AI PRACTICES10 June 2026
The Human Stays Responsible
AI drafts, summarises and suggests, but accountability never transfers. On oversight, primary sources and the defamation stakes.

When a claim goes out under your organisation's name, nobody asks which tool drafted it. Not the member reading the email, not the journalist checking the quote, not the court, if it comes to that. Responsibility is one of the few things AI genuinely cannot automate, and the organisations using these tools well have stopped trying. They've arrived at a simpler settlement: the machine drafts, and a human answers for it.
Judgment at every consequential step
The working principle is that human judgment belongs at every step where the outcome matters: a person decides what to ask, a person reads what comes back, a person edits it, a person approves it, and that person — not the software — remains accountable. Four kinds of work sit permanently inside that perimeter. Public-facing work, because credibility is spent in public. Legally significant work, because obligations don't attach to a model. Politically sensitive work, because a machine has no feel for what a sentence costs. And relational work, because the member who wrote to you deserves an answer from someone, not something.
None of this is a case against the tools. It's a case about where they sit. A draft is a proposal. Judgment is what turns a proposal into an act.
The hierarchy of sources
Verification needs a spine, and the spine is a hierarchy: primary sources outrank anything a model says about them. What an MP actually said is in Hansard, not in a summary of Hansard. What a donor actually gave is in the AEC's disclosure returns. What a council actually decided is in its minutes; what a court actually found is in the judgment. A model's account of any of these is a lead — often an excellent lead — but it is testimony from a witness who was never in the room, delivered with the confidence of one who was.
The practical rule follows: any fact that will carry weight in public traces back to a primary document, or it goes.
Grounded tools and open generation
Not all AI output stands the same distance from the truth. A source-grounded tool — one that answers only from documents you've supplied, pointing to where in them each claim comes from — is a different instrument from open generation, where the model composes from everything and nothing. For research work, grounding is worth seeking out: an answer pinned to page twelve of an actual report can be checked in a minute. But a citation is an address, not a verification. Models can cite real documents for claims those documents don't quite make. The citation tells you where to look; the looking is still yours.
Verification as a working practice
The difference between organisations that get burned and those that don't isn't scepticism as a mood — everyone claims that. It's verification as a practice: a named person responsible for checking, a habit of tracing every load-bearing fact before publication, and a standing rule that fluency is not evidence. Model output reads confident at every level of accuracy; that's a property of the writing, not the knowledge. The tell isn't in the prose, so the check can't be a vibe. It has to be a step — as unskippable as the one where someone approves the spend.
What's at stake here in particular
Australia gives this discipline a sharp edge. Our defamation law is friendlier to plaintiffs than most of the English-speaking world, and an organisation that publishes a damaging claim about an opponent — a rival candidate, a company, a lobby group — owns that claim entirely. 'The model said it' is not a defence, and honest belief without the checking to back it isn't much of one either. The scenario writes itself: a deadline, a confident summary of an opponent's record, one detail the model composed rather than found, and a concerns notice in the mail a fortnight later.
The applied version of this — how a communications team builds checking into its drafting pipeline — is its own subject, and we've written about it separately. The principle underneath doesn't change with the workflow: the tools can do more of the work than ever, and none of the answering. The human stays responsible, because there was never anyone else.
SEE WHAT PROGRESS CAN DO IN YOUR CAMPAIGN
Book a demo and we'll show Progress working inside one of your own campaigns.
