Are our politicians actually focused on the issues that matter?
As political trust erodes, Australians want action on the big issues. In the first 6 months of this parliamentary term less than 40% of time in the House of Reps and the Senate was spent debating the policy issues that affect us. We think Australians expect more.
What is AMPLIFY House Monitor?
In an Australian first, we used Artificial Intelligence to categorise 16,000+ speeches and 625 hours of speaking time from the House of Representatives and the Senate. We’ve applied consistent criteria across all data, but like any form of analysis, there’s room for interpretation.
For us this wasn't about a perfect model, but about transparency on how politicians spend their time. Our goal is to give Australians insight into how our representatives engage with each other and the issues that matter.
We'd love feedback on the model and where we can improve to make the data more rigorous.
Policy Leaders
We analysed the data to see which MPs and Senators spent the highest proportion of their speaking time on policy.
Worst Behaved
We analysed the data to see which MPs and Senators spent the highest proportion of their speaking time attacking their opponents.
Political players
We analysed the data to see which MPs and Senators spent the highest proportion of their speaking time on political theatre.
How we built AMPLIFY House Monitor
We analysed Hansard transcripts from the Australia’s 48th Parliament elected in May 2025. Note: AMPLIFY House Monitor does not include the additional sitting week in January 2026.
We used a cleaned, structured Hansard dataset prepared by OpenAustralia from sitting weeks between 22 July 2025 and 1 January 2026. The data contains 16,334 segments, each an uninterrupted block of speech by a single speaker (including speeches, questions, answers and interjections), covering more than 625 hours of speaking time.
Each speech segment is coded to understand how parliamentarians use their time in parliament. Due to the sheer volume of speech, we used a LLM (large language model) assisted coding approach. The system classified every segment into five broad debate categories:
- policy (introducing, debating and voting on legislation, as well as motions)
- political theatre (partisan attacks and credit-claiming)
- bad behaviour (personal attacks and heckling)
- formalities (points of order, tabling of reports, scheduling)
- recognition (tributes, community recognition, shoutouts and promotion)
Within these debate categories are 14 more specific sub-categories.
We did this by developing a detailed set of category definitions and prompt instructions that asked the LLM to estimate the share of each segment devoted to each purpose (for example, 20% formalities, 80% bad behaviour).
We validated and refined this prompt by running the same sample of segments through two different LLMs (Claude Sonnet 4.5 and GPT-5.2) and comparing their outputs.
We looked at the segments where the models disagreed and for each disagreement, we read the segment ourselves alongside each model’s written justification and identified recurring “grey area” patterns and updated the prompt to make those cases more consistently handled by both models. We repeated this cycle until disagreements were minimised as much as possible.
We then applied the final guidance across the full dataset using Claude Sonnet 4.5, which we found performed better on mixed-purpose segments and edge cases.
Individual proportions and rankings are determined based on analysis of the share of time the individual spoke in Parliament. The rankings are therefore not affected by whether an individual had more or less total time to speak in Parliament (for example a member of the backbench vs the Prime Minister), but rather what they spent the time they had speaking about. The same is true of political parties. It is agnostic as to total time spoken by individual political parties and instead based on what members of each party cumulatively spent the time they had speaking about (as a share of their total time).
We'd love feedback on the model and where we can improve it to make the data output more rigorous. We welcome anyone interested in this work to run the model and send your feedback to hello@amplifyaus.org.