Sovereign AI for the Australian Newsroom

Three layers of control. First, the model never publishes — every output is a draft for an editor, sub-editor or in-house lawyer to review. Second, the model is fine-tuned on the masthead’s own correction history, prior apologies and the relevant Australian defamation case law (the post-amendment uniform Defamation Acts, the public-interest defence under section 29A, the headline cases including Wilson v Bauer Media and Lehrmann v Network Ten), so it understands the actual standard the publisher is held to in Australia. Third, the workflow includes an explicit risk-screen step on contested draft material that surfaces imputations for the duty lawyer in plain language. None of this removes the responsibility of the in-house counsel, but it materially reduces the volume of obvious issues that reach them and gives them better-quality drafts to review.

Yes, and this is one of the highest-value workflows in a daily newsroom. The model can ingest hours of interview, court or parliamentary audio in a controlled environment, produce a diarised transcript with speaker attribution, then answer plain-language questions about the content with citation back to the timestamp. A court reporter can ask "what did the witness say about the second meeting" and get the answer with the verbatim quote and the timecode in seconds. The source audio never leaves the publisher’s infrastructure. The reporter still writes the story; the AI removes the hours that would otherwise be spent skimming the audio.

The journalism exemption in section 7B(4) of the Privacy Act 1988 applies to the publisher’s own handling of personal information for journalism — it does not extend to third-party AI providers the publisher might disclose source material to. That is the critical reason a custom LLM matters in this industry: by keeping all source material inside infrastructure the publisher controls, the journalism-exemption posture is preserved end-to-end. A public AI tool, by contrast, can be argued to be a third-party disclosure that takes the activity outside the exemption. We can produce the architectural documentation the publisher’s in-house counsel needs to confirm this in your specific context.

Yes. The model sits as a retrieval and reasoning layer on top of the systems the newsroom already operates — WordPress VIP, Brightspot, in-house CMS variants, Méthode, Newscycle, Avid iNews and ENPS for broadcast — using each system’s standard API. There is no migration of editorial content and no requirement for the newsroom to change its day-to-day workflow. Reporters and sub-editors interact with the model through whichever surface the newsroom prefers (chat, integrated CMS plugin, or a stand-alone newsroom-research interface) and the drafts flow into the existing CMS publication workflow.

Source-identifying material is handled under tighter access controls than general editorial material. Specifically: role-aware access so source-identifying notes are limited to the small group of people who actually need them; an immutable audit log of every interaction with source-identifying material; and an architectural guarantee that no third party — not the model provider, not the cloud provider, not any downstream service — retains, processes or has technical access to the material. For the most sensitive material (investigative pieces with high-risk sources) a fully air-gapped on-premises deployment is supported, with no external network connectivity at all.

Technically yes, but that is not the editorial pattern we recommend or that any serious masthead would deploy. AI-generated content published without human editing has produced predictable category errors (factually wrong, defamatory, off-style) across every publisher that has attempted it globally. The right pattern is the AI accelerates the journalist — research, transcript Q&A, archive recall, style-fluent first drafts of intros and standfirsts, fact-checking, headline variants — and a human journalist, sub-editor and (where relevant) lawyer remain in the publication chain. Used that way, the model adds capacity to the newsroom without compromising the masthead’s editorial standing.

A typical newsroom deployment runs eight to twelve weeks: two to three weeks of editorial, legal and archive scoping, four to six weeks of archive ingestion and style fine-tuning on the masthead voice, then a pilot on a single desk for a defined period (typically four weeks). The duty lawyer and chief of staff stay close to the rollout so editorial or legal issues surface immediately. Reporters on the pilot desk are doing useful archive retrieval and transcript Q&A inside the pilot window. Full newsroom rollout follows once the editor and in-house counsel are satisfied with the pilot outcomes.