Machine learning models are increasingly used to inform high stakes decision-making that relates to people. Although machine learning, by its very nature, is a form of statistical discrimination, the discrimination becomes objectionable when it places privileged groups at systematic advantage and certain unprivileged groups at systematic disadvantage, potentially causing varied harms. Biases in training data, due to either prejudice in labels or under-/over-sampling, yields models with unwanted bias.

• Diverse and representative data

Ensure that the training data used to build AI models is diverse and representative of the population it is meant to serve. Include data inputs from various demographic groups to avoid underrepresentation or bias. Regularly check and assess training data for biases. Use tools and methods to identify and correct biases in the dataset before training the model.

• Bias-aware algorithms

Incorporate fairness metrics into the development process to assess how different subgroups are affected by the model's predictions. Monitor and minimize disparities in outcomes across various demographic groups. Apply constraints in the algorithm to ensure that the model adheres to predefined fairness criteria during training and deployment.

• Bias mitigation techniques

Apply techniques like re-sampling, re-weighting and adversarial training to mitigate biases in the model's predictions.

• Diverse development teams

Assemble interdisciplinary and diverse teams involved in AI development. Diverse teams can bring different perspectives to the table, helping to identify and rectify biases that may be overlooked by homogeneous teams.

• Ethical AI review boards

Establish review boards or committees to evaluate the potential biases and ethical implications of AI projects. These boards can provide guidance on ethical considerations throughout the development lifecycle.

Robust AI effectively handles exceptional conditions, such as abnormalities in input or malicious attacks, without causing unintentional harm. It is also built to withstand intentional and unintentional interference by protecting against exposed vulnerabilities. Our increased reliance on these models and the value they represent as an accumulation of confidential and proprietary knowledge, are at increasing risk for attack. These models pose unique security risks that must be accounted for and mitigated.

Users must be able to see how the service works, evaluate its functionality, and comprehend its strengths and limitations. Increased transparency provides information for AI consumers to better understand how the AI model or service was created. This helps a user of the model to determine whether it is appropriate for a given use case, or to evaluate how an AI produced inaccurate or biased conclusions.  

Many regulatory frameworks, including GDPR, mandate that organizations abide by certain privacy principles when processing personal information. A malicious third party with access to a trained ML model, even without access to the training data itself, can still reveal sensitive personal information about the people whose data was used to train the model. It is crucial to be able to protect AI models that may contain personal information, and control what data goes into the model in the first place.