Established in May 2018, the Toronto Declaration says that artificial intelligence (AI) should not create new barriers to diversity, equality, or representation. The declaration further states that when an AI system discriminates, and the discrimination cannot be resolved, the system should be questioned and curtailed. Governments and companies that use AI are responsible for testing it before, during, and after deployment to ensure it behaves fairly and is not biased.
This guide discusses the feasibility, strategies, and challenges of AI ethics testing.
What is AI ethics testing?
AI ethics testing is a framework for assessing AI models for prejudice, bias, social and environmental impact, data privacy, transparency, explainability, safety, and security. AI ethics testing facilitates human oversight of AI and must be conducted alongside traditional AI testing. To pass an ethics test, AI models must provide affirmative evidence of fairness and also must be able to demonstrate the systems of decision-making and qualities of justice. If an AI system reveals biases or injustices, it fails the ethics test. Additionally, if the system cannot be explained, it fails the ethics test.
Is it possible to "test" AI ethics?
Although AI models may not have inherent ethical values, it is feasible to test AI ethics. In the context of AI, testing refers to evaluations of ethical implications and implications of AI systems before the systems are deployed and consistently throughout their implementation and use. AI ethics testing must also involve an assessment of the motivations, assumptions, and risks potentially posed by the entities conducting the testing.
Strategies for Testing AI Ethics
There are many approaches to testing AI ethics. Here are three distinct perspectives:
In UNESCO's Recommendation on the Ethics of Artificial Intelligence, UNESCO details eleven key policy areas for responsible AI developments. One of the policies is the Ethical Impact Assessment (EIA), a structured process that helps AI teams and communities address the impacts of AI systems. The following ethical questions are raised:
- Who is most likely to be adversely affected by this AI system?
- What form will these impacts take?
- What can be done to prevent these harms, and have resources been allocated to this harm prevention?
The computer software company, Adobe, created a standardized process for testing AI through the oversight of an AI Ethics Review Board guided by the company's AI Ethics principles. Adobe's strategy includes training AI models on data sets that pass ethical guidelines, conducting automated testing and human evaluation, assessing impacts, providing human oversight, and welcoming community feedback.
When given the prompt "detail strategies for testing AI ethics," followed by "write the strategies with less than 15 words per strategy," for the sake of brevity, ChatGPT suggested the following ten approaches to testing AI ethics:
- Scenario-based Testing: Simulate ethical dilemmas to assess system responses.
- Data Analysis: Evaluate training data for biases and underrepresentation.
- Evaluation Metrics: Define metrics to measure ethical performance.
- User Studies and Feedback: Gather user perspectives and feedback on ethical implications.
- Expert Reviews and Audits: Engage ethicists and auditors to assess compliance.
- Regulatory Compliance: Ensure adherence to legal and ethical standards.
- Red Team Testing: Simulate adversarial scenarios to test system resilience.
- Continuous Monitoring and Iterative Improvement: Monitor, analyze, and iterate for ethical performance.
- Collaboration and Diversity: Involve diverse perspectives for a comprehensive examination
- Ethical Frameworks and Guidelines: Refer to established principles for ethical evaluation.
Importance of AI Ethics Testing
AI enables countless opportunities and also presents startling ethical concerns. As UNESCO details, AI systems can potentially contain algorithmic biases, threaten human rights, and impact the planet.
A study conducted in February 2023 revealed that personal data privacy was the top AI-related issue concerning adults in the United States. However, recent events show that AI-related issues pertain to much more than personal data. Here are a few recent events:
Racial Bias in Healthcare AI
In 2019, a study found that a clinical healthcare algorithm recommended the same care to Black and White patients. However, the Black patients were much sicker than the white patients. The algorithm reflected long-standing wealth and income disparities between Black and white patients that resulted in Black patients having less to spend on healthcare and thus receiving less care, even for the same level of illness of white patients who had more economic privilege to receive more care.
Discrimination in Advertising
In 2018, The United States of America filed a complaint against Meta for violating the Fair Housing Act (FHA) with a housing advertising system that discriminated against Facebook users based on race, color, sex, disability, religion, family status, and national origin. The discrimination stemmed from algorithms that determined which users received housing ads based partly on characteristics and identities protected by the FHA. In 2023, the Justice Department reached a settlement with Meta that involved a civil penalty. It also resulted in the creation of the AI-based Variance Reduction System (VRS) to reduce bias in advertising and increase the equitable distribution of advertisements.
Inequity in Facial Recognition Algorithms
A study conducted in 2018, entitled "Gender Shades," examined the disparities in the accuracy of gender classification algorithms and showed that the algorithms were the least accurate with darker-skinned females who received error rates over 30% higher than algorithm performance for lighter-skinned males. In 2019, the National Institute of Standards and Technology (NIST) conducted an independent study of the demographic effects of face recognition algorithms. The study confirmed the results of "Gender Shades" across nearly 200 algorithms and showed that AI facial recognition technologies are the least accurate on women of color.
Challenges and Solutions for AI Ethics Testing
Testing AI ethics is challenging for many reasons. First of all, ethics requires a blend of intuition, experience, culture, circumstance, and situational variables that depend upon human sensitivity and intellect. Additionally, human sensitivity and intelligence are laden with identities, social locations, various privileged or marginalized statuses, and a wide range of influences that lead to biases, prejudices, and complexities that can be challenging to identify and address. Furthermore, the data sets that train AI models are embedded with biases, inequities, and practices that are unfair, dangerous, and unjust. Nothing about AI or AI ethics is straightforward.
The solution to these layers of complex challenges with AI ethics testing is an equally complex, nuanced, and dynamic approach to testing. This approach involves examining personal, institutional, and societal biases, inequities, and practices; critically assessing data sets; engaging in ongoing critical discussions and evaluations with stakeholders; taking corrective action when issues are detected; and advancing the field of AI ethics.
Companies can partner with Encora to adopt innovative and ethical AI solutions. Encora’s full Product Development Lifecycle and technology experts help build AI solutions that power several aspects of the emerging economy.