

Slightly Curtailed

People are knowledgeable and their data has not been collected and used without their consent

Grateful

Organization provides full services independent of what data the person opts in to provide and consent to be used

Privacy Calculus

Privacy Risk and Benefits have an inverse relationship

Privacy Paradox

The relationship between a person’s intentions to protect their data v how they actually behave online