A school library asks an AI assistant to chart its borrowing records. January: 200 loans. February: 240. Two tidy bars suggest a cheerful headline: borrowing is up 20%.
Then someone asks, “Does that hold for fiction?”
In this invented dataset, fiction loans fell from 100 to 90. Comics rose from 100 to 150. The total was correct, but it answered only the first question. A reader needs a way to examine the parts.
Try the second question
Change the category below. The scale stays fixed so that switching views does not quietly change the visual comparison.
INVENTED LIBRARY DATA
What changed between two months?
Loans · fixed scale from 0 to 250
All categories: 200 → 240 loans. Up 20%.
See every source number
| Category | January | February |
|---|---|---|
| Fiction | 100 | 90 |
| Comics | 100 | 150 |
| Total | 200 | 240 |
The chart still cannot tell us why comics increased, or whether more individual children borrowed books. We supplied neither causes nor borrower counts. Clicking a filter reveals a breakdown; it does not create missing evidence.
What Flint adds
Microsoft Research’s October 22 article about Flint-interactive describes a tool for giving AI-generated charts this kind of exploratory behavior. Its examples include selecting countries in one chart and highlighting the same countries in another, or connecting a table selection to a chart.
The useful idea is to describe what an interaction means. “Select this country” is more useful to an application than “the pointer landed at these screen coordinates.” Flint handles the connection between the visible marks and the data, while the author specifies what should change after a selection.
Interactive charts already exist: Vega-Lite’s documentation, for example, describes selections and controls that can filter data or alter its presentation. Flint’s contribution is a higher-level authoring approach for agents and developers. Microsoft says this reduces implementation difficulty; this article does not independently measure that benefit.
Can you use it today?
The public project dates Flint 1.0 and its interactive features to October 19, before the blog announcement. It offers a software library and a connector for compatible AI assistants. Setup is required. Its connector documentation makes clear that the embedded interactive view depends on the host supporting that interface. This is not a feature automatically switched on in every chatbot.
Ask for an inspectable chart
A useful request would be: “Chart monthly borrowing. Let me filter by category, keep the scale fixed, show the selected category, and include the underlying counts.”
Then test a case you can calculate yourself. Here, choosing Fiction should show 100 and 90, with a 10% decline. Returning to All categories should restore 200 and 240. If a label changes but the numbers do not, the interaction is broken.
Also ask what was counted. These are loans, not unique readers, reading time or books finished. The most valuable follow-up may be discovering that the dataset cannot answer the question you actually care about.