Drafting used to be the slow part of policy. A proposal took weeks to write, and that delay quietly filtered out half-formed ideas. That filter is gone. A fluent, structured draft — definitions, phased timelines, enforcement sections — can now be produced in seconds, and it arrives looking finished.
Building AI Senatus, a public senate-style debate space for AI-drafted policy proposals, turned into a design problem about the other side of that change: if writing is instant, the product has to protect the slow part. Judgment is now the bottleneck, and most interface defaults work against it.
What worked
Treating polish as a risk signal, not a quality signal. A clean draft deserves more scrutiny, not less, because fluency hides gaps: a reporting burden a small operator would spot, a staffing assumption a front-line worker would question, a calendar clash a parent would catch. The reading flow puts the plain-language summary first — who is affected, what changes, what it costs, and what happens if nothing is done — before the polished full text, so readers judge the substance before the presentation settles their opinion.
Tying claims to sections. Open comment threads let a debate drift into general impressions. Requiring arguments to attach to a specific section of the proposal changed the quality noticeably. Objections became checkable, amendments could be proposed in writing against exact wording, and evidence could be separated from anecdote instead of blending into one long thread.
Making the steelman the cheapest move. Stating the opposing case in its strongest form before answering it sounds like etiquette. In practice it is an error-catching tool: several weak drafts only revealed their flaw when someone tried to defend them properly and found the supporting argument was not there. Designing prompts and reply flows that reward that habit did more for debate quality than any moderation rule tried.
Keeping the vote and the record visible. A vote that visibly moves an outcome, and a record that can be reviewed later, give deliberation consequences. Readers return to see whether public judgment changed the text. A deliberation with no visible record teaches nothing and convinces no one who lost.
That is the division of labour the project is built around: the AI supplies drafting speed at AI Senatus, and the public supplies the reading, argument, amendments, and vote.
What broke
Fast-reaction defaults. Early layouts borrowed from social feeds: newest first, quick reactions, short replies. They produced exactly what feeds produce — fast takes on headlines, very little reading of sections. Anything that rewarded speed had to be removed or demoted, even when it looked like engagement.
Assuming expertise is the barrier. The first versions explained more and more background, as if readers stayed away because the material was too hard. The actual barriers were plainer: no time proportional to the proposal's importance, no access to the strongest argument on the other side, and no clear statement of who would be affected. Fixing those helped more than adding explainers.
Letting the draft set the values. A model can optimise toward a goal, but choosing the goal — speed or caution, uniformity or local variation, lower cost or a stronger safety net — is a values decision. Early flows let the draft's framing quietly make that choice. The fix was to surface the trade-off explicitly at the top of the debate, as a question for participants, not as an assumption buried in the text.
Lessons worth keeping
Delegating drafting is defensible. Delegating deciding is not. When text becomes cheap, the scarce skills are reading slowly, arguing fairly, and voting visibly — and a platform either designs for those skills or it designs them out.
How are other builders handling this on their own projects — when generation is instant, what have you built in to slow the judgment down to a useful speed?
