AI-generated insights

Some charts and report summaries on the Health Equity Tracker include AI-generated text. These insights are produced by a large language model. They are not written by our team and are not research findings.

What the model receives

When an insight is generated, the rendered contents of the chart or table (the topic, geography, demographic breakdown, and data rows) are sent to the model. The model has no access to anything else on the site and is not searching the web.

Limitations

Language models can restate what a chart shows fluently while saying nothing that the chart does not already say, and they can state incorrect values with confidence. Specific rates and group comparisons should be read directly from the chart, which is the authoritative source on the page. Insights describe correlation in published data and do not establish cause.

Data flow and third-party processor

Chart contents are sent to Google's Gemini API to generate insights. The request originates from our server, not your browser, so the provider sees our server's address rather than yours. Only public aggregate data that the site already publishes is included; no personally identifiable information is sent.

The model family is Gemini. The specific version may change over time as newer models become available.

Reporting a harmful or inaccurate insight

Each insight includes a way to report it as harmful or inaccurate. Submitting a report records the text and the reason for team review. After a report is submitted, a fresh insight is generated in its place so the chart remains usable. Reports are our primary signal for catching outputs that need attention, and we review them regularly.