Purpose of data review
During the testing of Benchling Intelligence beta features, we may request permission to review the inputs provided to the AI system and the outputs it generates. This review process is essential for understanding the AI system's behavior in various scenarios and identifying areas for improvement. For example, we might find that the AI system struggles with certain phrasing or question patterns. Reviewing specific examples enables us to address these issues more effectively.
How will data be stored?
All reviewed data is securely stored within our standard infrastructure (AWS) and is subject to advanced security and privacy measures, including stringent access controls. This data is not shared with third parties. Our review is limited to the data sent to and from the AI; we do not access any other information.
Types of data reviewed
The type of data sent to the AI system can vary depending on the feature being used and may change over time. Here is a brief overview of the data sent for each feature:
- Guided Search: The AI system receives the user's search query along with additional details such as tenant schemas, dropdown menus, and user information. It currently does not include data about specific entities or records.
- Notebook Check: The AI system receives the notebook entry and configured review guidelines.
- Notebook Writer: The AI system receives the user's prompt as well as the full contents of the notebook entry.
- Report Generation: The AI system recieves all notebook entries related to the report, whether they are entries selected by the user or all entries from a selected study.
- SQL Assistant: The AI system receives warehouse information (schemas, field name mappings, some sampled data), your SQL prompt and any configured parameters.
Use of collected data
Our goal with this review process is to enhance the AI system's performance and ensure it meets our standards and your expectations. A human may read through the inputs and outputs to the AI, and make notes of general patterns (for example, "Frequently misunderstands questions about container contents", "Uses entity field filters reliably", etc). These notes will not contain any customer-specific data, and will be used to guide the team on what kinds of questions and answers need to be improved. The data will not be used for AI training.