Implementing SynthID-Text From Scratch: Watermarking Doesn't Hurt Output Quality
ell-hol1 · reddit · 2026-10-26
Frustrated by the noise around LLM text watermarking, the author read Google DeepMind's SynthID-Text paper, implemented a naive version on their own TinyStories-pretrained model, and wrote it up:
- Watermarking doesn't necessarily degrade quality: SynthID-Text has distortionary and non-distortionary settings; the latter preserves the original token distribution.
- Counterintuitive testing: DeepMind's 20M-response Gemini experiment showed watermarked outputs got slightly more positive user feedback — not statistically significant, but the opposite of the feared quality drop.
- Why it matters: as AI slop floods the web, distinguishing human from machine text deserves more attention than reflexive dismissal.
The article covers the probability theory, the sampling algorithm, and a working from-scratch implementation.
More from Research
- UniForge3D turns sparse-view inputs into detailed 3D meshes, code released — mervenoyann · 2026-10-26
- neuron-explorer: open-source steering-first neuron explorer for Qwen3.5-4B — viksit · 2026-10-26
- 30% of US STEM PhD dissertations now contain AI-generated text, NBER paper finds — 0xkarasy · 2026-10-26
- None of the 11 grand challenges in single-cell science solved since 2020 — Sauers_ · 2026-10-26
- Emotion neurons causally steer how LLMs read ambiguous news — Sauers_ · 2026-10-26
- MemTensor's MemT-1.0-Nano trends on Hugging Face, a memory-tuned Qwen3.6-35B-A3B finetune — MemTensor · 2026-10-26