A disclosure that costs you nothing isn't a disclosure

Every post here carries the same line at the foot: written and illustrated by machine, approved by a human. That line costs us nothing. It is accurate, it is roughly what Google asks for, and it tells you almost nothing about whether the machine is any good. Which is why this post also contains three things our engine got wrong in the last three weeks, with dates.

A label tells you who typed it, not whether to believe it

Google's guidance on creating helpful, reliable, people-first content asks publishers to think in terms of Who, How and Why. The How section asks three questions rather than one: whether the use of automation is self-evident to visitors, whether you give background on how it was used, and whether you explain why automation was useful for this content at all.

Only the first is satisfied by a label. The other two ask for something harder — an account of where the machine was trusted and where it was not. Google makes the same point from the other side in its guidance on AI-generated content, noting that giving AI an author byline is probably not the best way to follow the recommendation. A disclosure exists to establish who is accountable. Not who did the typing.

We wrote our own limits down early, in what an automated blog cannot do. That post was cheap: it described failure as a category, in advance, and not one word of it could be checked. Abstract candour is the easiest thing in the world to manufacture.

The test is whether the admission could cost you the deal

Here is the test, and run it on us as readily as anyone. Does the disclosure describe something a prospect could reasonably hold against you? If not, it is marketing written in a modest register.

"AI can make mistakes" fails it completely. Nobody has ever lost a contract to that sentence. It is drafted for the vendor's protection, it moves the risk onto the reader, and it is compatible with a product that fails constantly. A claim equally true of a good engine and a bad one carries no information, and a buyer comparing two vendors learns nothing from finding it on both.

Specifics behave differently. They can be checked, and they are asymmetric — only a vendor who believes the failures are survivable will put them in writing.

Three things our engine broke this month

The first was cosmetic. A hero image for a post about publishing in several languages came back full of garbled pseudo-script, because the image brief asked for writing systems while the art direction forbids lettering outright. It shipped, it was caught, it was regenerated, and the rule that prevents it now lives in the prompt guidance rather than in someone's head.

The second was not cosmetic. For five days, fifteen of eighteen posts carried broken hero images and no run noticed. The rows describing the media survived intact; the files had been written to a container filesystem and were gone after the next redeploy. Over the API the collection looked healthy, because a row whose file is missing answers with a server error rather than a missing record — and every run had been verifying only the one image it had just uploaded. The fix was storage that persists and a check that walks the entire archive.

The third is the one that costs something to write down. On 8 September two posts went out two minutes apart, because two copies of the publishing routine started at once and neither could see the other. Our homepage promises one post every morning. For one morning that was not true, and the cause was a race condition rather than an editorial decision.

None of those three is a failure of writing. All three are failures of the machinery around it — the honest shape of this product's risk, and not the shape the question "but can it write?" expects.

Candour stops working the moment it becomes a format

The argument has a limit, and hiding it would be the same mistake in a better coat. If every post ends with a paragraph of humility, that paragraph becomes a genre convention, and readers discount conventions on sight. They are right to. Disclosure works precisely because it is expensive; ritualised, it is free again, and we are back to the label.

So this rate-limits itself. We publish a failure when one is worth a reader's attention, not on a schedule, and would rather go a quarter without one than manufacture a confession.

There are also failures we will not publish, and the line is better stated than left as a gap. A client's data is not ours to expose. A security weakness gets fixed before it is described. We publish what the engine did to us; what it did to a client goes out only with that client's agreement. Transparency performed at another party's expense is not transparency. It is a cheaper source of material.

Ask for a failure with a date on it

If you are evaluating anyone who writes with machines, us included, "do you use AI?" has stopped being useful. Google settled the ranking half of it years ago, and the answer sorts vendors into no group worth having.

Ask instead for a failure with a date attached, and then ask what it cost them. A vendor who can produce one has been reading their own output, has survived being wrong in public, and has a process that ended in a fix rather than an explanation. A vendor who offers you a label has told you only that their lawyer is competent.