Shipping generative AI features users can trust

With generative AI opening up new possibilities for web experiences, the responsibility falls on developers to make sure those features are safe, fair, and reliable. Google’s responsible AI practices provide a baseline, but your actual work is in the implementation—and in documenting your approach as you go. Keeping records of your decisions, results, and safeguards helps you and your team maintain and improve these applications over time.

If you are building sites, apps, or extensions that generate content, a few core practices should be part of your workflow:

  • Test AI tools and models rigorously before launch.
  • Leverage pre-existing safety tools and configuration options.
  • Give users a clear path to report offensive content—and act on those reports.
  • Make sure your advertising accurately reflects what your product actually does.

Testing is your first line of defense

You are accountable for the experience your application delivers, so it’s essential to understand the underlying AI models you depend on. That means testing against a wide range of use cases and being deliberate about guarding against prompts that might produce harmful or offensive output. Thorough testing also helps you anticipate failure modes before they become user-facing problems.

Turn on built-in safety filters

Many APIs ship with safety controls that act as a barrier against harmful output. For example, Vertex AI provides configuration options for safety attributes. Note that these filters don’t alter the model’s behavior—they only block output. Still, you should review the available thresholds and set them to a level that fits your particular use case.

Plan for what safety filters can’t catch

No testing scheme or filter is perfect, and offensive content can still slip through. Build a process for handling those cases before they happen. Monitor user feedback around your generative AI features closely and prioritize it, then establish a clear workflow for dealing with reports.

When you do encounter offensive output, you have several options for responding:

  • Block similar inputs that led to the problem.
  • Redact the model’s output.
  • Fine-tune the model to correct its behavior.
  • Switch to a different model entirely.

If your work includes browser extensions headed for the Chrome Web Store, you also need to follow the Chrome Web Store policies. For content published with AI assistance, aim for output that is helpful, reliable, and people-first.

Marketing matters too

Being responsible also extends to the ads you run. Review all marketing materials to make sure ads don’t overpromise what your AI features can deliver. Honest representation of your product’s capabilities is not just good ethics—it prevents users from feeling misled once they try the experience.

For more guidance, Google’s People + AI Guidebook covers a range of best practices for human-centered AI design. As the ecosystem evolves, staying deliberate about these principles will keep your generative AI work both innovative and responsible.