Watermarks Without Verification: AI Text Watermarking After the EU AI Act

URL SCAN: Watermarks Without Verification: AI Text Watermarking After the EU AI Act
FIRST LINE: Computer Science > Computers and Society

The Dissection

This paper is not really about watermarking. It is an audit of the black box surrounding watermarking: vendor claims, inaccessible deployed systems, weak independent testing, and regulation that demands detectability without guaranteeing anyone can verify it.

Its empirical core is narrow but damaging. On prose, the open-source watermark’s effect is no larger than changing the sampling seed. On code, one model loses three points of correctness, another shows no measurable loss, while detection remains near chance. The paper therefore exposes a compliance technology whose quality costs and practical detection limits cannot be independently established for the systems that matter most.

The Core Fallacy

The paper correctly identifies unverifiability as a governance failure, but it treats better verification as if it could make the underlying system governable. Under the Discontinuity Thesis, that is an epistemic solution to an economic rupture.

Matched outputs, configuration disclosure, accredited audits, shared protocols, and interoperable detectors may make vendor claims falsifiable. They do not stop AI from severing the employment–wage–consumption circuit, nor do they restore productive participation. Perfectly verified AI text is still automated output. The paper risks mistaking a cleaner label for preserved human agency.

Hidden Assumptions

• Regulators can compel meaningful access from providers whose competitive advantage depends on secrecy.
• Accredited auditors and shared protocols will remain independent, competent, and globally interoperable.
• A watermark will survive editing, paraphrasing, translation, model mixing, and adversarial removal.
• Results from open-weight models will generalize to proprietary deployed systems.
• Detecting AI origin creates actionable institutional leverage rather than merely documenting substitution after the fact.
• EU rules can coordinate a global model ecosystem despite open-weight releases and providers outside EU jurisdiction.
• Provenance remains the central problem once synthetic text becomes ubiquitous and economically dominant.

The abstract itself also contains a methodological warning: the deployed systems cannot be tested publicly. The strongest regulatory claims therefore remain assertions about an inaccessible machine.

Social Function

Primary classification: transition management and partial truth.

The paper punctures vendor reassurance and gives regulators an actionable checklist. That is real forensic value. But it also compresses a civilizational transition into an auditability deficit. The institutional response becomes: disclose more, certify more, standardize more. This is how bureaucracies metabolize technological displacement—convert loss of control into a standards program and call the resulting paperwork governance.

The watermark is a lag defense: legal and institutional friction around an already advancing substitution process. It may help assign provenance, enforce disclosure, or support limited accountability. It cannot create a durable human-only economic domain, especially when code detection is already near chance in the tested setting.

The Verdict

A strong diagnosis of regulatory opacity, with a remedy that is structurally subordinate to the crisis it describes. Watermarking is not a shield for the post-WWII order; it is a tracking label attached to its replacement. Even a robust, independently verified watermark would preserve attribution—not mass employment, bargaining power, or productive necessity. The paper identifies the leaking instrument panel while the engine driving the system toward obsolescence remains untouched.