Trust is the product
The brand is trust, so we treat neutrality as something to be built and checked, not claimed. Here is exactly how a summary is made, checked, and published — and why your vote can’t be traced to you.
Each neutral summary is generated by an AI model from the actual text of the bill as introduced to Parliament. The instructions require it to explain only — no opinion, no framing, not left, not right, not centre.
Links to the primary sources — the bill text, its explanatory memorandum and second-reading speeches — are shown on every bill, so you can check the summary against what Parliament actually wrote.
The left-leaning, independent and right-leaning panels are AI-generated political perspectives, clearly labelled as such. They are not factual analysis and do not represent the position of Australia Speaks.
All three are generated by the same model with the same word budget, and each is held to the same grounding standard as the neutral summary — so no perspective gets a smarter, longer or louder voice than another. The symmetry is deliberate, and it is enforced in the software, not left to good intentions.
The symmetry is about process, not merit: every perspective is written and checked under the same rules. It is not a claim that every side is equally persuasive on every bill — that judgement stays where it belongs, with you.
Alongside bills, Australia Speaks covers the political topics dominating the news. Topics are selected automatically from the headlines of a fixed set of outlets spanning the political spectrum — left-leaning press, right-leaning press and the national broadcaster among them. The outlet list is fixed in the method, and we will publish it on this page.
A topic only qualifies if several outlets are covering it, and every explainer links the articles it is based on. Talking Points go through the same neutral-explainer, three-perspective, independent-check and human-review process as a bill — and the model may only use facts the cited reporting actually contains.
Every summary and perspective is cross-checked by a second, independent AI model — a different model from the one that wrote it — for neutrality, balance and grounding. It rejects specifics that the source does not support.
Then a human editor reviews it. Nothing is published without a person approving all four pieces. No auto-publishing, no exceptions. When we are not confident a piece is neutral and grounded, it does not go out.
Your vote is anonymous — structurally
Votes are stored with no user identifier of any kind. There is no column anywhere linking a person to a choice — so even a full copy of our data could not reveal how you voted. This is enforced in the schema and in the code that records every vote, and it is guarded by automated tests that fail the build if anyone tries to change it.
We record only that you voted on a bill — so you can’t vote twice — and your state or territory, so results can be shown by region. Your own choice is kept only on your device. Anonymity here isn’t a promise we could quietly break; it’s the shape of the system.
Tallies show the votes of Australia Speaks users — people who chose to download the app and vote. They’re a real record of considered views, but not a scientific poll, and we’ll never present them as “what Australia thinks”.
When we get something wrong, we say so. Anything corrected or retracted after publication stays on the public corrections log, permanently. You can flag any summary or perspective as unfair from inside the app; every flag goes to the human review queue.
No advertising, no partisan sponsorship and no money from political parties — ever. Australia Speaks is funded by optional subscriptions. No investor or partner has any influence over content.
No comment threads or reply wars. No streaks, points or badges that reward voting more. No feed tuned to your politics — everyone sees the same items and all three takes. Nothing that could corrupt the honesty of the tally.
This is version 1.0 of our public method, kept under version control. Any change to how content is made or checked is a public, dated change to this standard — not a quiet edit.
That risk is exactly why we don’t rely on one model’s good behaviour. A second, independent model checks the first for balance and grounding, a human approves every piece before it publishes, and we publish our method and our corrections so you can judge us on the record rather than on our word.
No. The database has no column that links a person to a choice, so even a complete copy of our data could not reveal your vote. Your own choice is stored only on your device. We record that you voted on a bill (so you can’t vote twice) and your state or territory (so results can be shown by region) — nothing more.
No, and we will never say they do. Tallies show the votes of Australia Speaks users — people who chose to download the app and vote. That is a real record of real people’s considered views, but it is not a scientific poll.
Australia Speaks takes no advertising, no partisan sponsorship and no money from political parties — ever. It is funded by optional subscriptions. No investor or partner has any influence over content.
We’d rather you check us than believe us. Join the early-access list and be there when Australia Speaks opens.
