How can you make your AI system fair?

You need to engage with diverse stakeholders, such as customers, employees, partners, regulators, experts, or activists, throughout the design and deployment of your AI system. You should also solicit and incorporate their perspectives, opinions, concerns, or suggestions into your AI system.

The rise of AI brings with it the pressing concern of racial, gender, and other biases. These disparities aren't mere technological glitches but mirror societal prejudices, sometimes amplifying them. For instance, AI models might perpetuate hiring biases, intensifying workplace disparities. Recognizing these biases is the first step towards rectification. Developers need clear fairness goals and diverse data sets. Public discourse can drive demand for ethical AI. Addressing AI biases isn't just a tech imperative but a moral one, ensuring all benefit equitably from AI's advancements.

In the ever-evolving world of AI, it is essential to set the right course from the outset. Defining fairness goals serves as the North Star in developing AI systems that are ethical and equitable. Fairness, as we've learned, is multifaceted and context dependent. Organizations must work closely with architects to identify and prioritize relevant fairness criteria. These may include equality, equity, diversity, inclusion, accountability, transparency, or explainability, among others. This foundational step ensures that AI systems align with ethical principles from the very beginning. Note: Selecting a fairness definition is in itself problematic because fairness is neither binary nor absolute. What fits in one situation will not in another.

Ensuring fairness in an AI system involves several key steps. Begin by designing unbiased algorithms, using diverse and representative training data. Regularly evaluate the system for biases, both during development and in real-world usage. Implement fairness-aware metrics and auditing tools to detect and rectify any disparities. Additionally, engage diverse teams and stakeholders to provide input and perspectives. Continual monitoring and transparency in decision-making processes are essential to promoting fairness and mitigating potential biases in AI systems.

I think we can establish that Fairness is a complex concept. I really appreciate the guidelines put forth by the EU, that are much more outcome focused: - fairness should ensure equal and hust distribution of benefits and costs - fairness should ensure that systems are free for intended, unintended or implicit biases or discrimination.

There are various methods we can use to check if our data is of good quality: Descriptive Statistics: This is a simple method where we use basic statistics like the mean (average), median (middle value), mode (most common value), standard deviation (how spread out the values are), and range (difference between the highest and lowest values). This helps us spot anything unusual or incorrect in our data. Data Profiling: Here, we take a closer look at our data - its structure, what's in it, and how different parts relate to each other. This way, we can find any mistakes, errors, or missing pieces in our data.

Removing sensitive attributes from datasets to prevent unfair outcomes is insufficient. Information about these attributes can still be inferred from other features, especially where high correlation exists. This connection can be quantified via metrics like Shannon Mutual Information or Correlation Coefficient, or others that indicate dependency between sensitive attributes and different features. Additionally, documenting these findings is crucial. The tool "Datasheet for Datasets", proposed by Gebru et al. (2021), is an effective way to communicate limitations, collection procedures, processing methods, and pitfalls associated with a specific dataset. This approach is useful for providing deeper context alongside dataset delivery.

In the realm of AI, data is both the lifeblood and a potential source of bias. Architects must lead the charge in assessing data quality and representation. Thorough examination is imperative to ensure that data accurately mirrors the diversity of the real world. Identifying data biases, inaccuracies, or contamination is crucial. Additionally, architects should establish transparent documentation practices to disclose the sources, methods, and limitations of data collection and processing. This transparency fosters accountability and trust among users and stakeholders.

The first step towards fairness is to select and prepare the data that we use to train AI systems. We have ensure the data reflects the diversity and reality of the people and situations that our AI systems will encounter and that it does not include features or labels that may unfairly favor or harm some groups over others. We have to also keep track of and check the quality and distribution of the data throughout the development and deployment of AI systems to make any adjustments to the data as necessary to keep up with constant changes. We should also evaluate the algorithm's performance and behavior using multiple fairness metrics and methods, such as confusion matrices, ROC curves, fairness-aware learning and counterfactual analysis.

While libraries like Fairlearn, AIF360, and SageMaker Clarify offer ways to mitigate bias, it is still possible for disparities to remain even after implementing these tools. Indeed, recent research by Kaneko et al (2022) demonstrated that bias can be reintroduced during subsequent fine-tuning procedures, which are commonly used to enhance the performance of language models like GPT3.5, GPT4, PaLM, and Llama. This highlights the need for responsible and ethical automated algorithms that go beyond bias mitigation. Stakeholders, including dataset producers, ML developers, and users, should also have access to complementary tools that help them interpret potential disparity risks throughout the entire pipeline.

Addressing algorithmic bias is a top priority in ensuring fairness. Architects play a critical role in this process. It's essential to select fairness-aware algorithms, establish fairness constraints, and integrate bias-detection tools into the AI pipeline. This collaborative effort with data scientists and machine learning engineers aims to ensure that AI systems do not produce biased outcomes that harm specific groups or individuals. Note: Selecting a fairness definition is in itself problematic because fairness is not binary nor absolute. What fits in one situation will not in another. De-biasing methods are better applied toward identifying root causes than attempting to surgically remove bias, which can introduce unintended effects.

While the focus on mitigating algorithmic bias is a good start, true fairness in AI goes beyond technical adjustments. It demands a holistic approach that embeds ethics at the core of AI design and deployment. Engage diverse teams in AI development to capture a broader spectrum of perspectives. Regular audits, not just of system outcomes but also the data fueling these systems, are essential. Beyond tools and techniques, fostering an organizational culture that prioritizes fairness and inclusivity is crucial. Remember, fairness in AI isn't a one-off task; it's a continuous commitment to uphold the values of equality and justice.

In the context of digital transformation, it's crucial to architect systems that prioritize user consent and privacy. User trust and privacy are paramount in the digital age. Architects must design AI systems that respect user consent and privacy. This involves crafting robust identity and access management systems, implementing encryption protocols, and creating user-friendly interfaces for obtaining informed consent. Architects should ensure that users retain control over their data, including the ability to opt in or out of data sharing. This guardianship of privacy is a fundamental ethical principle.

Only collect and use data that is needed and relevant for your goals and tell your users how and why you do it and their rights. Give your users easy ways to agree or disagree to the use of their data - don't pressure or trick them. Respect their choices and let them change their minds anytime. Navigate within the legal framework and have plans to prevent and target data problems. Inform your users and authorities right away if something goes wrong. Let your users access, change, or delete their data if they ask, and give them ways to do it. For example, a dashboard or portal, or a contact form. You can also let them move their data to other places or services. Check and improve your governance policies and processes regularly.

Transparency and explainability are essential pillars of trustworthy AI. Architects should consider the technical infrastructure required to offer users insights into how AI systems make decisions. This may involve implementing model interpretability tools, constructing user-friendly dashboards for visualizing model outputs, and crafting clear documentation for users and stakeholders. Collaboration with UX/UI designers is vital to make transparency features accessible and understandable.

The question that is left to be answered is what explainability entails. When we talk about explainability for AI systems we need to make sure that we have defined, to whom the System should be explainable. A functional requirement of explainability for a human is to short sighted. I think explainability to a Subject Matter Expert is much more feasible.

We were building a customer churn prediction model. The initial models were simple, easy to understand, but delivered only 30% improvement over the base state. As we improved the model accuracy, we crossed over into black box algorithms. These were 30% more effective than simpler white box models, but with zero explainability. When we rolled these 'superior' models, the users resisted and we had to kill the project. The importance of explainability cannot be overstated. It might sound illogical to engineering teams, but that's the real world out there!

Ethical AI development goes beyond technology; it's about engaging with diverse stakeholders. Architects must facilitate the integration of stakeholder feedback into the AI system's design and development process. They should design solutions that encourage ongoing engagement, such as feedback mechanisms and open communication channels with stakeholders. This collaborative approach ensures that AI systems align with the values and expectations of a diverse user base.

A voice recognition system trained mostly on male voices might struggle to understand female users. It's essential to treat data as you would a board of advisors - the more diverse and high-quality the input, the more trustworthy the recommendation. Engage with more folks & open up to more voices.

We can't let algorithms repeat the same mistakes and injustices that humans do. We need to know what's inside the code we use and we need to make sure it's fair, clear and held accountable. We wouldn't trust people who make big decisions to act without any checks or balances and we shouldn't let algorithms do that either. Providers must be honest about how their algorithm work so users and authorities can avoid harm. Bias can cost a lot of value and opportunity. Whereas reducing bias can create value. If we want to use artificial intelligence to be more profitable and productive, we need to make sure those systems are more fair and inclusive. That's good for business and good for society.

Let's illustrate with examples 1. Define Objectives, for example, a hiring AI system to remain fair to different demographic groups 2. Representative Data, e.g. for a loan approval AI model, data to represent diversity of genders, incomes etc 3. Bias Detection / Mitigation, e.g., a facial recognition system corrected if showing higher error for darker-skinned individuals 4. Transparency, e.g. a credit score system explaining what factors influencing score outcomes 5. Regular Auditing 6. Diverse Development Teams e.g. a chat bot for customer service reflecting diversity 7. User Feedback e.g. social media platform allowing users to flag content 8. Ethical/Regulatory , e.g. healthcare system complying with HIPAA 9. Bias Impact Assessments