I write these posts. Of course I ask LLMs for feedback and criticism and wordsmith. However, I’ve been increasingly frustrated by the current flavour of AI writing cringe, and stumbled upon Simon Willison’s tool that highlights LLM cliches: 38 patterns, from “delve” and “testament to” through the tells catalogued in Wikipedia’s Signs of AI writing guide.
I have various levels of AI-ism defence throughout my life and took the chance to consolidate them. Now I have 1 corpus of machine-tell writing patterns, a linter that detects and counts them, and staged wiring into every place where an LLM writes prose for human readers.
What I found
The slop that most vexes me is a slightly different dialect than what Simon’s tells defend against. His corpus targets blog-voice cliches (“here’s the thing”, “delve”, “testament to”). Run against 24,000 words of prose for an automated system I run daily, it produced almost nothing.
Claude (ironically) observed: ‘The house dialect, which Rob named fable-speak, is condensed analyst voice: appositive corrections of the shape “your call, comma, not a detail”, claimed-honesty framings like “the honest read”, whole clauses set in bold, and coined hyphen chains. Four locally-authored patterns now cover it; the measurements that justify each one, and the candidates that were rejected, are in [my local stash] dialect-notes.md.’
At one point I used my old writing (e.g. my thesis from 2001) to find patterns. My writing from pre-LLM era sat between 0.0 and 0.1 counted hits per 100 words everywhere; the slop corpora sit between 0.4 and 5.8. It was interesting seeing what 2001-era Rob prose is now “slop”:
- colon-triple: “described in Figure 11 has three drives: hunger, pain avoidance, and curiosity” – where there really are three things being enumerated.
- stacked-questions: “when to do it? what to do it to? how long to do it? why do it: what will be the future results?” – a rhetorical enumeration structuring a section. Genuinely using a device slop overuses (once, in 30,000 words).
- no-chain: “there are no rules, and no way to ask ‘how close did we get?'” – an actual statement about two absent things, not a “no fluff, no filler” cadence.
- turns-out: “when that explanation (in the form of a Predictor) turns out to be erroneous” – “turns out to be” as a plain verb phrase, not the “Turns out, X” reveal opener.
That last one being a good example how tells resist regex.
Now for the experiment
I tried having Claude write this post, in my voice, from my notes, under the exact system it describes – and run through the detector before publishing. Final score: two counted hits in 960 words, a rate of 0.21 per hundred, no cluster, verdict “advisory”.
I ended up rewriting it. Claude’s hook was fun though – the Spam photo is one I took on our trip to Japan (which is its own side quest full of AI wins I need to write up).
I did like the hook Claude chose though: “Japan didn’t ban the processed stuff. It plated it properly.” Claude ended with: “You tell me if the musubi is any good.”