Pre-processing: filtering what goes into the generative AI model or tool

AI can be guided and grounded by controlling inputs before they reach the model. Pre-processing can strip out sensitive data like personally identifiable information (PII), spelling errors/typos, or filter out hate speech or proper nouns (like a brand name or celebrity) before it ever enters the model or tool. Pre-processing can also lead to clarifying questions for the user, like "did you mean Safari like the Apple web browser, or a trip through the wilderness?" before returning a generated image for the term 'safari.'

• Pros: Reduces risks without retraining the model.
• Cons: Adds overhead and may miss edge cases if filters are poorly designed.
• Think about: Input sanitization, moderation APIs, truth-checking workflows, and layered review systems.

​