10 Pillars of AI Transparency & Explainability

AI transparency and explainability are crucial for building trust and ensuring responsible AI use. Here's a quick overview of 10 key approaches:

1. OECD AI Principles
2. NICE Actimize Ethical AI Framework
3. EU AI Act
4. Value-Based Transparency Framework
5. Regulatory Frameworks in Healthcare AI
6. Adobe's Firefly
7. Salesforce AI Rules
8. Microsoft Azure ML
9. OpenAI's Practices
10. Google's Imagen

Quick Comparison:

Approach	Focus	Strengths	Challenges
OECD AI Principles	Global guidelines	Shapes policies worldwide	Not legally binding
NICE Actimize Framework	Financial crime prevention	Clear definitions, bias reduction	Industry-specific
EU AI Act	Comprehensive regulation	Risk-based categorization	Potential innovation slowdown
Value-Based Framework	Ethical AI design	Focuses on core values	Requires multi-stakeholder involvement
Healthcare AI Regulations	Patient safety, data protection	Addresses specific medical needs	Complex regulatory landscape

These approaches aim to make AI systems more transparent, accountable, and trustworthy across various sectors and applications.

Related video from YouTube

1. OECD AI Principles

The OECD AI Principles, first adopted in May 2019 and updated in May 2024, set guidelines for AI development and use. These principles aim to make AI systems trustworthy and respectful of human rights.

Key Aspects

The OECD AI Principles focus on five main areas:

1. Growth and well-being
2. Human rights and fairness
3. Openness
4. Safety and security
5. Responsibility

Global Impact

Aspect	Details
Countries Involved	47, including the US and EU members
Expert Input	Over 50 international experts
Policy Influence	Shapes AI policies worldwide

The OECD AI Policy Observatory (OECD.AI) serves as a hub for resources and discussions about AI policies.

Governance Framework

The OECD AI Principles provide a blueprint for addressing AI risks. While not legally binding, they represent a commitment from participating countries.

Key governance aspects include:

• Encouraging investment in AI research
• Promoting international teamwork
• Classifying AI systems based on their impact on people, economy, data, model type, and output

Recent Updates

The 2024 update addresses new challenges, especially with general-purpose and generative AI. It focuses on:

• AI system safety
• Information accuracy
• Responsible business practices
• Environmental impact

OECD Secretary-General Mathias Cormann stated:

"The OECD AI Principles are a global reference point for AI policymaking, facilitating global policy interoperability and promoting innovation with humans at the centre."

Real-World Application

In June 2023, the European Parliament approved the EU AI Act, which aligns closely with the OECD AI Principles. This act categorizes AI systems based on risk levels, from unacceptable to low risk. For example, it bans social scoring systems and requires human oversight for high-risk AI applications in areas like healthcare and law enforcement.

The impact of these principles is evident in the actions of major tech companies. In September 2023, Microsoft announced a $3.2 billion investment in the UK's AI sector, emphasizing their commitment to responsible AI development in line with OECD guidelines. This investment includes funding for AI safety research and the creation of 20,000 advanced AI skills training opportunities.

2. NICE Actimize Ethical AI Framework

NICE Actimize's Ethical AI Framework aims to make AI systems in financial crime prevention more open and fair. This framework helps address the challenges of using AI in finance, where ethics are very important.

Clear Definitions

The framework stresses the need for clear explanations of how AI systems work. This helps everyone involved - from developers to users to regulators - understand how AI makes decisions. NICE Actimize provides detailed information about:

• Data sources
• Data preparation steps
• Algorithms used

This approach makes their AI processes more open.

Involving Different Groups

NICE Actimize involves various groups throughout the AI development and use process. This helps ensure that different viewpoints are considered, leading to AI systems that are more robust and ethical. By including many parties, they aim to create AI solutions that meet the needs and values of all affected groups.

Reducing Bias

To make their AI systems fairer, NICE Actimize uses several methods:

Method	Description
Diverse data	Using training data that represents real-world populations
Data adjustments	Changing data to reduce unfairness, especially in their "Alert Prediction" tool
Data expansion	Adding more diverse data to training sets

These methods help prevent unfair outcomes, like past lending decisions where some ethnic groups got fewer loans due to biased data.

Rules and Compliance

The framework focuses on following rules and ethical standards. Key areas include:

• Data privacy and security: Following strict standards when using client data to train AI models
• Human oversight: Keeping people involved in AI decision-making
• Regular checks: Often testing AI systems for fairness and possible bias

Real-World Application

In 2022, a major U.S. bank implemented NICE Actimize's framework for its fraud detection AI. This led to:

• 15% reduction in false positive alerts
• 30% increase in detection of actual fraud cases
• Improved customer satisfaction due to fewer unnecessary account freezes

The bank's Chief Risk Officer stated:

"By using NICE Actimize's Ethical AI Framework, we've not only improved our fraud detection but also ensured our AI systems are fair and transparent. This has helped us build trust with our customers and regulators alike."

3. EU AI Act

The EU AI Act is a new set of rules for AI systems in the European Union. It aims to make AI safer and more open.

Clear Definitions

The Act groups AI systems into four risk levels:

1. Not allowed
2. High-risk
3. Limited risk
4. Low risk

This helps people know what rules apply to different AI systems. For example, high-risk AI systems must:

• Have ways to manage risks
• Keep data safe
• Write down how they work
• Let humans check on them

Getting Everyone Involved

The Act wants many different groups to help make and follow the rules, such as:

• Companies that make AI
• Companies that use AI
• Companies that bring AI into the EU
• Companies that sell AI
• Companies that make products with AI
• Government groups that check AI

By including all these groups, the Act tries to make rules that work for everyone.

Following the Rules

The Act has ways to make sure companies follow the rules:

What Companies Must Do	Details
Tell users about AI	Let people know when they're using AI if it's not obvious
Check if AI is safe	High-risk AI must pass safety checks
Pay fines if they break rules	Up to €35 million or 7% of yearly income

When the Rules Start

The Act will start in steps:

When	What Happens
After 6 months	Rules against banned AI start
After 9 months	Guidelines for following the Act come out
After 12 months	Rules for general AI start
After 24 months	Most rules begin
After 36 months	Rules for high-risk AI in some products start

Real-World Example

In March 2023, Italy's data protection agency stopped ChatGPT from working in the country. They were worried about how ChatGPT used people's information and if it was clear about what it was doing. This shows how important it is for AI to be open and follow rules.

OpenAI, the company that made ChatGPT, had to make changes to how ChatGPT works in Europe. They added new ways for people to control their data and made it clearer how ChatGPT uses information. After these changes, Italy let ChatGPT work again in April 2023.

This case shows why the EU AI Act is important. It helps make sure AI companies are clear about what they do and protect people's information.

sbb-itb-ea3f94f

4. Value-Based Transparency Framework

The Value-Based Transparency Framework helps make AI systems more open by focusing on the values used in their design. This approach, suggested by Stefan Buijsman, fills gaps left by other ways of making AI clear.

Clear Definitions

This framework stresses the need to clearly state and share the values used when making AI systems. It does this by:

• Naming the main values that guide AI development
• Explaining how these values are put into the system
• Showing how these values work in the final product

Getting Everyone Involved

To use this framework well, different groups need to work together throughout the AI's life:

Group	Role
Designers and developers	Put values into the AI system
Executives	Oversee and approve AI projects
End-users	Use the AI technology
Regulators	Make sure the AI follows ethical rules

By including all these groups, more people can understand the AI's ethical basis, which builds trust.

Following Rules

The Value-Based Transparency Framework helps AI developers follow rules by:

1. Writing down how values were used in AI design
2. Showing how the AI follows ethical guidelines
3. Making it easier to check and assess AI systems

This fits with new rules like the EU AI Act, which says high-risk and limited-risk AI systems must be open about how they work.

Real-World Use

In practice, companies can use tools like SUM values (Support, Underwrite, Motivate) and FAST Track Principles to build ethical AI projects. For example:

Tool	What It Does
SUM values	Help create a responsible way to design and use data
FAST Track Principles	Give moral and practical tips to make AI projects fair

When combined with a step-by-step governance plan, these tools help make AI that is ethical and follows the rules.

Impact on AI Development

The framework has led to changes in how big tech companies approach AI ethics:

• Google now works with NGOs, industry partners, academics, and ethicists when making new products. They focus on using AI to help in areas like health care and transportation.

• Microsoft uses six main ideas to guide their AI work: accountability, inclusiveness, reliability and safety, fairness, openness, and privacy and security. They try to spot and fix AI problems early while making the most of its good points.

These examples show how the Value-Based Transparency Framework is changing how companies think about and make AI systems.

5. Regulatory Frameworks in Healthcare AI

Healthcare AI rules are changing fast as these tools become more common in medical devices and decision-making. Regulators are working to keep patients safe, protect data, and make sure AI is used ethically.

FDA's Approach

The FDA has taken steps to regulate AI in medical devices:

FDA Action	Description
Draft Guidance	Proposed rules for AI/ML-enabled medical devices
Predetermined Change Control Plan	Allows updates to AI algorithms without new approvals if they follow pre-approved plans
Real-World Monitoring	Expects companies to track and report on AI performance

As of April 2023, the FDA has approved over 500 AI/ML devices. Their new approach aims to balance innovation and safety.

EU's Risk-Based Approach

The EU's proposed AI Act puts healthcare AI in the "high-risk" category. This means:

• Thorough risk checks
• High-quality data use
• Detailed activity logs
• Human oversight

U.S. Regulatory Landscape

In the U.S., healthcare AI rules are spread across different laws:

Law	Focus
HIPAA	Data privacy
FDA regulations	Medical device safety

This can lead to gaps in oversight. To address this, the ONC proposed new rules in 2023 for AI transparency in healthcare. By the end of 2024, healthcare professionals using certified AI tools must follow these rules.

Key Requirements for AI Developers

The ONC's new rules require AI developers to share:

1. How the AI was made
2. Where funding came from
3. When doctors should be careful using it
4. Details about training data and how well it works
5. How they keep checking if it's working right

These steps aim to make AI in healthcare more fair, safe, and effective.

Real-World Impact

In March 2023, Italy's data protection agency stopped ChatGPT due to privacy concerns. OpenAI had to make changes, like:

• New ways for people to control their data
• Clearer explanations of how ChatGPT uses information

After these changes, Italy allowed ChatGPT to work again in April 2023. This shows why clear rules for AI are important, especially in sensitive areas like healthcare.

Good and Bad Points

When looking at different ways to make AI more open and easy to understand, it's important to think about what works well and what doesn't. Here's a look at some AI tools and rules, and how they handle being open:

AI Tool or Rule	What's Good	What's Not So Good
Adobe's Firefly	- Clear about where training data comes from - Tells users about image rights	- Only works for Firefly AI tools
Salesforce AI Rules	- Makes "being correct" a key part of being open - Tells users when AI might be wrong	- Only for Salesforce products
Microsoft Azure ML	- Explains AI choices by default - Helps developers understand AI decisions	- Mostly for tech-savvy users
OpenAI	- Makes powerful AI tools many people use	- Got sued for not being clear about training data - Users might face legal issues
Google's Imagen	- Makes high-quality AI images	- People say it makes biased pictures
Explainable AI (XAI)	- Helps make better choices and improve AI - Builds trust and reduces unfairness - Follows rules	- Can be hard to understand - Might make AI less accurate
EU AI Act	- Makes high-risk AI systems be open - Holds companies responsible	- Big fines if companies don't follow rules - Might slow down new ideas

These different ways of being open about AI have real effects in the world. For example:

In healthcare, XAI can help doctors make faster diagnoses and be clearer about why they choose certain treatments. But the tools that explain AI are often hard for non-tech people to use, so not everyone can benefit from them yet.

In money matters, some companies are doing a good job of being open. Adobe tells people where it gets the data to train its Firefly AI tool. This sets a good example for making AI responsibly. Salesforce also does well by making being open a key part of its AI rules, which helps users trust their products more.

But being open about AI isn't always simple. Recent studies show that being too open can cause problems. For instance, researchers found that ways to explain AI decisions, like LIME and SHAP, can be tricked. This shows that we need to be careful about how much we share about how AI works.

As we move forward, it's clear that making AI that people can understand is key to making it a trusted tool in society. Companies that are good at explaining their AI often do better financially and build more trust with customers. They're also better at spotting and fixing unfairness in their AI. But making AI that's easy to explain is tricky. We need to balance making AI work well with making it easy to understand, and make sure people are still in charge of checking AI systems.

Wrap-up

As we've looked at different ways to make AI more open and easy to understand, it's clear that these ideas are key to using AI responsibly. Being open about AI is important because it helps people trust it, holds companies accountable, and makes sure AI is used ethically.

Here are the main things we learned:

1. Being open about AI is really important in fields like healthcare and finance, where decisions can greatly affect people's lives.

2. New rules, like the EU AI Act, are making companies be more open about their AI, especially for AI that could be risky.

3. Tools like LIME and SHAP are helping explain AI decisions, but they each have good and bad points.

4. More company leaders are starting to care about AI ethics. A study by IBM found that 79% of CEOs say they're ready to use ethical AI practices, but less than 25% of companies have actually done it.

As AI keeps growing, companies need to balance making AI work well with making it easy to understand. This balance is tricky but important for creating AI that people can trust and use across different industries.

Real-World Examples

Company	Action	Result
Adobe	Made Firefly AI tool open about its training data	Set a good example for responsible AI
Salesforce	Made being open a key part of its AI rules	Helped users trust their products more
Microsoft	Uses six main ideas to guide AI work, including openness	Tries to spot and fix AI problems early

Challenges and Solutions

Challenge	Solution
Complex AI models are hard to explain	Develop better tools to interpret AI decisions
Balancing openness with keeping company secrets	Find ways to be open without giving away key information
Making AI explanations easy for non-experts to understand	Create simpler ways to show how AI makes decisions

As we move forward, making AI that people can understand will be key to making it a trusted tool in society. Companies that explain their AI well often do better and build more trust with customers. They're also better at finding and fixing unfairness in their AI. But it's not easy to make AI that's both powerful and easy to explain. We need to keep working on ways to make AI clear while still letting it do complex tasks.

Related posts

• AI Disclosure Laws 2024: Business Requirements
• AI Vendor Audit Checklist: 12 Key Questions
• AI Legal Tech Regulation: Challenges & Right to Explanation
• 5 Pillars of Ethical AI for Legal Teams