Focus on Mitigating AI Prejudice Rather than Aiming for Its Eradication.
Navigating the field of Artificial Intelligence (AI) comes with a significant challenge: bias. Traditional safeguards may not be enough to eliminate this inherent bias in AI results. But that's not to say we can't manage it. Here's a practical approach to help reduce the chances of bias in AI outcomes:
Decide on data and design
Since eliminating bias altogether is an idealistic goal, set realistic expectations for bias-free results by determining an acceptable threshold. Prioritize issues based on importance, focusing on fairness across all aspects of your AI development work.
The selection of data plays a critical role in AI development. Ensure equal representation of each class within the dataset. However, fair representation may not yield equal results for all classes. Assess whether proportional representation is more appropriate for the problem at hand. Consider accuracy differences among various groups and address any concerns accordingly.
Align these choices with company policy to maintain consistency throughout the development process. Verify fairness by testing with a diverse group of individuals, ensuring that risks are minimized across the population.
To ensure data sets are diverse and representative of the broader population, consider the following steps:
- Ensure that all classes within the dataset are represented.
- Identify and manage biases that have occurred during data selection.
- Test the AI model with data that is reflective of the intended population.
Check outputs
After sound data and a design strategy, it's crucial to review the fairness of the AI output. A misguided approach could be detrimental. A two-model solution like generative adversarial networks (GANs) can be effective in this regard. In essence, GANs serve as a comparison between the original approach and a second model that checks for fairness. This combined approach should lead to more fair AI outcomes.
Monitor for problems
Ongoing monitoring is vital in identifying patterns that deviate from expected results. It's essential to keep a vigilant eye on models' performance, as even those that pass various tests could still yield biased outcomes. Individuals have become accustomed to bias, making it challenging to spot these biases. However, by monitoring regularly, it's possible to identify rare events and less common miscalculations or errors.
Companies can never completely eliminate bias, but they can monitor and correct their practices to foster more fair, diverse, and equitable AI results. Ultimately, the goal is to create AI systems that function ethically and contribute positively to society.
Reference:Sian Townson, MIT Sloan Management Review, 2023/01
In the context of AI development, it's important to harness technology like generative adversarial networks (GANs) to check the fairness of AI outputs, ensuring a more equitable approach to AI outcomes. Companies should also continuously monitor their AI systems to identify and address any deviations from expected results, working towards building AI that functions ethically and positively impacts society.
