AI labels under scrutiny: what the EU's new platform reports mean

What the new reports show

Democracy Reporting International reported research into TikTok during Hungary's 2026 elections. It says it identified 2,043 posts from 22 accounts that received more than 50 million views. According to its submission, 71.6 per cent of the posts contained AI-generated content and 37.5 per cent of that content was not labelled by either the user or the platform.

The organisation says it reported 199 unlabelled posts to TikTok. Its report records the removal of 19 videos and one account, visibility restrictions on two posts, and 177 posts remaining unlabelled. These are the reporting organisation's findings, not a regulatory decision against TikTok.

Google's submission says YouTube requires creators to disclose meaningfully altered or generated photorealistic content and applies labels for viewers. Google also refers to its participation in the Coalition for Content Provenance and Authenticity. Meta reports that it labels certain photorealistic images, AI-generated advertising created with its own tools, and a wider range of content where it detects industry-standard indicators or receives a user disclosure. Meta says it may use a more prominent label where altered content presents a particularly high risk of materially deceiving the public.

The reports therefore expose both control measures and claimed gaps. They should be read as auditable evidence and stakeholder submissions, not as a single verified dataset or a regulator's final assessment. The Commission has said it will now open dialogues with VLOP and VLOSE signatories about their subscription levels and implementation.

Two legal frameworks, two different functions

The DSA Code addresses disinformation risk management by participating services. Its publication does not make every Code commitment a free-standing legal duty for every platform, AI developer or content business. The underlying DSA duties, the organisation's designation and role, its chosen commitments, and the facts of the service must be assessed separately.

Article 50 of the EU AI Act is different. Its transparency obligations have applied since 2 August 2026. Providers of generative AI systems must add machine-readable marks so that generated or manipulated outputs can be detected, subject to the legislation's detailed scope and exceptions. Deployers must disclose deepfakes and certain AI-generated or manipulated text published to inform the public on matters of public interest, again subject to the applicable exceptions.

The reporting period ended on 30 June 2026, before Article 50 became applicable. The new reports should therefore not be presented as proof of a breach of Article 50. They are, however, highly relevant operational evidence for controls that must work now.

A platform label and an embedded provenance mark are not interchangeable. A provider may need to mark an output at source; a deployer may need to make a disclosure to people; and a platform may separately need to assess and mitigate systemic risks. One measure may support several objectives, but it does not automatically discharge every duty.

Practical steps for founders and technology businesses

1. Map the role performed in each content flow: AI-system provider, deployer or publisher, platform, advertiser, agency or technical intermediary. Legal duties follow the actual function, not the product label.
2. Classify the output before release. Distinguish generated or manipulated audio, image, video and text; deepfakes; public-interest publications; political material; and ordinary editing that may fall outside the relevant rule.
3. Preserve machine-readable provenance through export, compression, editing and re-upload. Test whether marks remain detectable on the services and devices where content will actually appear.
4. Use a clear user-disclosure workflow for realistic synthetic media and relevant public-interest content. Do not rely on a term hidden in general conditions or assume that a platform will add the required label later.
5. Build an escalation route for elections and other high-risk events, covering detection, user reports, rapid review, evidence preservation, appeals, repeat misuse and communication with relevant platform or regulatory contacts.
6. Measure label coverage, failed detection, false positives, response times, removals, distribution restrictions and language or Member-State gaps. Record missing-data periods and methodology limits rather than presenting incomplete metrics as comprehensive.
7. Contract for preservation and transfer of provenance information, disclosure responsibilities, notice of policy or model changes, audit evidence and cooperation when content is challenged.
8. Keep marketing claims narrow. Signing a voluntary code, applying a visible label or using a provenance standard does not by itself establish compliance with the DSA or the AI Act.