AI Ethics and Accountability: Critical Artificial Intelligence Incidents

Name: AI Ethics and Accountability: Critical Artificial Intelligence Incidents
Author: Briana Brownell

Learn about the negative outcomes of AI through specific case studies

Preview the first lesson free — get full access to all 3 lessons.

Course: On-Demand

There have been numerous incidents where artificial intelligence systems have been deployed with significant ethical implications. Many of these incidents involve bias and discrimination. Historically, bias and data sets and model building was considered a minor issue, subordinate to model accuracy, but several significant incidents have made practitioners more aware of the issues and the ethical requirements that they should be applying in their system designs. Let's consider two cases of bias and discrimination. In 2014, Amazon was trying to create an AI system that would help their hiring managers more easily identify top job candidates. To do this, the company began building a tool that could automate the review of resumes and use keywords to identify the top candidates. The company used data on resumes and Amazon hires from the past 10 years as part of a supervised machine learning model to essentially codify the company's past hiring practices into an automated AI system. What they did not realize was that the decisions of the hiring managers in the past had been discriminatory. The AI system was hence codifying this discrimination into the system they were building. The company eventually recognized that the machine learning model was discriminating against women, perpetuating a bias that was present in their training data. It was downgrading resumes with words like women's, such as ones that might occur in a resume statement like captain of the women's soccer team and mentions of certain women's colleges. Because Amazon's workforce was overwhelmingly male, alumni of these colleges were rare or non-existent in the training set. The AI system learned that graduates of these colleges were undesirable. Whether or not the company intended to discriminate against female applicants, the use of an algorithm that unfairly removes candidates of a particular gender, race, or religion is illegal. The company denied that the tool had been used exclusively to evaluate candidates and discontinued its use after the problems of discrimination came to light. Another common issue in AI discrimination and bias has to do with race in relation to image captioning and classification. During her research on facial recognition MIT grad student Joy Buolamwini found that the AI tools she was using had difficulty recognizing her face. As a black woman, her challenge using facial recognition was common. Many facial recognition systems are trained on data that contains more white men compared to other groups. As a result these systems are much less accurate for certain groups. She was not able to use the technology as easily as her white male peers. The problems with image classification are pervasive. Image captioning systems can perpetuate bias that may already exist in the training data. As in the example of Amazon's hiring system. In 2020 as the pandemic was raging around the world, citizens were encouraged to take their temperature to see whether they had a fever, one of the primary covid symptoms. One of the most common ways to measure fever was a handheld thermometer. But when Google's Cloud Vision produced image captions for people holding one of these devices, photographs with dark-skinned hands were caption as containing guns. While photos of Asian people holding them were labeled as an electronic device. Even when an identical image of a hand holding the device was used and the skin color changed, the device held by the darker-skinned hand was captioned as a gun, while the device in the light-skinned hand was captioned as a tool. These examples illustrate the importance of using unbiased data sets for training and for ensuring that systems do not perpetuate and scale human biases from the past.

Intermediate Provider : Briana Brownell 3 Lessons · 14m in Arabic, German, English, Spanish, French, Portuguese, Chinese

Course Description

These lessons on Critical Artificial Intelligence Incidents looks at significant AI incidents through a series of real-world examples of negative outcomes in AI, how they were resolved, and how they could have been prevented. Specifically, you’ll work through case studies focusing on issues of identity and discrimination, offensive content, and legal implications. You’ll first review a case of AI hiring discrimination where female applicants were excluded, and then you’ll look at a case of biased image AI that resulted in offensive photo descriptions.

The second series of case studies includes a generation and amplification of offensive content on popular internet platforms because of bad actors. Through the final case studies, you’ll explore cases where AI systems intersect with the law, and you’ll gain insight into the legal implications of such interactions.

What You'll Learn

Describe the ways in which AI systems can fail or be problematic
Recognize the ethical issues raised by real-world AI use cases
Examine AI failures involving identity and discrimination through case studies
Analyze how AI systems generate or amplify offensive content
Explore cases where AI systems intersect with the law and their legal implications
Learn from past AI incidents to understand how they were resolved and prevented

Key Takeaways

The course studies significant AI incidents through real-world examples of negative outcomes, examining how they were resolved and how they could have been prevented.
Case studies cover identity and discrimination, including a hiring AI that excluded female applicants and a biased image AI that produced offensive photo descriptions.
One series of case studies addresses the generation and amplification of offensive content on popular internet platforms caused by bad actors.
The final case studies explore how AI systems intersect with the law and the legal implications of those interactions.
The lessons aim to build understanding of AI ethics by learning from past problems with AI.

Frequently Asked Questions

What does this course cover?

It examines significant AI incidents through real-world case studies of negative outcomes, focusing on identity and discrimination, offensive content, and legal implications, including how the incidents were resolved and how they could have been prevented.

What real-world examples are discussed?

Examples include a case of AI hiring discrimination where female applicants were excluded, a biased image AI that produced offensive photo descriptions, the amplification of offensive content on popular internet platforms by bad actors, and cases where AI systems intersect with the law.

What will I be able to do after taking this course?

You will be able to describe ways AI systems can fail or be problematic, recognize the issues involving AI ethics in real-world use cases, and learn from past issues with AI.

What topics or lessons are included?

The lessons are AI Issues in Identity and Discrimination, AI and Offensive Content, and AI Translators and Legal Implications.

What skills does this course develop?

It develops skills in Applications of Artificial Intelligence, Artificial Intelligence Systems, Artificial Intelligence, Artificial Intelligence Development, Data Ethics, and Explainable AI (XAI).