8 hours testing Pangram: anti-AI-detection tricks all failed; only self-typed drafts pass
PawelHuryn · x · 2026-08-16
The author spent 8 hours experimenting with Pangram, the AI detector Substack built into its platform. Conclusion: the popular "de-AI-ify" tricks do nothing against detectors — the only thing that passes is typing (or dictating) the first draft yourself.
Key experimental findings
- Common advice (kill your em dashes, rewrite in Simplified Technical English) changes style but makes prose read like an IKEA manual — and doesn't fool detectors anyway.
- He rewrote one draft three times, swapping punctuation and words: no impact. He polished an AI draft line by line, keeping his own typos: still flagged, 25% human. Even feeding his own collected ideas, takes, and experiment results to AI got flagged.
- The only thing that passed: writing the first draft himself. AI could then reword it heavily — still 100% human. Dictation counts too.
Three separate problems
- AI slop is defined by readers — they don't run detectors; they feel the texture: buzzword framing ("paradigm shift" five times a post), self-clapping ("Here's the kicker"), artificial enthusiasm, recapping a point two sentences later. Simplified English fixes none of this.
- Detection is about who wrote the first draft — the drafter determines the verdict, regardless of later edits.
- Watermarking is about which model chose the words — Anthropic will watermark future Claude models via a secret key that influences word choice, invisible to readers and verifiable only by key holders. So his "100% human" AI-reworded draft is exactly what a watermark would catch. The only technique that works everywhere is not using AI at all.
Two closing takes: the loudest AI-slop critics are largely people who built their positions on writing better than others; most readers care far less than expected — what matters is whether you had something to say before opening the agent.
More from Models
- Gemini 3.7 Flash debuts at #7 on Vals Index v2 — aronchick · 2026-08-16
- Gemini 3.7 Flash Matches GLM 5.2 in Price and Quality; GLM 5.3 to Be Open-Sourced — zainhas · 2026-08-16
- Qwen4-27B predicted to run Fable-level graphics on laptops — SumitGup · 2026-08-16
- DeepSeek V4 Max Completes Challenge with Just $23 Cost — burny_tech · 2026-08-16
- Grok 4.6 beats GPT-5.6 Sol on coding agent efficiency, 35% lower cost — rohanpaul_ai · 2026-08-16
- Uncensored vision-enabled Qwen3.6-27B finetune trends on Hugging Face with GGUF release — HauhauCS · 2026-08-16