Responsible AI

This guide describes the responsible AI features in Vertex AI, potential model limitations, and best practices for developing generative AI applications safely and responsibly. This page covers the following topics:

  • Safety filters and attributes: Learn about the built-in content filtering and safety attribute scoring available in Vertex AI generative AI APIs.
  • Model limitations: Understand potential limitations of generative models, such as hallucinations, bias, and limited domain expertise.
  • Recommended practices: Follow recommended steps to assess security risks, perform safety testing, and monitor your application.
  • Report abuse: Find out how to report suspected abuse or inappropriate generated content.

When you integrate the generative APIs into your use case and context, you might need to consider additional responsible AI factors and limitations. To promote fairness, interpretability, privacy, and security, follow the recommended practices.

Safety filters and attributes

To learn how to use safety filters and attributes for an API, see Gemini API in Vertex AI.

Model limitations

Generative AI models have limitations that you might encounter, including the following:

  • Edge cases: Unusual, rare, or exceptional situations that are not well-represented in the training data. These cases can lead to performance limitations, such as model overconfidence, misinterpreting context, or generating inappropriate outputs.

  • Model hallucinations, grounding, and factuality: Generative AI models can produce outputs that sound plausible but are factually incorrect because they can lack real-world knowledge or an accurate understanding of physical properties. To reduce hallucinations, you can ground models to your specific data. To learn more, see Grounding overview.

  • Data quality and tuning: The quality, accuracy, and bias of the input prompt or data significantly impact model performance. Inaccurate or biased input can lead to suboptimal performance or false outputs.

  • Bias amplification: Generative AI models can amplify existing biases from their training data, leading to outputs that reinforce societal prejudices and the unequal treatment of certain groups.

  • Language quality: Model performance can be inconsistent across languages, dialects, and language varieties. Languages or dialects that are underrepresented in the training data might have lower performance. While models have impressive multilingual capabilities, most benchmarks, including all fairness evaluations, are in English. For more information, see the Google Research blog.

  • Fairness benchmarks and subgroups: Google Research's fairness analyses don't cover all potential risks. For example, the analyses focus on biases related to gender, race, ethnicity, and religion, but use only English language data and model outputs. For more information, see the Google Research blog.

  • Limited domain expertise: Models can lack the deep knowledge required for highly specialized or technical topics, which can lead to superficial or incorrect information. For specialized use cases, consider tuning models on domain-specific data and including meaningful human supervision, especially in contexts that can impact individual rights.

  • Length and structure of inputs and outputs: Models have a maximum token limit (pieces of words) for inputs and outputs. If the input or output exceeds this limit, the model doesn't apply safety classifiers, which can lead to poor model performance. In addition, unusual or complex input data structures can negatively affect performance.

To use this technology safely and responsibly, consider risks specific to your use case in addition to the built-in technical safeguards.

We recommend that you take the following steps:

  1. Assess your application's security risks.
  2. Perform safety testing that is appropriate for your use case.
  3. Configure safety filters, if required.
  4. Solicit user feedback and monitor content.

Report abuse

To report suspected abuse of the service or any generated output that contains inappropriate material or inaccurate information, use the Report suspected abuse on Google Cloud form.

Additional resources