Generative AI Google Brain

Responsible AI - Best Practices

3 min read

Responsible AI — Best Practices

Distilled from 4 notebooks tagged Responsible AI in the GoogleCloudPlatform/generative-ai repository. The From the notebooks section below cites the per-notebook source for grounding.

Do this

  • Be concise, specific, and well-defined in prompts.
  • Ask one task at a time to reduce ambiguity and output variability.
  • Use system instructions to guardrail the model away from irrelevant or unsafe responses.
  • Convert generative tasks into classification tasks when lower variability is important.
  • Use zero-shot, one-shot, or few-shot prompting based on the goal, with one to five representative examples.
  • Keep example distribution aligned with the real distribution for classification tasks.
  • State that provided context is the only source of truth when grounding matters.
  • For long documents or data, place the question after the context and ask the model to answer based on the full document above.
  • For complex prompts, place the most important constraints at the end.
  • Use persona carefully because it can override other instructions.
  • Inspect safety_ratings and finish_reason instead of relying only on generated text.
  • Set safety thresholds according to business policies and use case needs.
  • Check both input and output with DLP before sending content to Gemini or returning content to users.
  • Treat user-provided links, binaries, files, retrieved documents, and tool data as untrusted and validate them.
  • Use defense in depth: combine safety filters, DLP, embeddings-based dangerous-prompt detection, mission checks, schema validation, OCR when needed, and user consent for dangerous operations.

Avoid this

  • Assuming Gemini has real-time information without additional integrations, which can lead to confident hallucinations.
  • Treating prompted citations as a reliable hallucination fix, even though citations can be false or inaccurate.
  • Using too many examples, which can overfit the prompt and reduce response quality.
  • Ignoring response metadata such as safety_ratings and finish_reason, especially when response.text is empty because filters blocked output.
  • Turning safety filters to BLOCK_NONE outside a demonstration context.
  • Relying on simple regex ReAct parsing or small embedding datasets as production-grade security controls.
  • Assuming Cloud DLP catches all sensitive data, including transformed or encoded values.
  • Forgetting required setup such as Google Cloud project configuration, Vertex AI API enablement, Colab authentication, quota project setup, or dependency/runtime restarts.

From the notebooks

Prompt Design - Best Practices

  • Be concise in prompts.
  • Be specific and well-defined.
  • Ask one task at a time.
  • Use system instructions to guardrail the model from irrelevant responses.
  • Turn generative tasks into classification tasks to reduce output variability.

Gen AI & LLM Security for developers

  • Do not store sensitive information in the prompt.
  • Use low temperature for reproducible results in security demonstrations.
  • Check both input and output with DLP before sending to or returning from Gemini.
  • Treat links, binaries, and files from users as untrusted and validate them.
  • Use Responsible AI safety filters and understand how to configure safety attributes.

Responsible AI with Gemini API in Vertex AI: Safety ratings and thresholds

  • Inspect safety_ratings instead of relying only on generated text.
  • Test prompts against safety categories before deployment.
  • Set safety thresholds according to business policies and use case needs.
  • Use low-variability generation settings when comparing safety behavior.
  • Check finish_reason to understand why generation stopped.

Gen AI and LLM Security - ReAct and RAG attacks & mitigations

  • Use ready Agents and RAG libraries such as Agent Builder, LangChain Agents, Vertex AI Search, and LangChain RAG.
  • Use strict schema validation of tool input and output.
  • Use out-of-band user consent for dangerous operations.
  • Apply defense in depth by layering multiple filters.
  • Use OCR for documents if concerned about invisible text.

Back to Responsible AI · Best Practices Map

Graph View

Backlinks