CASRAI Dictionary

Tag: ai act compliance

  • AI Labeling Act: Article 50 for Research Tools

    The EU’s working equivalent of an AI labeling act is Article 50 of Regulation (EU) 2024/1689 (the AI Act): it requires providers of interactive AI systems to disclose that users are talking to a machine, and requires generative AI outputs to be marked as artificially produced, with these disclosure duties applying from 2 August 2026. For university-built chatbots, literature-review assistants, and text-generation tools, that turns a US-style headline into a concrete EU compliance checklist.

    Article 50 is the AI Act’s transparency article: it sets disclosure duties for AI systems that talk to people, generate synthetic media, or publish AI-written text on public-interest matters — duties that apply whether or not the same system is separately classified as high-risk.

    What Does Article 50 of the AI Act Actually Require?

    Article 50 sets four distinct transparency duties, split between the organisations that build AI systems (“providers”) and the organisations that use them (“deployers”). None of these duties depend on high-risk classification — they apply to limited-risk systems such as chatbots and content generators precisely because those tools interact directly with the public.

    • Providers of AI systems intended to interact directly with people must ensure users are informed they are dealing with a machine, at the latest by the first interaction — unless this is already obvious to a “reasonably well-informed, observant and circumspect” person.
    • Providers of generative AI systems must mark synthetic audio, image, video, and text outputs in a machine-readable format that is detectable as artificially generated or manipulated.
    • Deployers of systems that create deepfakes must disclose that the image, audio, or video content has been artificially generated or manipulated.
    • Deployers publishing AI-generated text on matters of public interest must disclose this, unless the text has undergone human review with editorial responsibility.

    Which University-Built AI Tools Fall Within Scope?

    Article 50 catches tools that most research offices would not think to classify as “AI regulation risk” at all. Coverage already published elsewhere addresses high-risk classification and the AI Act’s literacy obligations (Article 4); this article is narrower, covering only the transparency/labelling layer that limited-risk academic tools commonly trigger.

    In practice, the following university-built or -deployed tools are typically in scope:

    • Chatbots and virtual research assistants embedded in library, admissions, or grants-support portals.
    • AI writing, summarisation, and literature-synthesis tools offered to staff or students.
    • Generative tools used to produce images, audio, or video for research communications and outreach.
    • AI-assisted drafting of public-facing text — press releases, policy briefs, funder reports — that has not passed through human editorial review.

    What Must Research Chatbots and Virtual Assistants Disclose?

    Any research-facing chatbot, virtual advisor, or conversational interface built or procured by a university is a “provider” (or, if procured from a vendor and deployed as-is, effectively acts as one for disclosure purposes) under Article 50(1). The system must make it clear the user is interacting with AI, not a human colleague, and that disclosure must land no later than the first exchange.

    The exemption is narrow: it applies only where AI involvement is already self-evident to a reasonably attentive user — a defence institutions should not rely on for anything branded as a “research assistant” or “advisor”, since that framing invites exactly the confusion Article 50 targets.

    How Must Generative Research Tools Mark Synthetic Content?

    Providers of generative AI research tools must mark outputs — text, image, audio, video — in a machine-readable format detectable as AI-generated. The European Commission’s Code of Practice on Transparency of AI-Generated Content, whose closing plenary ran on 10 June 2026, splits this work into two tracks that map cleanly onto institutional roles.

    Role under Article 50 Core duty Who this typically is at a university
    Provider Mark generative outputs in machine-readable form; disclose AI interaction to end users Teams building or fine-tuning chatbots, writing aids, or research-summarisation tools
    Deployer Disclose deepfakes and AI-generated public-interest text, unless human-edited under editorial responsibility Communications, press, and outreach teams using third-party generative tools

    The Commission has also published an EU icon set that deployers can use to label AI-generated content consistently, and several providers already implement machine-readable marking through the Coalition for Content Provenance and Authenticity (C2PA) “Content Credentials” standard, adopted by Adobe, LinkedIn, and Meta. Neither is mandatory in itself, but both give institutions a documented, defensible approach to the “machine-readable format” requirement rather than inventing one from scratch.

    When Do These Duties Take Effect — and Could the Timeline Slip?

    Article 50’s transparency obligations are due to become applicable on 2 August 2026 — one month from today. That said, institutions should not treat this as fixed: the European Commission’s “Digital Omnibus” proposal has floated delaying parts of the AI Act’s implementation into 2027, a live and unresolved question at the time of writing.

    The safer institutional posture is to build disclosure into procurement and development now, rather than wait for a delay that may not materialise, or may apply unevenly across provisions. The UK, by contrast, has no equivalent statutory labelling requirement: the House of Commons Library’s briefing on AI content labelling (20 January 2026) confirms proposals for wider UK AI regulation remain delayed, with labelling requirements unresolved.

    Answer-First Q&A

    What is the AI labeling act?

    In the United States, the AI Labeling Act is a Senate bill (S.2691) requiring disclosure for AI-generated content. In the European Union, the binding equivalent is Article 50 of the AI Act (Regulation (EU) 2024/1689), which sets transparency duties for providers and deployers rather than existing as a single named statute.

    What are the rules for AI label?

    Under Article 50, providers of AI systems that interact with people must disclose that users are talking to a machine, and providers of generative AI must mark outputs in a machine-readable format. Deployers must disclose deepfakes and AI-generated public-interest text, unless a human editor takes responsibility for it.

    Does AI content need to be labeled?

    Yes, where EU law applies. Under Article 50, synthetic audio, image, video, and text content generated or manipulated by AI must be marked as such, and deepfakes must be disclosed to viewers. The UK currently has no equivalent statutory labelling requirement, per the House of Commons Library.

    How does AI labeling work?

    Labelling combines technical and disclosure methods: visible notices such as captions and watermarks, embedded metadata, and machine-readable markers like the C2PA Content Credentials standard used by Adobe, LinkedIn, and Meta. The EU’s draft Code of Practice on AI-generated content is standardising these methods for Article 50 compliance.

    What This Means for Research Administrators and Tool Builders

    Article 50 does not require the conformity assessments, risk-management systems, or CE-marking-style processes that apply to high-risk AI. It requires something simpler but easy to overlook: honest, timely disclosure baked into the user interface and content pipeline. For research administration teams, that means three concrete actions.

    • Inventory every AI-enabled tool that talks to students, applicants, or the public — chatbots, virtual advisors, drafting assistants — and confirm each carries a first-interaction AI disclosure.
    • Check generative tools used in communications and outreach for machine-readable output marking, and prefer vendors that already implement C2PA or an equivalent provenance standard.
    • Route AI-drafted public-interest text (funder updates, policy explainers, press material) through documented human editorial review, which is the only recognised exemption from the public-interest text disclosure duty.

    This compliance layer sits alongside — and is simpler than — the wider AI-governance obligations research administration teams are already tracking under the AI Act’s high-risk and literacy provisions.

    The Bottom Line

    Article 50 is a transparency obligation, not a ban: universities can keep building and deploying chatbots and generative research tools, provided users are told they are talking to AI and synthetic outputs are marked as such. With application due from 2 August 2026 — and a possible Commission-proposed delay to 2027 still unresolved — institutions that build disclosure in now avoid having to retrofit it under deadline pressure later.