How we make each story

What “fact checked” means here

We are not a traditional fact-checking organisation, and we do not rate claims as true or false. When we say a fact is verified, we mean it was reported independently by at least two separate news organisations. That is a useful signal, but it is not proof: outlets can all repeat the same official statement, and that statement can still be wrong.

What we add is the comparison itself: what the outlets agree on, what only one of them reported, and where their accounts conflict.

1. Where the stories come from

Every three hours we collect newly published, publicly available articles from 7 Australian news outlets, owned by 5 separate organisations:

We stopped collecting from News Corp Australia mastheads (including News.com.au) in August 2026, because their sites ask not to be collected automatically. Articles gathered from them before then are no longer shown or counted as a source.

Every story links to the original articles. We encourage readers to read them: the reporting belongs to those newsrooms, and our summaries depend entirely on it.

2. Grouping coverage of the same event

Each article is converted into a numerical representation of its meaning (an “embedding”, produced by Amazon’s Titan model). Articles whose meaning is very close are grouped together as one story, regardless of headline wording or how each outlet framed the event. Groups that turn out to be broad round-ups of several unrelated events, rather than one story, are filtered out because they cannot be compared fairly.

3. Counting independent sources

Two articles from the same company are not two independent sources. So we count owners, not mastheads:

• The Sydney Morning Herald and The Age are both owned by Nine and count as one source. 7News and The West Australian are both Seven West Media and count as one.
• Where articles from different owners are near-identical (typically the same wire-service copy), they are collapsed into one source as well.

A story needs at least two independent sources to exist on this site at all. The number shown on each story (“from 3 independent sources”) is this owner-level count.

4. Writing the summary

Summaries are written by an AI model (Mistral's Ministral 3 language model, run on Amazon Bedrock) from the full text of every article in the group. It works in two passes:

1. Anchor the details. The model first copies out, word for word, every date and number the articles contain. In the second pass it may only use dates and figures from that list, not from its own knowledge. Where outlets give different dates or figures, the difference is reported rather than averaged.
2. Compare and summarise. The model then produces:
   • Verified facts: each must name at least two outlets that explicitly reported it.
   • Points of difference: details only one outlet reported, labelled with that outlet.
   • Where the reporting differs: direct conflicts between outlets, shown side by side and never resolved by us.
   • A consensus summary of 400–600 words covering what happened, the background, who is involved, where coverage diverges, and what is still unresolved, with outlets named where it matters.

The model is instructed to paraphrase rather than copy, to add nothing from outside the articles, and not to speculate or editorialise. If fewer than two articles in a group have usable text, no summary is written.

5. What gets published

Every story is checked against fixed rules before it appears:

• Not published: stories with fewer than two independent sources, summaries under 100 words or with fewer than four verified facts, near-duplicates of an existing story (these redirect to the original), and broad umbrella topics.
• Published and listed in search engines and the archive: stories with three or more independent sources, or two independent sources and at least four articles between them.
• Published but not offered to search engines: stories that pass the minimums but have thinner coverage. They stay readable from links on the site.
• Retired: short summaries (under 300 words) more than 90 days old. We don’t keep the source text long enough to expand them, so they are taken down rather than left thin.

6. What people do, and what they don’t

Stories are produced and published automatically. A person does not review each story before it goes live; with new stories every few hours, that is not something we can promise honestly.

The process is overseen by Ed, Founder & Editor at The AI Playground, who:

• chooses which outlets are included and sets the publishing rules above;
• monitors what the system publishes and tightens the rules when problems appear;
• reviews every reported error and decides on corrections, under our corrections policy.

7. Known limitations

• AI makes mistakes. A summary can misstate or over-simplify what the articles said. Every claim should be checked against the linked sources before you rely on it.
• Agreement is not truth. If every outlet repeats the same mistaken claim, it will appear here as a verified fact.
• Our outlets shape our coverage. Stories these 7 outlets don’t cover, or cover in only one newsroom, won’t appear.
• Timing. A story reflects the articles available when it was last updated. Later reporting or outlet corrections may not be included yet.

Our videos

The short videos on our YouTube channel are made automatically from the same published stories, with the same rules. Their narration uses a synthetic (computer-generated) voice.

Questions or errors

If you think we’ve got something wrong, please tell us how to fix it. For anything else about how the site works, get in touch.