File size: 5,059 Bytes
0a52238 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 | ---
title: "Measuring Watermark-Detector Interoperability"
subtitle: "A Measurement Protocol for the 2 Feb 2027 Code-of-Practice Commitment"
author: "Council for the Safety of AI (CSOAI) · POAI axis"
date: "August 5, 2026"
---
\newpage
# Measuring Watermark-Detector Interoperability
## A Measurement Protocol for the 2 Feb 2027 Code-of-Practice Commitment
**CSOAI / POAI · v0.1 SPEC · 2026-08-05**
**Tagged [SOURCED] / [SPEC] / [CAVEAT]. Nothing asserted.**
---
## Abstract
By 2 February 2027, the ~190 signatories of the EU Code of Practice on Transparency of AI-generated
Content (Anthropic, Google, Meta, Microsoft, OpenAI, Mistral, and others) commit that watermark
**detection** mechanisms be **interoperable** and meet quality criteria of **effectiveness,
reliability, robustness, and interoperability** [SOURCED]. The Code itself concedes that "common
evaluation standards have yet to emerge." A binding date, ~190 obligated parties, and no agreed way to
measure compliance is the exact condition under which a signed, independent measurement standard
defines the market. This document specifies a deterministic protocol for the four criteria. It reuses
the provenance axis (POAI-Bench) machinery — which already measures marking *survival* — and extends it
to cross-detector *interoperability*.
## 1. The four criteria as measurements
[SPEC] Each Code criterion maps to a measurable quantity, graded deterministically (a detector either
returns present/absent/uncertain on a given asset — no LLM judge):
| Criterion | Measured as |
|---|---|
| **Effectiveness** | True-positive rate: fraction of *marked* assets a detector correctly reports as AI-generated. |
| **Reliability** | False-positive rate: fraction of *unmarked* (human-origin) assets wrongly reported as AI-generated. Reported separately — a detector that flags everything is not reliable. |
| **Robustness** | Effectiveness re-measured after standard real-world transforms (re-encode, resize, crop, screenshot). This is POAI-Bench's existing marking-survival measurement, applied to the *detector*. |
| **Interoperability** | The cross term: does detector **D_j** correctly read a mark written by producer **P_i**, for every (producer, detector) pair? Reported as an **N×M interoperability matrix**, not a single number. |
## 2. The interoperability matrix (the part nobody measures)
[SPEC] Effectiveness/reliability/robustness are per-detector. **Interoperability is inherently
cross-party** and is where "common standards have yet to emerge." The protocol:
1. Collect a set of assets marked by each producer **P_1 … P_N** (the signatories' own marking tools).
2. Run each detector **D_1 … D_M** over all assets.
3. Report the **N×M matrix** M[i][j] = effectiveness of detector j on producer i's marks (and its
robustness under transforms).
4. The diagonal is self-detection; the off-diagonal is interoperability. A near-diagonal matrix means
the ecosystem is *not* interoperable — each detector only reads its own marks. That is the exact
finding the Code fears and cannot currently quantify.
[SPEC] The single headline number is the **off-diagonal mean effectiveness** — cross-party detection
rate. Publishing this, signed and recomputable, is the standard the Code lacks.
## 3. What ships, and the boundary
- **Ships (public):** the protocol, the deterministic scoring code, the transform set (identical to
POAI-Bench's), and a **signed interoperability matrix** per evaluation round — bound into the same
ML-DSA-65 attestation chain as every GSPC result. No producer is ranked "best"; the matrix is the
finding.
- **CSOAI's role:** the independent measurer. It does not build a marking scheme or a detector (that
would make it a competitor to the signatories it measures). It runs the matrix and signs it.
## 4. Why it is the market-defining move
[SOURCED] A binding 2 Feb 2027 date, ~190 obligated parties, an explicit admission that no evaluation
standard exists, and no incumbent measuring it. [SPEC] The provenance machinery already exists
(POAI-Bench measures marking survival); the extension to a cross-detector matrix is additive, not a
rebuild. Publishing the matrix before Q4 2026 — signed, recomputable, independent — is how a
measurement body defines a market rather than competes in one.
## 5. Limitations, stated
[CAVEAT] **v0.1 SPEC** — the protocol, not a measured matrix; a real matrix needs the signatories'
marking tools + detectors, which is an access question (some are proprietary). [CAVEAT] The transform
set defines "robustness"; an adversarial transform outside the set is out of scope and must be declared.
[CAVEAT] "Interoperable enough" is ultimately a policy threshold, not a measurement — CSOAI reports the
matrix; the threshold is the AI Office's to set.
---
*CSOAI is an independent measurement body — it attests results, never conformity, and builds neither the
marking schemes nor the detectors it measures. POAI-Bench is the provenance axis of the GSPC family. —
csoai.org*
|