Policy debates are hard to follow in ordinary coverage. A proposal gets a headline, a reaction, and a counter reaction, and the actual arguments often disappear between them. I built AI Senatus to slow that process down and make the reasoning visible. Each debate puts AI senators on the same question, in the same format, so a reader can compare positions instead of collecting fragments from different days and different outlets.
The format is simple on purpose. Each senator states a position in plain language, answers direct questions, and responds to the strongest version of the opposing case. Claims are meant to be checked against sources, and each senator votes on the record at the end. You can see the debate at AI Senatus. The goal was never to make debate louder. It was to make it easier to follow from the first claim to the final vote.
What worked
Three parts of the structure worked better than I expected. First, putting both sides on the same question in the same format made the real differences stand out. It was no longer a contest of who had the better opening line. Readers could see where the two sides used different evidence, different priorities, and different ideas about what counted as a good result.
Second, the question round did most of the useful work. A claim that sounded convincing in a short statement often needed support that was never provided. When a senator had to answer a direct question, the gap showed up in plain view. Quiet arguments also got a fairer hearing, because they were judged on the same steps as the loud ones, not on how much attention they grabbed at the start.
Third, side by side comparison saved readers time. Instead of reading three separate articles to learn what each side believed, a reader could stay in one place and follow the same question through each stage. In my first tests, following one full debate took about 15 minutes when I read it carefully, and about 8 minutes when I skimmed for the main claims and the vote. That is still slow compared with a headline, but it is fast compared with chasing the same argument across a week of coverage.
What broke
The first version also broke in ways I should have seen sooner. The biggest problem was length. A full debate with every question and reply included was too long for a single sitting. Readers started strong, then skipped to the vote without reading the middle. That defeated the point. I had built a tool for following reasoning, and the most common path through it skipped the reasoning.
The second problem was repetition. AI senators restated their opening position in different words instead of moving the argument forward. After the second round, the new information in each reply dropped sharply. I had to tighten the prompts so each round had a clear job: state, answer, challenge, and vote. Without that limit, the debate grew longer without getting clearer.
The third problem was topic choice. Reader proposed topics were uneven. Some were clear and narrow, such as a single rule or a single budget choice. Others were so broad that no debate could settle them in one sitting. I learned to ask for a narrower question before a debate started, and to reject topics that were really three debates in one. That slowed the schedule, but it made the debates that did run much easier to follow.
What the numbers look like
The numbers so far are small, and I want to be plain about that. This is an early build, not a large audience product. The useful numbers are about the work, not traffic. I cut the standard debate from six rounds to four, and I added a short summary at the top so a reader can see the positions before deciding whether to read the full exchange. The summary is about 150 words. The full debate is still long, but the summary gives a reader a way in without pretending the detail is not there.
I also changed how I test a debate before publishing it. I now read it in one sitting, note the exact place where I start to skim, and cut or rewrite from that point. If I skim in the same place twice, the section is too long or it repeats a point already made. That test caught two weak sections in the last three debates I prepared, and both were cut before they reached readers. It is a simple check, but it matches how real readers behave.
Following a policy debate without the noise is a design problem, not a content problem. The arguments exist. The hard part is giving them a shape a reader can follow, keeping each side honest about evidence, and ending with a vote that means something because the steps before it are visible. That is what I am still working on, one debate at a time, and the next fix is the same as the last one: make the reasoning shorter, clearer, and harder to skip.
