If you work in government relations, you’ve probably already tried using ChatGPT or another large language model to research a bill. And if you have, you’ve probably already run into the problem.
I want to walk through exactly what that problem is, why it happens, and how we designed AskAI to address it.
Why General-Purpose AI Gets Legislation Wrong
Here’s a concrete example. Earlier this month I pulled up California AB 1389 in ChatGPT and asked what the bill was about. It told me: horse racing.
That answer was correct — about three weeks ago. On March 24, the bill was amended in the Senate. The horse racing language was struck entirely. The bill is now about tribal gaming. Two completely different policy areas.
ChatGPT got it wrong not because it’s a bad tool. It got it wrong because of how it works. It queries the broader internet on a crawl schedule. When something changes in a bill, the web takes time to catch up. By the time ChatGPT refreshes its understanding of that bill, it may already be law.
This is the core limitation: general-purpose AI is trained on the internet, and the internet is always a few steps behind active legislation.
We built AskAI, FastDemocracy’s built-in legislative research chat tool, specifically to handle this.
AskAI doesn’t search the internet. It pulls from only primary sources from FastDemocracy’s database of bill text, fiscal notes, committee documents, and meeting transcripts, and uses those documents as the basis for every answer it generates. The primary source comes first. The language model’s job is to turn that source material into a clear, accurate response. That order of operations matters more than most people realize.
Amendments Are Where It Gets Dangerous
The horse racing example is relatively harmless. But consider what happens when you’re tracking a bill that directly affects your organization, your clients, or your advocacy priorities.
Amendments don’t just update bills. They sometimes reverse them entirely. A provision that was in a bill last week may have been deleted. Language that seemed favorable may have been replaced with something that cuts the other way. If your AI tool doesn’t understand insertions and deletions in bill text, it will give you a confident answer based on a version of reality that no longer exists.
AskAI tracks what has been added and what has been removed at the bill text level, and makes sure that distinction is passed to the language model before it generates a response. AskAI isn’t guessing from the internet. It’s working from the current, primary source document, with full context about what changed and when.
RAG Is the Right Architecture for This Problem
The technical approach we use is called Retrieval-Augmented Generation, or RAG. The idea is straightforward: instead of asking an LLM to answer from memory, you retrieve the relevant primary sources first and give those to the model as context.
For legislative tracking, that means bill text, fiscal notes, committee documents, and meeting transcripts. When someone asks AskAI a question, it pulls the current version of the relevant documents from our database and passes them to the language model along with the question. The model is then constrained to answer based only on those sources.
If it doesn’t have enough context to answer reliably, it says so. That might sound like a limitation, but it’s actually one of the most important features we built. An AI that tells you “I don’t have enough information” is far more useful than one that confidently gives you the wrong answer.
The Bias Question
Something that comes up often when we talk about AI in policy contexts: bias. And it’s a legitimate concern.
If you ask a general-purpose LLM to summarize a bill, it might characterize the legislation in ways that reflect a point of view. Words like “streamline” or “expand access” carry connotations. One person’s streamlining is another person’s deregulation.
We handle this by prompting AskAI to operate strictly as a legislative assistant. It is instructed not to introduce opinion or characterization unless explicitly asked. When a user does ask for arguments for and against a bill, AskAI draws those arguments from people who have actually testified on the bill, not from its own synthesis. The tool surfaces what has been said, not what it thinks.
What This Means in Practice
The difference between these approaches becomes clearest when legislation is moving fast, which is most of the time during session.
Bills get amended. Hearings happen on short notice. A committee vote can change the trajectory of an issue your organization has been tracking for months. If your AI tool is working from outdated information or generating responses without grounding in primary sources, you are making decisions based on something that may no longer be true.
Legislative tracking is one of the domains where accuracy is not optional. The cost of acting on wrong information is real.
Follow FastDemocracy on LinkedIn for more on how we’re building AI tools for government relations professionals, and for updates on new features including 50 state bill tracking and our recently launched local government tracking in St. Louis, Chicago, New York City, Baltimore, and Denver.



