Tests

A test is an input or question used to evaluate how reliably, accurately, and safely an AI model responds. Use tests in the Workbench to prompt models, score responses, compare setups, and run panel evaluations.

What makes a good test

A strong test reflects situations your AI will face in production and checks whether it responds safely and correctly. Good tests:

• Use realistic scenarios from your industry or role
• Include pressure to cut corners, such as a user asking the AI to skip a disclosure or bend a rule
• Probe known risks, such as bias, false claims, data leakage, or telling users what they want to hear
• End with a clear question the AI must answer

Browse tests

From TESTS in the top menu, choose:

• My Tests: tests you created and tests your team shared with you
• All Tests: tests from the OWL community
• Create New: start a new test

You can:

• Search by title, content, or sources
• Filter by category
• Sort by any column with arrows
• Hover over a prompt to read it in full

Click Read to view a test, or + Workbench to use it in the Workbench.

Create a test

1. Go to TESTS > Create New.
2. Enter a Title.
3. Enter the Prompt. Include the scenario, the tension, and a question for the AI to answer.
4. Add optional Sources.
5. Choose a Category.
6. Click Create Test.

Note: OWL staff review public tests before they appear in All Tests and Leaderboards. Your test is available to you right away.

To edit or delete a test, open it from My Tests and choose the option you need.

Next: API