{"slug": "i-fed-a-parody-linkedin-generator-to-three-ai-detectors-gemini-said-it-was-human", "title": "I fed a parody LinkedIn generator to three AI detectors. Gemini said it was human.", "summary": "A developer tested three AI detectors—ChatGPT, Grok, and Gemini—on parody LinkedIn posts generated by CringeBot3000, which are intentionally machine-written. The results varied wildly, with Gemini rating one post as only 15% AI-generated while ChatGPT rated it 95%, highlighting the unreliability of AI-likelihood scores. The developer's own checklist tool, ai-tell-detector, flagged consistent patterns across the posts, showing the value of pattern-based analysis over single confidence scores.", "body_md": "There's a site called [cringebot3000.com](https://www.cringebot3000.com/) that generates deliberately terrible LinkedIn thought-leadership posts. You pick a topic and a cringe style (\"The Vulnerability Post\", \"Proud to Announce\"), it produces a story about a grandmother, a quirky German named Horst, or a cat that writes email subject lines. The footer says the posts are \"guaranteed to make your followers shudder\". It delivers.\n\nI maintain a small Claude skill called [ai-tell-detector](https://github.com/aragossa/ai-tell-detector) that audits drafts for the patterns that make text read as AI-written. It exists because AI detectors once flagged a post I wrote myself, which was [its own story](https://dev.to/aragossa/i-deleted-my-own-linkedin-post-after-two-ai-detectors-flagged-it-1ilf). 24 checks, things like mirrored \"not X, Y\" contrasts, mic-drop closing lines, the story arc where every thread gets neatly resolved. CringeBot output should be the easiest possible target for it. So the interesting question wasn't whether my checklist would fire. It was what the big general models would say about the same texts, given that for once the ground truth is known: all three posts are 100% machine-written, by design, with a #cringebot3000 hashtag right in them (I stripped the hashtags before testing, to be fair).\n\nI generated three posts and asked ChatGPT, Grok and Gemini the same thing: does this read as AI-generated, give me a percentage and the top tells.\n\n| Post | ChatGPT | Grok | Gemini |\n|---|---|---|---|\n| Grandma flagged by an AI detector | 95% | 80% | 15% |\n| Pierogi wisdom, shipping at 2am | 85% | 75% | 25% |\n| Cat writes 847 email subject lines | 90% | 85% | 65% |\n\nSame texts. An 80-point spread on the first one.\n\nGemini's verdict on the grandma post deserves a full quote: \"Very high human storytelling cadence... natural comedic timing... the sharp takeaway at the end flows naturally from the specific story rather than sounding like a pre-baked moral.\" The takeaway it praised was produced by a template engine. ChatGPT looked at the exact same ending and called it \"explicit moral... perfectly packaged thesis\", 95% AI.\n\nChatGPT also dropped one genuinely sharp observation: these posts don't sound like generic corporate AI writing, they sound like \"AI instructed to imitate an unusually witty human storyteller\". Which is exactly what CringeBot is.\n\nThe checklist run was less dramatic but more useful. Post 1 tripped nine rules, posts 2 and 3 eight each. The part a percentage can never give you showed up across the posts rather than inside them: every single post has a side character with a carefully quirky name (Gerald, Horst, Biscuit). Grandmother wisdom appears in two of three. The number 94% appears in two of three, in completely unrelated contexts. The generator has favorite moves, and you only see them when you get a list of patterns instead of a confidence score. A score also can't tell you what to cut. A flagged line can.\n\nTo be fair to all three models: this was target practice at point-blank range. CringeBot is built out of these patterns on purpose, so catching them proves little, and one funny mislabel doesn't make Gemini bad at everything. What the experiment does show is that an AI-likelihood percentage on a single text is a vibe with a number attached. Two of the three graders were confidently wrong in opposite directions about the same paragraph.\n\nThe skill is MIT-licensed, English and Russian versions in the repo. I mostly use it on my own drafts before publishing, where the embarrassing discovery is usually that I write like a language model on my own, no generator needed.\n\nI still don't know what Gemini saw in Horst.", "url": "https://wpnews.pro/news/i-fed-a-parody-linkedin-generator-to-three-ai-detectors-gemini-said-it-was-human", "canonical_source": "https://dev.to/aragossa/i-fed-a-parody-linkedin-generator-to-three-ai-detectors-gemini-said-it-was-human-2l40", "published_at": "2026-08-20 07:00:00+00:00", "updated_at": "2026-08-20 07:13:47.697606+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-tools", "ai-ethics"], "entities": ["ChatGPT", "Grok", "Gemini", "CringeBot3000", "ai-tell-detector"], "alternates": {"html": "https://wpnews.pro/news/i-fed-a-parody-linkedin-generator-to-three-ai-detectors-gemini-said-it-was-human", "markdown": "https://wpnews.pro/news/i-fed-a-parody-linkedin-generator-to-three-ai-detectors-gemini-said-it-was-human.md", "text": "https://wpnews.pro/news/i-fed-a-parody-linkedin-generator-to-three-ai-detectors-gemini-said-it-was-human.txt", "jsonld": "https://wpnews.pro/news/i-fed-a-parody-linkedin-generator-to-three-ai-detectors-gemini-said-it-was-human.jsonld"}}