Google announces new tools in AI Studio to explore, debug, and share logs. If you build apps with Gemini models or use the GenerateContent API, you can now turn on centralized logging without touching your code and convert real interactions into reusable datasets for evaluation.

What the new tool brings

The main feature is simple and powerful: with one click on Enable logging from the AI Studio panel, all supported calls to the GenerateContent API for that billing-enabled project become visible in the dashboard. No code changes required, and it records both successful responses and errors.

You can obtain logging at no monetary cost in all regions where the Gemini API is available.

The logs include key attributes: response codes, inputs (prompts), outputs (responses), and API tool usage. That lets you filter by status, dive into specific interactions, and trace a user complaint back to the exact model call.

Observability and debugging (technical)

The observability improvements hit the pain points for people building AI systems:

• Fast response to failures: filter by status codes to group network errors, quota limits, or model faults.
• Traceability of inputs/outputs: inspect the prompt, metadata, and response to reproduce the issue locally.
• Tool usage: see when and how the model relied on external tools or intermediate steps.

This shortens the debug cycle: instead of guessing what went wrong, you locate the problematic interaction and reproduce it in a controlled environment.

Exportable datasets and reproducible evaluation

Logs can be exported as CSV or JSONL. What does that do for you in practice?

• Create reproducible baselines: pick examples where quality dropped and build a test suite that defines expected behavior.
• Prompt refinement: use real failing examples to tweak templates, system instructions, or few-shot strategies.
• Batch evaluation: with the Gemini Batch API you can run large-scale evaluations against the accumulated datasets and compare model versions or changes in your app logic.

Concrete example: if your e-commerce assistant starts mixing up returns and refunds, export interactions labeled as problematic, fix the prompt, and run a batch evaluation to confirm the improvement before deploying.

Share datasets with Google and privacy considerations

You have the option to share specific datasets with Google to provide feedback on end-to-end model behavior. Google will use that data to improve products and models.

Before sharing, consider these best practices:

• Anonymize sensitive information and remove PII.
• Apply sampling and review content to avoid exposing legal or medical data without permission.
• Check your organization policies and the service agreement: sharing data should align with compliance and privacy requirements.

Also remember to confirm whether there are cost implications for storage or transfer when exporting datasets out of the dashboard.

Quick steps to get started (checklist)

1. Go to Google AI Studio in the project with billing enabled.
2. Click Enable logging to start capturing GenerateContent API calls.
3. Explore the dashboard: filter by response codes, inputs, outputs, and tool usage.
4. Export selected logs as CSV or JSONL and create an evaluation dataset.
5. Run evaluations with Gemini Batch API or use the data to refine prompts and automated tests.
6. If you decide to share, anonymize and validate the data before sending it to Google.

Practical recommendations for technical teams

• Integrate a regular process to extract problematic examples: weekly export the most relevant cases and add them to your test pipeline.
• Keep version tracking: tag each exported dataset with the model and prompt versions used to make comparisons easier.
• Automate alerts based on log patterns (for example: a spike in a certain error code) to catch regressions in production.

Google AI Studio is providing infrastructure to make the observability-to-improvement loop more direct: from capturing real interaction to validating changes with batch evaluations.

Ready to turn real logs into measurable improvements? Start by prototyping in Build mode and keep monitoring throughout your app's lifecycle.

Original source

https://blog.google/technology/developers/google-ai-studio-logs-datasets