AI usage in this blog

There seems to be a kind of dividing line among bloggers. Some blogs make it obvious that they lean heavily on machine content generators. The human being more prompt-giver than author. Others give the impression that their position consists of nothing but blanket opposition. That’s the spectrum. I think I owe you a note on where in it I place myself.

TL;DR: Somewhere in between.

Transparency, and why it’s needed

Realistically, slop is unavoidable once you spend any time on a social network¹. Slop is everywhere. I have no problem with slop as such. Some slop is good enough (or at least entertaining enough) to pass as a form of expression in its own right². It’s not unlike a greasy burger and fries: certainly not healthy, but sometimes it just fills the hole in your stomach.

I also think that machine content generation is getting a lot of stories told whose tellers previously lacked the skills to tell them to an audience. That doesn’t make any of those stories less worth telling.

I can’t draw or paint to save my life. Even though many a situation would work beautifully as a comic. Why don’t I use machine content generation for that? Why a picture, when I can just as well write a thousand words instead?

I only want to know beforehand when I’m looking at an LLM-generated text. And not just when the text is sloppy, but whenever it mainly came out of an LLM. That’s the transparency I ask for when I’m spending my lifetime consuming this content. I want to be able to choose machine-generated content actively — knowing full well that I may be taking in slop. The same way I actively choose to eat a burger.³ Or to watch the sixteen-thousand-and-twenty-ninth episode of General Hospital.

But if I’m the one asking for this, I can hardly be opaque about it myself. So: here’s my position, which doubles as a disclaimer for my own content.

Incompatibility

There’s a very simple reason why I don’t stand on the side of the spectrum that uses LLMs for everything. I’ve written before that I write for writing’s sake. That’s why I write the way I write. But “writing for writing’s sake” is fundamentally incompatible with using an LLM. It would be like taking up oil painting to relax by telling a painter what to put on the canvas.

Workflow

The texts here are written by me. That’s also why the output fluctuates so much. Stretches of almost manic productivity follow quieter phases. It isn’t mania, though. It’s a fairly reliable indicator of how well I’ve been sleeping.

A text usually starts as an idea in a notebook (no, not an app — the paper kind, to be used with some sort of writing implement. I’m partial to Leuchtturm’s A4 notebooks) and then becomes a fully drafted article in Word. For the spell check.⁴ I’ve noticed that the grammar and spelling of my articles have improved considerably since I brought my blogging workflow in line with the one I use for professional writing during the day. I’m a child of the red and blue squiggly lines, it turns out.

That’s also where my en dashes come from. In German en dashes are not optional and have a distinct use case stipulated in the German orthography rules. Unlike in English where the usage of em dashes (the equivalent of en dashes in German) is a decision. The em dash seems to be a choice that is rare enough to be suspicious. In German you don’t get a choice. Thus many word processors in German have a rule that enforces them. The problem starts where correctness is used as a marker for LLM use. I’m not sure those people using this heuristic spend much time in word processors, especially those set to German. Since in some quarters an accusation of slop, or of LLM assistance, can carry real weight, I’d appreciate it if those people also checked their instrument for false positives and false negatives. Not every text with dashes is an LLM, and not every text without them was written without one.⁵

You could have a fine argument here about whether spell checking in Word already counts as the frowned-upon automated text generation, or whether it’s still acceptable. Honestly, I assume that spell checkers in word processors will be local LLMs before long. And then it’ll get genuinely hard to maintain any pure doctrine on this.

That I don’t use LLMs for the urtext (the text that i’m typing into the keyboard, in this context the finalized german version of it) doesn’t mean I don’t use LLMs at all. A few weeks ago I decided to run this blog bilingually from now on. Historically it has had a fairly international readership. Between reader feedback and some quiet signals in my logfiles, I suspected the texts were being handed to some translation tool or other fairly often. Bilingualism at the necessary quality, however, isn’t sustainable if I do the translating by hand. That became clear to me while translating the “The Day My Heart Stood Still” series. Somewhere in the middle of it I started writing the translation with increasing LLM support.

By now I have a settled workflow for it. Since German is obviously my native language, the text is written in German first (the mentioned “urtext”). The English version is then produced with LLM support.

Limits

Which is where I run into the limits of machine translation. I recently used the German mnemonic “333 – bei Issos Keilerei” in a text — a rhyme that fixes the date of the Battle of Issus. The LLM translated it dutifully, and at high quality. Except the line doesn’t work in English. So the English text usually needs substantial rework afterwards. For instance, turning 333 into “In fourteen hundred ninety-two, Columbus sailed the ocean blue.”⁶

I also don’t always recognise myself in the automatically translated texts. So a manual pass over the translation follows anyway. And after that, one more run through the LLM to fix the errors I introduced during that pass.

Both of these are why I now provide the translation myself. I can’t control what becomes of my article when some arbitrary tool is handed nothing but the URL. I can control what I put on my own blog. And with LLM support, the time it takes has become acceptable.

Transparency in practice

I decided to be open about this. So I built myself a Liquid tag in my blogging software that outputs a disclaimer stating the text was produced with LLM assistance. It goes at the top of the translated text.

On an article translated from a German source text, this is the note you’ll find at the top: