Model safety

AI model safety can be broadly defined as the measures taken to ensure that AI is deployed in ways that do not harm humanity. It ensures preventing accidents, misuse, or other harmful consequences that could result from artificial or machine learning systems. While the problem with AI safety is complex to solve, some factors can be considered when ensuring model safety:

• Data quality: Data-related issues such as sending the wrong data in production, data sanity issues, or data drifts can cause model failure and unfair or discriminatory outcomes. To ensure the reliability of the model, it is important to train it on high-quality, diverse datasets. 
• Continuous monitoring: Monitoring AI models regularly helps identify precisely when the model performance starts diminishing. Monitoring the automated workflows helps to maintain the required accuracy and keeps transformations error-free.
• Model explainability: Explainability is important not only for transparency but also for auditability and interpretability so stakeholders involved in the process can identify and intervene in case of any potential  issues like data drift, model drift
• Human-in-the-loop processes: Humans should be involved in workflow activities like developing and deploying AI models. This ensures that the model is used ethically and responsibly.
• Robustness and Stress-testing:  It is important to conduct stress testing before deploying the model in production to prevent issues such as sensitive information leakage and bias. Conducting thorough stress testing on AI/ML models helps to identify gaps and fill them with feedback or separate policies. These policies can reinforce or override the model's output, particularly in highly sensitive use cases.