This month, Deloitte Australia agreed to refund the federal government for a report it had been paid to write.
The report, commissioned by the Department of Employment and Workplace Relations, ran to several hundred pages. It also contained citations to academic work that does not exist and a quote attributed to a Federal Court judgment that the judgment never contained. The firm disclosed it had used a generative AI tool chain to help produce the document. The fabrications were caught not by Deloitte, and not by the department, but by an academic at the University of Sydney, Dr Christopher Rudge, who read the thing closely enough to notice the sources were invented. Deloitte agreed to refund the final instalment, worth about AU$440,000. You can read The Register’s account or Fortune’s.
The easy reading of this is a story about one tool, one firm, one embarrassing quarter. It is a price signal.
The cost did not fall, it moved
The promise sold with generative AI was lower cost. Anyone can now produce a competent-looking report, a press release, a board paper, a policy submission, in minutes rather than days. On the production side, that is broadly true. The marginal cost of polished output has collapsed.
What has not collapsed, and has in fact gone up, is the cost of being sure the output is right.
When a junior analyst wrote a weak report, you could usually tell. It read thinly, the structure wandered, the gaps were visible. The poor quality was its own warning label. A fabricated citation produced by a capable model carries no such label. It is formatted correctly, cited in the right house style, sitting in a paragraph that argues its point well. The only way to know it is false is to go and check, source by source, claim by claim.
So the work did not disappear. It moved. It moved from production, where it used to live, to verification, where most organisations have no process at all. That move is the AI credibility tax. Every team now generating output faster is also, whether they have noticed or not, accruing a verification liability they are mostly not paying.
Deloitte is a useful case precisely because Deloitte is not careless. It is a large, sophisticated firm with review processes built over decades. The fabrications got through anyway, into a government deliverable, and stayed there until an outsider read it properly. If it can happen there, the question for a forty-person fintech or a stretched NGO comms team is not whether it could happen to you. It is whether you would catch it before a customer, a regulator, or a journalist did.
Who pays the tax, and how
The tax falls unevenly, and it falls hardest where you can least afford it.
A consumer brand posting cheerful copy can absorb the occasional wrong fact. A fintech cannot. When you are asking people to move money on your word, every published claim is implicitly a claim about your competence with detail. A fabricated statistic in a blog post is not just a blog problem. It tells a prospective customer something about how carefully you handle the things that matter, and they are right to draw the inference. NGOs sit in the same exposed position for a different reason: their entire asset is being believed. A foundation that publishes a fabricated figure in a funding appeal has not made a small editorial error. It has spent down the one thing it cannot easily rebuild.
This is the wider backdrop to the Deloitte story. Through 2025 the same pattern has surfaced across publishing and professional work: retracted articles built on invented sources, a rising tide of fabricated citations passed off as research, a general sense that a lot of what is being published was generated and never checked. The word that has attached itself to all this is slop. The more precise description is output that was produced but never vouched for.
Vouching is the part that matters. It is also the part AI cannot do for you.
A model cannot stand behind anything
A model can generate a claim. It cannot be accountable for it. It has no licence to lose, no reputation that survives the next prompt, no relationship with the reader that a falsehood would damage. Accountability is not a feature you can add to the tool, because accountability is precisely the thing a tool does not have. It belongs to a person, and through that person, to an organisation.
This is why the response to the credibility problem is not better prompting or a more careful model. Those help at the margins. The actual answer is structural. Someone has to own the question of whether a given thing is true before it goes out, with the authority to stop it if the answer is no, and the seniority to be believed when they say so.
Most organisations have quietly let that role decay. The sign-off that used to sit with an experienced editor or a communications director got distributed, then automated, then skipped, because the output looked finished and finished-looking output invites a light touch. The credibility tax is the bill for that skipped step, presented later and with interest.
The honest counterargument is that verification is expensive and slows everything down, and that the whole point of the new tools was speed. That is true, and it is the wrong frame. You are not choosing between fast and slow. You are choosing between paying for verification deliberately, as a designed-in cost, or paying for it by accident, in refunds and retractions and the slow erosion of being taken at your word. Deloitte is now paying the second way. It is the more expensive of the two.
Verification is a leadership decision
The instinct, faced with all this, is to reach for a tool. An AI detector, a fact-checking plugin, a policy document. None of those is the answer, because the problem was never really technical.
Deciding what your organisation will and will not put its name to is a judgement call. It needs someone senior enough to hold the line when a deadline is pushing the other way, who understands both the subject and the stakes, and who can tell the difference between a claim worth checking and one safe to wave through. That is editorial judgement, and it is a leadership function, not a software one. The organisations that come through this period with their credibility intact will be the ones that treated verification as something a senior person owns, not something a plugin handles.
This is the work A&C does. We sit inside a communications function as senior counsel and take responsibility for what goes out: setting the standard for what gets checked, holding the sign-off, and being accountable for the claim, not just the copy. The tools have made producing words nearly free. The scarce thing now is someone whose name on it means something.
If your output has run ahead of your ability to stand behind it, let’s have a conversation.