ML has the potential to transform society in many meaningful ways, either positively or negatively. It's critical to consider the ethical implications of your models and the systems they're a part of.
Your ML projects should benefit society. They shouldn't cause harm or be susceptible to misuse. They shouldn't perpetuate, reinforce, or exacerbate biases or prejudices. They shouldn't collect or use personal data irresponsibly.Google's AI principles
Google advocates developing ML and AI applications that adhere to its AI principles. Google's principles center around the following concepts:
- Fairness
- Privacy
- Transparency
- Safety
Fairness
Avoid creating or reinforcing unfair bias. Models exhibit bias when their training data has some of the following characteristics:
Doesn't reflect the real-world population of their users.
Preserves biased decisions or outcomes, for example, criminal justice decisions like incarceration times.
Uses features with more predictive power for certain groups of users.
The previous examples are just some ways models become biased. Understanding your data thoroughly is critical for uncovering and resolving any potential biases it contains. The first step for developing fair models is verifying the training data accurately reflects the distribution of your users. The following are further practices to help create fair models:
Identify underrepresented groups in evaluation datasets or groups that might experience worse model quality compared to other groups. You might need to oversample a subgroup of your users to increase their presence in the training data.
Use golden datasets (also known as benchmark datasets) to validate the model against fairness issues and detect implicit bias.
Avoid including sensitive features in datasets, like gender or ethnicity.
Avoid including features with little empirical or explanatory power, but especially in sensitive contexts where the trained model is used to perform high-impact tasks in areas such as healthcare, finance, education, employment, and so forth. For example, in a model for approving home loans, don't include names in the training data. Not only is an applicant's name irrelevant to the prediction task, but leaving such an irrelevant feature in the dataset also has the potential to create implicit bias or allocative harms. For instance, the model might correlate male names with a higher probability for repayment, or vice versa.
Measure potential adverse impact a model's predictions might have on particular groups, and consider intentional bias correction techniques if you find adverse impact in a sensitive context.
Privacy
Incorporate privacy design principles from the beginning.
The following are privacy related laws and policies to be aware of and adhere to:
European Union's Digital Markets Act (DMA) for consent to share or use personal data.
European Union GDPR laws.
Moreover, be sure to remove all personally identifiable information (PII) from datasets and confirm your model and data repositories are set up with the right permissions, for example, not world-readable.
Transparency
Be accountable to people. For example, make it easy for others to understand what your model does, how it does it, and why it does it. Model cards provide a template to document your model and create transparency artifacts.
Safety
Design models to operate safely in adversarial conditions. For example, test your model with potential hostile inputs to confirm your model is secure. Furthermore, check for potential failure conditions. Teams typically use specially designed datasets to test their models with inputs or conditions that caused the model to fail in the past.
Check Your Understanding
Always consider the broader social contexts your models operate within. Work to be sure your handling of sensitive data doesn't violate privacy issues, perpetuate bias, or infringe on someone else's intellectual property.