BREW: Block-wise Reliable Embedding for Multi-bit Text Watermarking
> Paper: Block-wise Codeword Embedding for Reliable Multi-bit Text Watermarking > Authors: Joeun Kim, HoEun Kim, Dongsup Kim, Young-Sik Kim > arXiv: 2605.00348 | 2026-04-29
The Problem: Watermarks Can Be Embedded, But Not Reliably Decoded
Why watermark AI-generated text?
- Identify AI-generated content
- Prevent misuse
- Copyright protection
- Content provenance
- Embed many bits
- But reliability suffers
- Detection and decoding get conflated
- Catastrophic false-positive rates
- Difficult threshold setting
- Raising the threshold → detection rate collapses
- Effectively random-guess level
- Multi-bit watermarks = more information stored
- But reliability degrades
- At detection time, you can't tell "watermarked but decoded wrong" from "no watermark at all"
- Stage 1: Block-level detection — confirm the text is watermarked with high reliability
- Stage 2: Specified verification — decode only on confirmed blocks, avoiding false positives
- Text is divided into blocks
- Each block is embedded independently
- Local verification, global aggregation
- Not blind decoding of all text
- First confirm "a watermark exists"
- Then "decode the payload"
- Detection and decoding are separated
- Low false-positive rate
- High detection rate
- Multi-bit capacity
- All three achieved simultaneously
- Traditional methods = receiving a letter and immediately reading its contents
- But the letter may contain no hidden message — you "read" something that doesn't exist (false positives)
- BREW = first check "does this letter have invisible ink?"
- If yes → read the message
- If no → don't read
- This dramatically reduces false positives
- Detection and decoding are conflated
- You decode without knowing whether a watermark exists
- You may "decode" information that isn't there → very high false-positive rates
- The threshold dilemma: raising the threshold reduces false positives but also tanks detection — a lose-lose situation
- Two separate stages: detect first (highly reliable), then decode (only when watermarked) → extremely low false positives
- Improved reliability: sensitive detection, no false alarms, reliable even for multi-bit payloads
- Practicality: suitable for real deployment; users can trust detection results
- Detection ≠ decoding
- Confirm existence first
- Extract content second
Limitations of existing methods:
Capacity-focused approaches:
Problems with ECC-based extractors:
The core issue:
The BREW Approach
The paper proposes BREW (Block-wise Reliable Embedding for Watermarking):
Core idea: > Shift from "detect-then-decode" to "specified verification" — a two-stage mechanism that ensures multi-bit watermark reliability.
Technical design:
1. Two-stage mechanism
2. Block-wise embedding
3. Specified verification
4. Reliability guarantees
An analogy:
Why "Specified Verification" Beats "Blind Decoding"
The problem with blind decoding:
Advantages of specified verification:
A Signal-Processing Principle: Detection ≠ Decoding
> Trying to decode a watermark from all text is like trying to "hear music" in all noise — you'll hear many things that don't exist. BREW's insight: first ask "is there music here?", then ask "what is the music?". Separating these two questions is the key to reliability.
This reflects a fundamental signal-processing principle:
Takeaways
If you're building watermarking or content-provenance systems, ask yourself:
1. Does my watermarking system conflate detection with decoding? 2. Is the false-positive rate acceptable? 3. Does multi-bit capacity sacrifice reliability? 4. Could a two-stage design improve reliability?
BREW reminds us: a reliable watermark is not about how much information you can embed, but how certain you are at detection time.
When a watermark system learns to "confirm first, decode second," it evolves from an information hider into a trustworthy content passport. In the future of AI content governance, the best watermark isn't the most invisible — it's the most reliable.
In an ocean of information, a lighthouse of certainty is worth more than a sail of capacity.