Turnitin recent updates: NMT Fingerprinting findings & potential collaboration

[Acnologia7021]

Hey, we've been doing extensive research on bypassing Turnitin using an n8n chain-translation workflow. We noticed your recent commit about detector limitations. We made our findings open source here: https://github.com/Acnologia7021/turnitin-bypass-research. Would love to collaborate to find a way around the new NMT fingerprinting!

[molly554]

Hi @Acnologia7021 — thanks for the detailed writeup, and for putting the workflow out in the open. We read through the n8n chain properly, and we're open to comparing notes. A few things jumped out that probably explain the regression back to 96–100%:

1. The whole chain runs at temperature: 0.1. This is likely the single biggest lever working against you. Low temperature forces the model toward the highest-probability token at every step, which flattens sentence-length variance — it actively destroys the "burstiness" you correctly identify as the missing ingredient. Deterministic translation converges on the same canonical phrasing every run.

2. Every node translates; none rewrites. All of the LLM nodes are prompted as strict, faithful translators ("preserve proper nouns, output only the result"). Faithful translation swaps the surface lexicon but preserves the deep syntactic structure — and that skeleton is a big part of what detectors actually fingerprint. There's no stage in the pipeline whose job is to restructure rather than translate.

3. The real language path doesn't match the README. The DeepL node labelled "Translate to DE" is set to target_lang: "ES", so the chain you're actually running is EN→ZH→JA→DE→ES→EN, not the EN→ZH→JA→DE→EN you document. Worth pinning down which path your original 0% results came from before tuning anything further.

4. Pure NMT is a fingerprint, not a launder. You're right that DeepL trades an LLM fingerprint for an NMT one. NMT engines optimise for the most probable, standard-register output, so a chain built entirely from faithful-translation steps will drift back toward a detectable distribution no matter how many languages you stack. The lever isn't more hops — it's introducing controlled variance somewhere in the chain.

Rather than hash the details out in a public thread, it's easier to take this to email — reach us at support@lynote.ai and we can compare notes properly. One thing worth covering first: how are you scoring results — a fixed Turnitin/GPTZero test set, or ad-hoc runs? Hard to iterate on any of this without a stable benchmark to measure against.