When Is Content "AI-Generated Enough"? Labelling Synthetic Media under the Digital Services Act and the AI Act

URL SCAN: When Is Content "AI-Generated Enough"? Labelling Synthetic Media under the Digital Services Act and the AI Act
FIRST LINE: Computer Science > Computers and Society

The Dissection

This paper is not solving synthetic media. It is designing a classification-and-liability apparatus around it. Its central move is to treat labelling as a socio-technical chain: legal threshold → provenance signal → interface exposure → recipient interpretation → responsibility allocation.

The abstract correctly identifies that “AI-generated” is not a self-evident fact. It is an administrative category shaped by generation, manipulation, authenticity, statutory exceptions, platform design, and reporting practices.

The Core Fallacy

The paper’s failure under the Discontinuity Thesis is narrower than claiming that labels will save capitalism: it treats better classification as a potentially central lever against a transformation whose decisive mechanism is economic ownership and automation.

It mistakes epistemic visibility for control. A label can indicate that content passed through an AI system. It cannot establish whether the claim is false, whether human contribution was meaningful, or whether unlabeled content is authentic. Under P1, synthetic production makes cognitive output cheap and abundant. Labelling decorates that output; it does not restore scarcity, bargaining power, or productive participation. It manages deception at the margin while leaving P2 and P3 untouched.

Hidden Assumptions

• Recipients will notice labels, understand them, and update their beliefs accordingly.
• Provenance systems will be accurate, interoperable, persistent, and resistant to stripping or forgery.
• Legal thresholds for “generated,” “manipulated,” and “authentic-looking” can remain stable enough to administer.
• Providers, deployers, platforms, and uploaders can be assigned responsibility without strategic evasion or regulatory arbitrage.
• Label proliferation will not produce fatigue, universal suspicion, or a background signal that users ignore.
• Machine-readable markings will survive contact with reposting, editing, screenshots, compression, and hostile actors.
• Contestability is meaningful when classification is automated, opaque, high-volume, and backed by unequal resources.
• European obligations can materially govern actors and distribution channels operating beyond European enforcement.
• Disclosure can mitigate systemic risk without addressing who owns the models, compute, platforms, and distribution infrastructure.

Social Function

Primary classification: transition management and partial truth. Secondary classifications: prestige signaling and ideological anesthetic.

The partial truth is genuine: definitions, provenance, interface placement, and responsibility assignment determine whether a disclosure works. But the institutional function is to make synthetic media appear governable through compliance artifacts—marks, notices, databases, and reporting practices—while preserving the legitimacy of the existing regulatory order.

This is bureaucracy building a dashboard for a machine that is changing the ownership structure underneath it. The framework may make liability more legible and suppress some low-grade deception. It also gives institutions a respectable way to demonstrate action without confronting the underlying concentration of productive power.

The Verdict

Useful as a compliance map; strategically inadequate as a response to the discontinuity. Labelling may reduce particular harms, but it cannot prevent synthetic media from becoming the default condition of information or AI from erasing the scarcity of cognitive labor. The paper documents how institutions will administer the transition. It does not alter who owns the machinery producing it.