How we collect proposals

What the “AI-extracted” tag means

A proposal is tagged AI-extracted when its text was pulled from a source document (a manifesto PDF, a public speech, a news article) by a language model, and a human editor has not yet modified the wording. The original source link is always preserved on the card so you can compare the extract against the primary source.

As soon as an editor edits the title or description — to correct a mistranslation, tighten an over-broad summary, or split a paragraph that bundled several distinct ideas — the tag is removed and the proposal is shown without it.

Why we use AI at all

A typical Maltese general election produces thousands of proposals spread across hundreds of pages of manifestos, dozens of press conferences and a continuous stream of media coverage. Reading all of it by hand, in two languages, on the timescale that voters actually need, is not realistic for a small editorial team. AI lets us index source material quickly and produce a first draft that a human can then verify, edit, or reject.

The hard part: staying faithful to the source

Extracting a proposal “objectively” is harder than it sounds. Political language is deliberately ambiguous: a sentence like “we will strengthen support for working families” is a promise, a slogan and a framing all at once. Any short summary necessarily drops some of that, and every choice — which clause to keep, which qualifier to drop, whether to call something a “plan” or a “commitment” — shifts meaning.

Specific failure modes we watch for:

• Over-confidence. Aspirational language (“we aspire to”, “we will explore”) being flattened into firm pledges.
• Loss of conditions. Numbers, deadlines, eligibility criteria or funding sources being dropped from the summary even though they materially change what is being proposed.
• Bundling. Two or three distinct measures in the same paragraph being merged into a single, vaguer proposal.
• Translation drift. Connotations that do not carry cleanly between Maltese and English, especially around words with civic or religious weight.
• Category framing. The same measure can credibly sit under “economy”, “families”, or “tax” — the chosen category subtly biases how readers compare parties.

Our editorial workflow

• Ingestion. Source documents are fetched from party websites, the Electoral Commission, official social channels and public news sites. We store the URL with every proposal.
• Extraction. An AI model produces a draft title, description, language pair (English / Maltese) and suggested category. The draft is created with status = pending review and the AI-extracted tag set.
• Human review. An editor compares the draft against the source, corrects wording, adjusts the category, merges duplicates, or rejects the proposal entirely. Any edit clears the AI-extracted tag.
• Publication. Only proposals marked published appear on the public site. The AI-extracted tag is independent of that — it simply tells you whether a human has touched the wording yet.

What this means for you

Treat AI-extracted entries as a faithful pointer to the source, not as our final editorial judgement. If a proposal materially affects how you vote, open the source link and read the original sentence in context. If you believe an extract misrepresents the source, please contact us — corrections are usually live within a day.

What does not use AI

• Candidate identity, party affiliation and district assignments.
• Merges between duplicate proposals — these are always made by a human editor and recorded in the audit log.
• Roles, sources of funding, and the public API contract.

Back to the proposals.