Civic applications of generative AI

17 Jul, 2026

Recently I pointed Fable at my town's Web site and asked it to scrape transcripts of recent school board meetings, answer some questions I had about them, and preserve those transcripts as a stable artifact (so I can keep adding to, and asking about, this record).

The experience was striking in a few ways. From less to more surprising:

1. Fable gave me reasonable-looking results with minimal hassle.
2. Those results are, as far as I can tell (including cross-checking and follow-up questioning), extremely accurate and useful.
3. The differences between the best models and even slightly outdated ones are very important in this kind of project.

Specifically, here are things that Fable did that seemed to me much better than the results I would have gotten a month or two ago:

1. It found the relevant subset of archived videos without asking a lot of questions or complaining about inconsistent URL formatting.
2. It cleaned up errors in automated transcripts.
3. It handled proper names quite well, including in figuring out when the automated transcript system had translated a person's name into different strings of text at different times.
4. It researched recent local news¹ and figured out which parts of the meetings had to do with which other events. (This sometimes required reading between the lines.)
5. It applied basic arithmetical and economic reasoning to various arguments.
6. It reasonably integrated basic contextual information (Is a claim made by a board member or a visiting parent? Is this meeting happening during the school year or at the beginning of summer vacation?).

For a computer system to get some stuff like this right occasionally is, at least in 2026, not new. That level of quality and consistency would also be nowhere near good enough to get actually useful outputs in this project, which required the system to get tons of stuff like that correct, with a very high accuracy rate. (It doesn't take many misconceptions, translation errors, or straight-up mistakes in a project like this to make its results useless or worse.)

Local politics can be tough to understand in a way that often reminds me of security through obscurity. Even when people want to communicate with you, the terms in which they're speaking are often only semi-familiar (at best). The aggregation of all the essential contextual information is so often just a bit too much for a layperson to keep in mind, especially if they can't constantly interrupt to ask clarifying questions. (I try to stay informed, but sometimes need reminders of, e.g., what precisely the different tiers are in the Massachusetts system of educational support, and you can't really understand some of these arguments if you don't know what's going on with the tiers and state policy. Again, this is just one of many fragments of context.)

The newest models, though, are very accurate, very good at getting through drudgery, and (still) infinitely patient with my questions. There's a large set of civic questions that are now practically accessible to me in a whole new way. (And this generalizes beyond local politics.)

#generative AI #lifestyle