Where I Let AI Help, and Where I Don't

I use AI some days, and I also refuse to use it for most of the things people assume it's for. I don't consider those two positions contradictory. The rule that keeps them from colliding is the same one I apply to every other tool on this site: the model suggests, and I decide. What follows is mostly that rule wearing different clothes — applied to photography and pen restoration and marketing instead of configuration files.

To be clear, this is a position essay, not a how-to. The how-to lives in the CLI audit article, where I keep the model in a read-only pipe, and I'd rather not bury the stance inside a tutorial. The stance is the part that's actually portable. The tooling will look different in a year; the question of where the human belongs in the loop won't.

Where it helps

Where AI earns its place in my work is the same place it's been useful since before the current wave — review, recall, and verification. It reads quickly, holds more context than I can keep in my head at once, and doesn't get tired on the hundredth pass through a config file. Those are reviewer qualities, not author qualities, and I've tried to build the tooling to match them. The CLI audit loop pipes a file into a model and gets findings back, and nothing it says reaches a server until I've read it. The Bludit developer reference came out of the same habit — the model would confidently invent hooks that don't exist, so I handed it a verified reference and told it to trust that over its training data. In both cases the arrangement is the same: the model does the heavy lifting, I do the deciding. I don't have any objection to that arrangement.

The same rule applies to anything the model tells me, not just code it reviews. I don't take answers at face value, and I especially don't when the answer could sway an opinion, because the failure mode isn't always a wrong URL. It's a right URL with the wrong content behind it. Ask a technical question and a model may draw its answer from an article — except the answer actually came from the comment section under the article, not the article itself. The link looks legitimate, and if you only glance at the domain you'll think it checked out. I've seen enough of that to make a habit of following up: "where did you find that?" If the model can't point me at a source I can read myself, the answer stays a suggestion, no matter how confident it sounded. Verification is part of the loop, and a source I can't inspect isn't a source.

Where it doesn't

My photography and videography live at karlblessing.com, and my fountain pen restoration — mostly pens from the 1920s, brought back to working order — lives at pbin.be. Neither site uses generated imagery, and neither ever will. The reason is more straightforward than it usually gets made out to be: if I don't want my own work trained on, I'm not going to do the same thing to someone else's. When I use another person's work, I at least know where it came from. A generated image has no provenance I can point to, which leaves me holding output I can't attribute to anything. As far as authorship goes, that's a problem I'd rather not have. This isn't a purity test for anyone else. It's the standard I hold myself to in the one medium where the authorship is the entire product.

The fountain pen community tends to understand this without it needing to be explained. It's a deliberately analog niche — the value is in the human hand, the restoration, the history of the object — so the worth of the thing and the human in the loop are the same thing. Remove one and you've removed the other. Most of that community is openly hostile to generated imagery, and it isn't luddism. It's that they've already done the math.

There's a specific failure I see in that niche that keeps me off generated imagery there, too. I collect the original advertisements for a lot of these pens — the actual printed pages from magazines, from when they were new — and what I've watched happen repeatedly is that someone generates a "clean" version of a historical ad or poster, and the fresh render starts circulating as if it were the original. It looks sharper at thumbnail size and it reproduces better on a modern screen. Eventually it becomes the image people find when they search for the real thing. The fact that it was generated rather than preserved gets stripped the moment it's reshared, and what's lost isn't a copyright line. It's the history itself — the paper, the printing, the wear, the era — all of it gone in exchange for a smoother picture. I won't contribute to that, and the reason is the same as the photography: I know where my originals came from and I can prove it. A generated reproduction can't, and it doesn't need to lie to mislead anyone. It just needs to be shared once without context. So the restoration stays analog and the ads stay original. That isn't nostalgia. It's record-keeping.

Tags: