Large language models (LLMs) can translate language, summarize text, generatecreative writing, generate code, power chatbots and virtual assistants, andcomplement search engines and recommendation systems. At the same time, as anearly-stage technology, its evolving capabilities and uses create potential formisapplication, misuse, and unintended or unforeseen consequences. Largelanguage models can generate output that you don't expect, including text that'soffensive, insensitive, or factually incorrect.

What's more, the incredible versatility of LLMs is also what makes it difficultto predict exactly what kinds of unintended or unforeseen outputs they mightproduce. Given these risks and complexities, Vertex AI generative AI APIs are designed withGoogle's AI Principles in mind. However, it is important for developers to understandand test their models to deploy safely and responsibly. To aid developers, theVertex AI Studio has built-in content filtering, and our generative AI APIs havesafety attribute scoring to help customers test Google's safety filters anddefine confidence thresholds that are right for their use case and business.Refer to theSafety filters and attributessection to learn more.

When our generative APIs are integrated into your unique use case and context,additional responsible AI considerations andlimitationsmight need to be considered. We encourage customers to promote fairness,interpretability, privacy and securityrecommended practices.

Safety filters and attributes

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

Model limitations

Limitations you can encounter when using generative AI models include (butare not limited to):

• Edge cases: Edge cases refer to unusual, rare, or exceptional situationsthat are not well-represented in the training data. These cases can lead tolimitations in the performance of the model, such as model overconfidence,misinterpretation of context, or inappropriate outputs.

• Model hallucinations, grounding, and factuality: Generative AI modelscan lack factuality in real-world knowledge, physical properties, oraccurate understanding. This limitation can lead to model hallucinations,which refer to instances where it can generate outputs that areplausible-sounding but factually incorrect, irrelevant, inappropriate, ornonsensical. To reduce this chance, you can ground the models to yourspecific data. To learn more about grounding in Vertex AI, seeGrounding overview.

• Data quality and tuning: The quality, accuracy, and bias of the promptor data input into a model can have a significant impact on itsperformance. If users enter inaccurate or incorrect data or prompts, themodel can have suboptimal performance or false model outputs.

• Bias amplification: Generative AI models can inadvertently amplifyexisting biases in their training data, leading to outputs that can furtherreinforce societal prejudices and unequal treatment of certain groups.

• Language quality: While the models yield impressive multilingualcapabilities on the benchmarks we evaluated against, the majority of ourbenchmarks (including all of fairness evaluations) are in the Englishlanguage. For more information, see theGoogle Research blog.

  • Generative AI models can provide inconsistent service quality todifferent users. For example, text generation might not be as effectivefor some dialects or language varieties due to underrepresentation inthe training data. Performance can be worse for non-English languages orEnglish language varieties with less representation.

• Fairness benchmarks and subgroups: Google Research's fairness analysesof our generative AI models don't provide an exhaustive account of thevarious potential risks. For example, we focus on biases along gender, race,ethnicity and religion axes, but perform the analysis only on the Englishlanguage data and model outputs. For more information, see theGoogle Research blog.

• Limited domain expertise: Generative AI models can lack the depth ofknowledge required to provide accurate and detailed responses on highlyspecialized or technical topics, leading to superficial or incorrectinformation. For specialized, complex use cases, models should be tuned ondomain-specific data, and there must be meaningful human supervision incontexts with the potential to materially impact individual rights.

• Length and structure of inputs and outputs: Generative AI models have amaximum input and output token limit. If the input or output exceeds thislimit, our safety classifiers are not applied, which could ultimately leadto poor model performance. While our models are designed to handle a widerange of text formats, their performance can be affected if the input datahas an unusual or complex structure.

Recommended practices

To utilize this technology safely and responsibly, it is also important toconsider other risks specific to your use case, users, and business context inaddition to built-in technical safeguards.

We recommend taking the following steps:

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

Report abuse

You can report suspected abuse of the Service or any generated output thatcontains inappropriate material or inaccurate information by using the followingform:Report suspected abuse on Google Cloud.

Additional resources

• Learn aboutabuse monitoring.
• Learn more about Google's recommendations forResponsible AI practices.
• Read our blog,A shared agenda for responsible AI progress