---
title: "Readers can't identify watermarked AI text"
url: https://daily.dev/posts/readers-can-t-identify-watermarked-ai-text-me1g7n4eb
source_url: https://seangoedecke.com/readers-cant-identify-watermarked-ai-text
type: article
source: "sean goedecke"
published: 2026-08-21T07:30:19.228Z
updated: 2026-08-21T07:30:43.474Z
tags: ["llm", "qwen"]
reading_time: 4
upvotes: 0
comments: 0
language: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Readers can't identify watermarked AI text

**[sean goedecke](https://daily.dev/sources/seangoedecke)** · 4 min read · 0 upvotes · 0 comments

## Summary

An informal experiment tested whether readers can distinguish AI text watermarked with Google's SynthID-Text from unwatermarked text generated by Qwen3-30B-A3B-Instruct-2507. After an initial run revealed a flawed quiz design (the watermarked answer was consistently option A, letting straight-A guessers score artificially high), a reshuffled version with 73 participants produced a mean score of 3.4/10, essentially matching random chance (3.33/10). The conclusion: readers cannot reliably identify watermarked AI text, supporting the argument that watermarking doesn't meaningfully degrade output quality or make AI text detectable by inspection. The quiz itself was built as a quick 'vibe coded' static site, and the setup cost about two dollars in rented H200 GPU time.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://seangoedecke.com/readers-cant-identify-watermarked-ai-text>

## Questions this post answers

### Can people actually tell the difference between watermarked and unwatermarked AI text?

No, an informal quiz asking 278 then 73 participants to pick the watermarked response among three answers to the same prompt produced scores close to random guessing (3.4 out of 10 versus an expected 3.33 from pure chance). The text was generated with Qwen3-30B-A3B-Instruct-2507 and watermarked using Google's SynthID-Text, suggesting watermarking does not create a detectable quality difference.

_Following how AI watermarking affects output quality and detectability is easier with daily.dev._

### What is SynthID-Text and how does it watermark AI generated text without changing quality?

SynthID-Text is Google's watermarking method that biases which of several equally likely next tokens a language model picks, rather than forcing lower-quality choices. Since models already randomly select among top candidate tokens, watermarking simply replaces that randomness with a predictable-but-still-random bias, such as picking a token based on a hash of prior tokens, leaving output quality unaffected.

_daily.dev helps developers track how watermarking schemes like SynthID-Text actually work under the hood._

## Similar posts on daily.dev

- [Text AI watermarks will always be trivial to remove](https://daily.dev/posts/text-ai-watermarks-will-always-be-trivial-to-remove-borbz7x9u) · sean goedecke · 6 upvotes · 4 comments

---

Tags: [#llm](https://daily.dev/tags/llm), [#qwen](https://daily.dev/tags/qwen)

[View this post on daily.dev](https://daily.dev/posts/readers-can-t-identify-watermarked-ai-text-me1g7n4eb)