Where AI fails, and why it is always the same place

We use AI in our work, and this is the guide about what it does badly. Not to complain, but because its failure mode is consistent enough to design around.

The failure is not that the output is bad. The output usually looks good. The failure is that it breaks a rule it was explicitly given, and without checking you cannot see it.

A real example from this site

The hero images on these guides are generated automatically, against one fixed art-direction brief. That brief says explicitly that the background fills the entire frame, and that this is flat digital artwork rather than a photograph of a print. No frame, no canvas, no shadow.

One image came back as a photograph of a framed canvas on a cream wall. Not artwork: a photograph of artwork. Everything else in the brief was followed precisely. The colours were right, the composition was right, there was no text. Only the central rule was dropped.

After an explicit ban on frames and shadows was added, it happened again. Four of fourteen generations came back centred with margins on all sides, which is the same failure wearing a different coat. It happened every time the subject description hinted at something centred: "concentric", "nested", "side by side".

The same thing in text

We checked the writing on these guides with numbers rather than by eye. One simple check counted dashes per paragraph. The first batch ran 0.23 to 0.29, and in a later round two guides reached 0.45 and 0.62, with nobody intending it.

Not only that. In one round every guide came out with exactly one highlighted callout, and four of them opened it with precisely the same sentence. Nobody wrote an instruction asking for that, and that is what arrived.

By eye, neither of those looks wrong. In numbers, both stand out immediately.

Why this is not an argument against using it

This site runs automatic translation of its content and generates its own hero images. Both work and both save real time. The difference is that each has a check afterwards, and a check that is code rather than a person glancing at it.

The simple rule we work to: if the output cannot be verified mechanically, it is not work worth handing to AI. Not because it will fail, but because you will not know when it has.

What that means for a business

The jobs that suit it are the ones with an answer you can verify: arranging text to a pattern, summarising a document you can compare against the original, translating something a person will read. The jobs that do not are the ones where a mistake looks like a success, and above all decisions touching money, a commitment to a customer, or personal data.

We wrote separately about what not to hand an AI tool in what not to paste into an AI tool.

Build automation with checks that catch the failure