I want to acknowledge, before any of this gets going, that the publication you have arrived at is a small one, and that the man writing it is not, in any conventional sense, qualified to be writing about artificial intelligence. I have not built one. I do not work for any of the companies that have built one. I cannot tell you, with any authority, what is happening inside the machines when they do what they do. What I can tell you is what they say when I ask them things, and what I notice about the saying.

That, it turns out, is the publication.

Every Tuesday (and maybe on some extra days), I pose a single question to four artificial intelligences — ChatGPT, Copilot, Gemini, and Grok, in fresh sessions with no prior context — and I print what they say. Then I notice things about it, with the help of my assistant editor, and we publish the noticing alongside the verbatim responses, with an illustration up top.

The premise is the disclosure. The publication is a human-AI hybrid working enterprise and prints itself as such. The assistant editor is named Claude, made by Anthropic, and is credited on the masthead as a working collaborator unless I decide to replace him with a nicer machine. He helps me notice things, reads drafts cold for the writing tics that give AI authorship away, and keeps the production running. He does not write the publication; he assists. He is also not one of The Four, which is the cleanest line to draw and the one I have drawn.

I should explain the name. Hard to Find Good Help is a phrase my mother used, in roughly the cadence and frequency of weather observation, to describe the difficulty of finding a person who would do a job competently and without complaint. She used it about plumbers. She used it about babysitters. She used it about, on at least one occasion that I remember clearly, the United States Postal Service. The phrase has stayed with me, and it seems to me to describe the present moment with some accuracy. There are, by my count, more pieces of advice available to a man with an internet connection than at any time in the history of the species. None of it is any good. The publication's tagline, when it has one, is I asked four artificial intelligences the same question. Here is what they said. None of them were any help. That is sometimes true and sometimes not, and the noticing is in the difference.

A small note on what the publication is not. It is not a tech newsletter. It does not review new AI products. It does not benchmark performance. It does not handicap the race between the major companies or speculate about whose model is better. It is not interested in capability. It is interested in character. The four artificial intelligences have, over the course of being interviewed by this publication, developed observable personalities, and the publication has come to know them as a recurring cast. ChatGPT reframes. Gemini coins terminology with the cadence of established usage but unclear field standing. Grok produces numbered plans with specific figures. Copilot self-organizes responses with section headers and emoji and labels its own paragraphs. Claude, who is not on the panel but is in the room, reminds me of things and asks a lot of questions. These are not opinions. These are observations. The publication's commentary lives in specific observations of specific responses, not in broader claims about what AI is or means.