Skip to content
Voto de Verdade

Beta version. Voto de Verdade is under development: it may contain errors and change without notice. When in doubt, check the Chamber's official text.

See our goals and how the site works

SupportPTEN

AI transparency

The subject and the summary of the 902 bills on the site are written by artificial intelligence, and nobody reviews these texts before they go live.

Updated on October 9, 2026. Here is what it does, how its work is checked and where it can be wrong. How the site works

What the artificial intelligence does

  • Chooses the subject of each bill.
  • Writes the title and the summary in plain language and lists the key points.
  • Says how sure it is and warns when it has doubts.

It does not record or change votes: each deputy's vote, the date and the result come straight from the Chamber's Open Data. It also does not grade, rank or compare deputies or parties.

How it is done

  1. A model (Claude Sonnet 5.5) reads the official text of the bill, chooses the subject, writes the summary and says how sure it is.
  2. A second model (Claude Haiku 4.5) chooses the subject without seeing the first one's answer. If the two disagree, confidence in the subject drops.
  3. A check on the words of the official summary (ementa) confirms that the subject makes sense.
  4. The second model checks the summary against the original text. If it finds parts with no support in the text, the bill shows a warning. If the summary does not hold up, it is discarded and the screen shows only the official summary (ementa).

Last text generated on October 7, 2026. The site only redoes the summary of new bills or bills whose official text changed.

Confidence in each text

Each subject and each summary gets a confidence level calculated by a machine, not by people. High means the automatic checks agreed, not that the text is correct.

Subject. High: 841 (93.2%) · Medium: 35 (3.9%) · Low: 26 (2.9%)

Summary. High: 855 (94.9%) · Medium: 46 (5.1%) · Low: 0 (0%) · No summary: 1

View as a table of the confidence levels

Warnings the site shows

  • 61 bills with a warning that the subject may be wrong.
  • 46 bills with a warning that the summary may contain errors.
  • 1 bill with no AI summary, because it could not summarize the text or the check found no support for it in the text. The screen shows only the official summary (ementa).
  • 539 bills whose vote was on a substitute text or amendment: the text voted on may differ from the official summary (ementa).
  • 17 votes with a warning that the classification was made automatically and may be wrong.

Limits and corrections

  • Artificial intelligence can be wrong, even when it says it is sure.
  • The official text of each bill is always next to the summary. When in doubt, the official text prevails.
  • For now the site has no way to request corrections.