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.
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.
Recommended practices
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:
Assess your application's security risks.
Perform safety testing that is appropriate for your use case.
Configure safety filters, if required.
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.
[[["容易理解","easyToUnderstand","thumb-up"],["確實解決了我的問題","solvedMyProblem","thumb-up"],["其他","otherUp","thumb-up"]],[["難以理解","hardToUnderstand","thumb-down"],["資訊或程式碼範例有誤","incorrectInformationOrSampleCode","thumb-down"],["缺少我需要的資訊/範例","missingTheInformationSamplesINeed","thumb-down"],["翻譯問題","translationIssue","thumb-down"],["其他","otherDown","thumb-down"]],["上次更新時間：2025-08-28 (世界標準時間)。"],[],[],null,["# Responsible AI\n\nLarge language models (LLMs) can translate language, summarize text, generate\ncreative writing, generate code, power chatbots and virtual assistants, and\ncomplement search engines and recommendation systems. At the same time, as an\nearly-stage technology, its evolving capabilities and uses create potential for\nmisapplication, misuse, and unintended or unforeseen consequences. Large\nlanguage models can generate output that you don't expect, including text that's\noffensive, insensitive, or factually incorrect.\n\nWhat's more, the incredible versatility of LLMs is also what makes it difficult\nto predict exactly what kinds of unintended or unforeseen outputs they might\nproduce. Given these risks and complexities, Vertex AI generative AI APIs are designed with\n[Google's AI Principles](https://ai.google/principles/) in mind. However, it is important for developers to understand\nand test their models to deploy safely and responsibly. To aid developers, the\nVertex AI Studio has built-in content filtering, and our generative AI APIs have\nsafety attribute scoring to help customers test Google's safety filters and\ndefine confidence thresholds that are right for their use case and business.\nRefer to the [Safety filters and attributes](#safety_filters_and_attributes)\nsection to learn more.\n\nWhen our generative APIs are integrated into your unique use case and context,\nadditional responsible AI considerations and\n[limitations](#limitations)\nmight need to be considered. We encourage customers to promote fairness,\ninterpretability, privacy and security\n[recommended practices](https://ai.google/responsibilities/responsible-ai-practices/).\n\nSafety filters and attributes\n-----------------------------\n\nTo learn how to use safety filters and attributes for an API,\nsee [Gemini API in Vertex AI](/vertex-ai/generative-ai/docs/multimodal/configure-safety-attributes).\n\nModel limitations\n-----------------\n\n*Limitations you can encounter when using generative AI models include (but\nare not limited to):*\n\n- **Edge cases**: Edge cases refer to unusual, rare, or exceptional situations\n that are not well-represented in the training data. These cases can lead to\n limitations in the performance of the model, such as model overconfidence,\n misinterpretation of context, or inappropriate outputs.\n\n- **Model hallucinations, grounding, and factuality** : Generative AI models\n can lack factuality in real-world knowledge, physical properties, or\n accurate understanding. This limitation can lead to model hallucinations,\n which refer to instances where it can generate outputs that are\n plausible-sounding but factually incorrect, irrelevant, inappropriate, or\n nonsensical. To reduce this chance, you can ground the models to your\n specific data. To learn more about grounding in Vertex AI, see\n [Grounding overview](/vertex-ai/generative-ai/docs/grounding/overview).\n\n- **Data quality and tuning**: The quality, accuracy, and bias of the prompt\n or data input into a model can have a significant impact on its\n performance. If users enter inaccurate or incorrect data or prompts, the\n model can have suboptimal performance or false model outputs.\n\n- **Bias amplification**: Generative AI models can inadvertently amplify\n existing biases in their training data, leading to outputs that can further\n reinforce societal prejudices and unequal treatment of certain groups.\n\n- **Language quality** : While the models yield impressive multilingual\n capabilities on the benchmarks we evaluated against, the majority of our\n benchmarks (including all of fairness evaluations) are in the English\n language. For more information, see the\n [Google Research blog](https://ai.googleblog.com/2022/04/pathways-language-model-palm-scaling-to.html).\n\n - Generative AI models can provide inconsistent service quality to different users. For example, text generation might not be as effective for some dialects or language varieties due to underrepresentation in the training data. Performance can be worse for non-English languages or English language varieties with less representation.\n- **Fairness benchmarks and subgroups** : Google Research's fairness analyses\n of our generative AI models don't provide an exhaustive account of the\n various potential risks. For example, we focus on biases along gender, race,\n ethnicity and religion axes, but perform the analysis only on the English\n language data and model outputs. For more information, see the\n [Google Research blog](https://ai.googleblog.com/2022/04/pathways-language-model-palm-scaling-to.html).\n\n- **Limited domain expertise**: Generative AI models can lack the depth of\n knowledge required to provide accurate and detailed responses on highly\n specialized or technical topics, leading to superficial or incorrect\n information. For specialized, complex use cases, models should be tuned on\n domain-specific data, and there must be meaningful human supervision in\n contexts with the potential to materially impact individual rights.\n\n- **Length and structure of inputs and outputs**: Generative AI models have a\n maximum input and output token limit. If the input or output exceeds this\n limit, our safety classifiers are not applied, which could ultimately lead\n to poor model performance. While our models are designed to handle a wide\n range of text formats, their performance can be affected if the input data\n has an unusual or complex structure.\n\nRecommended practices\n---------------------\n\nTo utilize this technology safely and responsibly, it is also important to\nconsider other risks specific to your use case, users, and business context in\naddition to built-in technical safeguards.\n\nWe recommend taking the following steps:\n\n1. Assess your application's security risks.\n2. Perform safety testing appropriate to your use case.\n3. Configure safety filters if required.\n4. Solicit user feedback and monitor content.\n\nReport abuse\n------------\n\nYou can report suspected abuse of the Service or any generated output that\ncontains inappropriate material or inaccurate information by using the following\nform:\n[Report suspected abuse on Google Cloud](https://support.google.com/code/contact/cloud_platform_report).\n\nAdditional resources\n--------------------\n\n- Learn about [abuse monitoring](/vertex-ai/generative-ai/docs/learn/abuse-monitoring).\n- Learn more about Google's recommendations for [Responsible AI practices](https://ai.google/responsibilities/responsible-ai-practices/?category=general).\n- Read our blog, [A shared agenda for responsible AI progress](https://blog.google/technology/ai/a-shared-agenda-for-responsible-ai-progress/)"]]
