AI Ethics and Accountability: Transparency and Explainability in Artificial Intelligence

Name: AI Ethics and Accountability: Transparency and Explainability in Artificial Intelligence
Author: Briana Brownell

Learn about principles related to transparency and explainability of AI systems

Preview the first lesson free — get full access to all 5 lessons.

Request a demo Sign in

Course: On-Demand

(gentle music) When we think about transparency as a principle for building an AI system, most of us already have an idea of what it means. Transparency allows for oversight. That's because a transparent system can be examined by virtue of its transparency. So the principle of transparency suggests that oversight of an AI system should be possible. There are several ways in which an AI system should be transparent in its design and implementation. In terms of design, there should be full disclosure around any data used to train a machine learning model, including how the data was collected, pre-processing steps, model parameters, and the algorithm used to generate the model. Transparency also applies to AI systems that don't use machine learning to disclose how expertise or knowledge is used to inform the decision system design. It's important to note that concern about transparency centers on the system's actions as well as on its design. These actions must be possible to oversee. It's tempting to think about transparency as a yes or no attribute of the AI system. Either the system is transparent or not. But in fact, there are many gradations between perfect transparency and opacity. A system's decision structure may be transparent in terms of how the decision is made, but the data used to generate it might not be transparent because the data collection and processing steps taken might not have been properly disclosed. Consider an AI system built around a decision tree. Decision trees are a common supervised machine learning method that essentially generate a flow chart based on input training data. Once generated, a decision tree might be completely straightforward and easy to follow, but the database justification for why the decision tree was created in that way depends wholly on the data used to produce the model in the first place. Using an incomplete or biased dataset could then generate decisions that are completely transparent in terms of how the system is making decisions, but not in the reasons why and whether those decisions are reflective of the landscape in which this system is deployed. Contrast that with a machine learning model trained on a standardized dataset like the Modified National Institute of Standards and Technology, or MNIST dataset. This dataset contains images of handwritten digits used for the supervised learning task of character recognition, and it has been made available by academic research teams. AI practitioners can use their own specific machine learning algorithm to create a supervised machine learning model based on this dataset. As a result, many different methods have been used to pre-process this data, and many different modeling techniques have been applied. As such, the classification decision for a specific input and its accuracy range widely. So now, we have an example of a case in which the dataset may be completely transparent but the model creation may not be. Practitioners might begin with different pre-processing techniques, use different algorithms, and otherwise structure their model building process differently. Those decisions all matter to the transparency of the end system. Transparency is important in the design of any AI system, but it's especially important when decisions have a high impact on those using the system. Any case where the result of the system can significantly impact the livelihood and opportunity of those affected by it should expect a high degree of visibility. But even though we might want to expect it as consumers and users of these systems, historically transparency has not been the norm. Systems like credit scoring and those that attempt to predict the likelihood of someone's relapsed or criminal behavior are rarely transparent. In some cases, they're deliberately opaque. In a business context, transparency might run counter to other values including keeping trade secrets and intellectual property. Businesses may not wish to share their secret sauce with outsiders or even within the company itself. Ultimately, there must be a balance between the expectations of transparency and the impact of a system.

Intermediate Provider : Briana Brownell 5 Lessons · 14m in Arabic, German, English, Spanish, French, Portuguese, Chinese

Course Description

In these lessons on Transparency and Explainability in Artificial Intelligence, you’ll learn about explainability and interpretability and its intersection with the concepts of black box and glass box AI. You’ll review the advantages and disadvantages of open-source software in the development of an AI system. You’ll also learn about the principles of notification of decision and interaction, and their importance with systems that use natural language or facial recognition.

The principle of transparency suggests that effective oversight of AI systems should be possible. These lessons will develop your understanding of the concept of transparency in decision systems, including different levels of opacity, key concepts around transparency including explainability, interpretability, black box and glass box AI using human-in-the-loop systems.

Explainability is one of the principles for AI ethics. Explainability is a key factor that allows humans to analyze the functioning of an AI system. This section will introduce you to the concepts of black box and glass box AI, and how these models can be used to analyze and explain AI decision-making. You’ll also gain insight into how a black box model is used in the context of interpretability.

You’ll review concepts of transparency, error checking, democratization and access as each relates to open-source software. Then, you’ll discover the principles of notification of decision and interaction, and their importance when interacting with systems using natural language, such as chatbots, or facial recognition being used in public.

What You'll Learn

Describe the principle of transparency and its relationship to black box and glass box systems
Identify why explainability is important in AI
Explain the advantages and disadvantages of open-source software in AI development
Distinguish explainability from interpretability and their intersection with black box and glass box AI
Recognize the principles of notification of decision and interaction for systems using natural language or facial recognition
Examine transparency, error checking, democratization, and access as they relate to open-source software

Key Takeaways

The principle of transparency suggests that effective oversight of AI systems should be possible.
Explainability is one of the principles for AI ethics and is a key factor that allows humans to analyze the functioning of an AI system.
Transparency in decision systems involves different levels of opacity and key concepts such as explainability, interpretability, and black box and glass box AI using human-in-the-loop systems.
Black box and glass box models can be used to analyze and explain AI decision-making, and a black box model is used in the context of interpretability.
Notification of decision and interaction matters when interacting with systems using natural language, such as chatbots, or facial recognition used in public.

Frequently Asked Questions

What topics does this course cover?

It covers transparency and explainability in AI, including explainability and interpretability, black box and glass box AI, the advantages and disadvantages of open-source software, and the principles of notification of decision and interaction.

What will I learn about open-source software in this course?

You will review the advantages and disadvantages of open-source software in developing an AI system, including the concepts of transparency, error checking, democratization, and access as each relates to open-source software.

How does the course address natural language and facial recognition systems?

It covers the principles of notification of decision and interaction and their importance when interacting with systems using natural language, such as chatbots, or facial recognition being used in public.

What lessons are included in this course?

The lessons are Transparency and Oversight; Explainability and Interpretability; Accountability in Black Box and Glass Box; Open Source and the Value of Common Algorithms; and Notification of Decision and Interaction.

What skills does this course help develop?

It develops skills in Artificial Intelligence Systems, Artificial Intelligence, Artificial Intelligence Development, Explainable AI (XAI), and Transparency in Human-Computer Interaction.