No (humans) read the damn bills

Will generative AI lead to a golden age of government transparency? Elon seems to think so.

Elon Musk

In the coming weeks, Grok will summarize these mammoth laws before they are passed by Congress, so you know what their real purpose is

The idea is seductive. It’s an open secret that lawmakers generally don’t write let alone often read the bills they vote on (or even supposedly “authored”). With zillions of bills and many pieces of legislation running in the thousands of pages, actually reading those bills exceeds the capacity of any single human (barring perhaps a Dune mentat or Tyler Cowen type).

LLMs have no such limitation on length, particularly the latest models like Gemini, which are a big deal for expanding inference. Such tools offer anyone the ability to interrogate and ask questions of arcane dense documents like legislation and other policies. It’s easy these days to dream of an automated services that scans every bill introduced in the CA legislature and sends a personalized summary of how that would affect a given type of organization, group or individual.

Or more immediately, such tools enable the ability to ask questions about what a bill means for different groups. For example, I asked Chat GPT what is in AB 1668 / SB 606, a piece of legislation I’m familiar with from when I helped launch the CaDC. It does a pretty good job of getting the gist.

But Chat GPT misses the mark on nuances like what the gallons per capita per day means for CA households. The regulations apply to local water utilities, not actual households, a nuance that’s lost in the following reply.

The response also ignores outdoor water usage, a critical part of AB 1668 / SB 606. Let’s consider another timely example — the new bill raising minimum wage requirements with a big carve out that’s seemly custom-sized for Panera, a company whose CEO just so happens to be chummy with and a big donor for Governor Newsom.

Joe Mathews

@joemmathews

"I inherited the bill" gives game away. Californians shud understand this. Your lawmakers often don't write their own bills. Interests do.... Now Chris is using this reality to defend a carve out for Panera on bill he "authored."

If only we columnists could use this defense...

Ashley Zavala

@ZavalaA

NEW: Assemblyman Chris Holden, author of California’s fast food worker law, says he did NOT negotiate it in its final stages in 2022, when Panera/bakery exemption was added.

He tried to say he inherited the language. But he was the author when carve out was added. Watch:

Here is what Chat GPT had to say about the bill:

Notably Chat GPT does not flag the exemption for baking bread on premises unless explicitly prompted:

This type of test makes me think (for now at least) LLMs will not necessarily lead to a golden age of legislative transparency. I do see the tools being extremely useful in helping to augment a bill analysis. It is also easy to see such tools lowering the cost of writing bills and increasing their volume still further.

What does this mean for political power? Let’s consider a few possibilities.

LLM’s lower the barrier to entry for legislative analysis, democratizing access to actionable information on what potential legislation means. A win for outsiders.
LLM’s muddy the waters of what bills do. Any analysis vastly increases the amount of information and thus the ratio of noise in the system. Thus personal relationships, context and closed door maneuvering become even more valuable. A win for insiders.
LLM's lower the cost of writing comment letters and other advocacy tools to influence the legislative process. This one is a bit of a muddle, as it'll make it easier to flood comments with nonsense, potentially making staff and legislators more agnostic about such tools.
LLM's lower the cost of actually authoring bills, empowering and enabling individuals and small groups to advocate for and enact legislation, potentially through direct democracy channels. A win for outsiders.

Regardless, I’m excited to mess around and find out.