(Lack of) AI disclosure

I am not hostile to use of AI. I am against arbitrage on expectations. It is wrong to trick people into believing a human wrote or created something where none did.

No AI was used to write the pages on this site. That was me, in a text editor. If you come to this site, you came to read a human being write something, and it would be a betrayal of the reader to instead serve them machine-generated text. And this site attempts to be as factual as possible; large language models will invent new facts that did not exist before. That isn’t acceptable for something that should reflect the historic record to the best of my knowledge. This means the errors are mine, too.

(I know some genius out there will exclaim, “but the em-dashes!” That’s actually a feature of the software I use to compile a bunch of Markdown text files to HTML. Normalise non-ASCII punctuation!)

I do not use AI-generated or AI-enhanced images for most of the same reasons. That’s just not what anyone came here for, and they do not reflect the historic record, so I will not use them.

On the other hand, I have found that AI is brutally effective at proof-reading, and I use it for that on this site.

Until mid-2025, I resisted any kind of computer assistance to correct technical problems with my writing (grammar & spelling). That isn’t because my writing is perfect. It’s really not! But software tools invariably made it worse. They mostly follow rules invented for bad reasons, and hammered out all the individuality and personality from my writing.

The mid-2020s generation of AI is different, because you can tell it specifically not to re-write your text according to made-up rules, and it listens. I use a prompt like this for a command-line tool:¹

Proof-read the file [FILE]. Only correct outright errors, misspellings, poor grammar, and so on. Don’t fix idiomatic or conversational grammar - it is a personal reflection, not an academic paper or newspaper article. Write corrections out to [TEMPORARY FILE] - give a few words of context and what should be fixed.

That finds only genuine syntactical errors in my text, and I manually apply the suggested changes instead of allowing the AI to re-write it. Style-aware proof-reading was just not possible before the large language model era without paid human proof-readers – who often exhibit the same failure modes of pre-LLM software.²

Most importantly, this is not trust arbitrage. Nobody would be upset if they found out I use one of the aforementioned pre-LLM tools; AI is an immeasurably-better way of doing the same thing.

Elsewhere, I do not think it is a betrayal of the reader to use AI assistance for the code that backs this site.³ Nobody came here expecting hand-written HTML and CSS; if that ship ever sailed, it was in the 1990s. So if needs be, I don’t mind using AI assistance with the invisible parts of the code, but as it happens I enjoy building it without. It was my day job for many years and I can do it with my literal eyes closed.

What about research?

The more niche the subject, the worse at research today’s large language models usually are. Many of the facts written on this site have not appeared on the Internet before, and this site has entire pages on one wagon and one semaphore signal, so like, niche. When faced with a paucity of data, LLMs will make something up that sounds plausible instead of saying “dunno mate”. Treat them as a generator of starting points for actual research and not as the research itself.