EU AI rules require NGOs to label unreviewed public-interest text

The European Commission has brought the EU AI Act’s transparency regime into practical operation, with Article 50 obligations applying from August 2, 2026. The rules require providers and deployers to disclose certain AI interactions and synthetic content, while placing a specific burden on organizations that publish public-interest text without meaningful human review. The change matters to NGOs because much of their public communication concerns subjects explicitly identified by the Commission as matters of public interest, including public health, fundamental rights, environmental protection, public administration, and economic or political developments. The Commission’s Article 50 FAQ says the central question is not whether an organization is commercial, but whether it uses an AI system under its authority and how the resulting material reaches the public.

The Commission published its implementation guidance on July 20, followed by updated explanations and practical material later in the month. Under the framework, providers must design systems to tell people when they are directly interacting with AI and must apply machine-readable marks to synthetic audio, image, video, or text content. Deployers must disclose deepfakes and inform people exposed to emotion-recognition or biometric-categorization systems.

For NGOs, the most operationally significant provision concerns public-facing text. AI-generated or manipulated text intended to inform the public on matters such as public health, humanitarian conditions, rights, environmental protection, or public policy must be clearly labeled if it has not received human review or editorial control.

The Commission says a publication can qualify for an exemption from the text-labeling duty when a person with relevant knowledge deliberately examines its substance, or when a responsible editorial entity has authority to approve, alter, or reject it on substantive grounds. Fact-checking and assessing source reliability are included in the Commission’s description of editorial control.

A superficial check is not enough. Spell-checking, grammatical correction, or a procedural click-through does not count as human review. That distinction is especially important for smaller civil-society teams that may use AI to draft situation reports, advocacy statements, donor-facing public updates, or crisis communications before a rapid publication deadline.

Most NGOs will more commonly encounter the rules as deployers rather than providers. The Commission defines a deployer as a legal person or other body using an AI system under its authority, excluding personal, non-professional use. An organization can remain responsible even when contractors or freelancers operate the system on its behalf.

That does not mean every AI-assisted document must carry a public label. The guidance includes exceptions for standard editing and other uses that do not substantially alter input data or its semantics. It also distinguishes machine-readable marks applied by providers from visible or audible disclosures that deployers must give people when labeling deepfakes or certain public-interest text.

The immediate governance task is to map where AI enters public communications and assign responsibility for the final decision. Organizations should identify which tools generate or transform content, which outputs reach external audiences, what counts as a matter of public interest in their work, and who is authorized to perform substantive review.

The Commission says the rules are enforced mainly by national market-surveillance authorities, while the AI Office has a more limited role. Fines can reach €15 million or 3% of worldwide annual turnover for companies, although proportionality applies to smaller entities. A grace period extends certain marking obligations until December 2026 for systems placed on the market before August 2, but it does not remove the need to establish clear editorial controls now. The next test will be whether organizations can show not only that a human touched AI-generated content, but that accountable judgment shaped what was ultimately published.