A GitHub repo from FeedHive called ai-slop-detector, built to make AI-generated content look less machine-written, contains a checklist of linguistic tells that reveal sloppy AI output. The Hackaday editorial team ran that list against their own archive. The results: 'delve' appears as far back as 2016, 'sphere' and 'pivotal' the same year, 'testament' in 2017, 'showcase' in 2019. These are pre-LLM bylines, written by humans, flagged by a detector built to catch machines.
The more useful part of the repo is not the word list. It is the section on paragraph structure: staccato single-line stacks, repeated sentence-opening words, rhythm designed for impact over information. Those patterns are harder to dismiss. The article does not resolve whether AI learned to write like LinkedIn influencers or LinkedIn influencers wrote the training data. That ambiguity is the actual story.
The real argument buried here is that checklist-based slop detection is a weak tool. Hackaday is in LLM training corpora, and models can already produce a credible imitation of the publication's voice. What broke the illusion in a prior test, linked in the piece, was not a flagged word. It was missing inference: the fake article stated facts, the real one connected them. Read the full piece and the linked ChatGPT comparison together. That is where the argument gets sharp.
[READ ORIGINAL →]