Mitigate AI Hallucinations: Best Practices for Engineers

Understanding AI Hallucinations

The speaker defines AI hallucinations as the AI making up information or getting things wrong, but emphasizes that new models significantly decrease their occurrence.

Treat AI Like an Intern

An analogy is used to suggest treating AI like an intern, requiring human review of outputs rather than blind trust, but noting that major errors are becoming rare.

Mitigating Hallucinations: Use Newest Models

The first mitigation strategy is to always use the most recent versions of AI models (e.g., Gemini 3 Pro, GPT 5.1) for optimal results.

Enable Web Search & Source Linking

The speaker advises turning on web search tools and instructing the AI to link URLs to its sources, explaining its choice for verification.

Verify AI Sources for Accuracy

An example demonstrates how to prompt AI to find information, link sources, and explain its choices, allowing for quick human verification of reliability.

Require Specificity for Internal Documents

For internal document searches, the AI should be prompted to provide exact locations (folder, document, page, paragraph) and direct quotes to ensure accuracy.