AI Gremlins
I use AI tools every day across two businesses, and I read a lot of AI-produced content that is terrible in ways the authors do not seem to notice. This site is where I work the whole thing out: an evolving, sometimes messy record of learning how to work with AI rather than against it.
What this site is about
This is not a finished body of work with a thesis to defend. It is an amalgam, and an iterative one: research, argument, frustration, the odd hunch, all of it part of working out how to use these tools well. I change my mind in public here, and the older pieces stay up even when a newer one moves past them.
The name comes from where I started. Every large language model ships with persistent behavioural tics baked so deep that no amount of instruction, custom prompting or stern user feedback can fully kill them. I call them gremlins. Researchers call them excess vocabulary, stylistic artefacts, statistical fingerprints. The effect is the same: your reader can tell. The gremlins are in the weights, not the prompt.
I started writing because I got frustrated beyond belief. I was producing content with AI tools, running it through my own writing protocol, editing line by line, and the output was still being flagged by detection tools. Then I tried degrading the writing to pass detection, and the result was worse than the original. That process became an article. The articles became a series. The series became this: two strands, one preoccupation, the gap between AI used well and AI used badly.
The articles stay close to the evidence: where the gremlins come from, how they behave, what the research says. The thinking-out-loud pieces are looser, where I work out what it is actually like to build and think alongside these tools every day. Both are part of the same process, and the process is the point.
The articles
Each is researched, sourced and written with AI tools while trying to keep the gremlins out. The irony is deliberate.
-
Where did the AI gremlins come from
A short history of AI slop, a field guide to the gremlins, and my position on using AI to write.
-
The gremlins living in AI
The vocabulary fingerprints, structural habits and punctuation tics that AI cannot suppress, no matter what you put in the prompt.
-
The gremlins are eating each other
What happens when AI models train on AI-produced content. Model collapse, knowledge collapse, the farmed salmon problem.
Coming -
I tried to fool the detector and the detector won
The process of humanising AI output, what detection tools actually measure, and why degrading the writing to pass is a losing trade.
-
The protocol is the point
Building a writing protocol that keeps the gremlins out. What it catches, what it misses, and why specificity is the only real defence.
Coming -
What the reader already knows
Your audience is better at spotting AI output than you think. The research on human detection, and what it means for professional credibility.
Coming
Thinking out loud
Less forensic, more reflective. Working-out pieces on building with AI agents day to day, and learning to work with the tools rather than against them.
-
Your AI frustration is a communications gap
Most of us treat AI as something to control, and control is the wrong tool. A note on building with AI agents and trying a different stance.