Transparency Obligation

Official/legal meaning: A transparency obligation denotes duties imposed on AI actors (providers, deployers and, for certain systems, users) to communicate specific information about an AI system or its outputs so that affected natural persons and other stakeholders can identify AI involvement, understand relevant capabilities and limitations, and access information needed to exercise rights or pursue remedies. In the EU context these duties are set out for particular AI uses in Regulation (EU) 2024/1689 (the "EU AI Act") — notably in the chapter on transparency obligations (e.g., Article 50 for certain systems and Article 13 for high‑risk systems) and accompanying recitals (see Recitals 132–135). (Regulation (EU) 2024/1689, Art. 13; Art. 50; Recitals 132–135.)

Context and scope: Transparency obligations appear across international guidance and standards as a core element of trustworthy AI. In the EU AI Act they target defined categories and uses: providers must ensure persons interacting with AI are informed that they are interacting with AI; providers of systems that generate synthetic audio/image/video/text must mark outputs as artificially generated (machine‑readable where required); and deployers of biometric categorisation or emotion‑recognition systems must inform exposed persons — with timing, accessibility and limited exceptions (e.g., lawful criminal‑justice uses). For high‑risk AI systems, transparency is also operationalised through requirements such as instructions for use, technical documentation and traceability to enable users to interpret outputs. (Regulation (EU) 2024/1689, Arts. 13, 50.)

Practical implications for businesses: A transparency obligation compels organisations to adopt organisational, technical and communication measures. Typical business actions include: adding clear front‑facing notices (e.g., chatbot disclaimers; labels for AI‑generated content); implementing machine‑readable provenance or watermarking for synthetic outputs; producing and maintaining technical documentation, model cards and data sheets for internal and external stakeholders; ensuring accessibility of notices (including for vulnerable groups); and embedding transparency into procurement, contracts and vendor management. Where the EU AI Act applies, providers and deployers must also honour timing requirements (information at or before first interaction/exposure) and comply with exceptions (e.g., certain law‑enforcement authorisations, editorial responsibility exceptions). Non‑compliance risks regulatory enforcement, fines and reputational harm. NIST guidance treats transparency as an organisational characteristic tied to documentation and stakeholder‑tailored disclosure practices (NIST AI RMF, Part 1: Accountable and Transparent). (NIST AI RMF 1.0, Part 1; Regulation (EU) 2024/1689, Art. 50.)

Key requirements / criteria:

• Who: specified AI actors (providers; deployers; in some cases users) — see EU AI Act allocation of obligations. (Regulation (EU) 2024/1689, Art. 50.)
• What to disclose: e.g., notice of AI interaction, labelling of AI‑generated or manipulated content, information on emotion/biometric systems, and for high‑risk systems additional operational details (instructions for use, limitations, performance, human oversight). (Regulation (EU) 2024/1689, Arts. 13, 50.)
• How: clear, distinguishable, intelligible communications; machine‑readable marking for synthetic outputs where required; documentation tailored to stakeholder roles and accessible to vulnerable users. (Regulation (EU) 2024/1689, Art. 50; NIST AI RMF.)
• When: at the latest at first interaction or exposure (timeliness requirement in the EU AI Act). (Regulation (EU) 2024/1689, Art. 50.)
• Exceptions and limits: proportionate exceptions where disclosure would conflict with law enforcement mandates or other legal authorisations, or where editorial responsibility and human review apply for public‑interest publications. (Regulation (EU) 2024/1689, Recitals 132–135; Art. 50.)

Examples and cross‑references: common examples include a customer being told a virtual assistant is AI‑driven; a news outlet marking AI‑generated images or deepfakes; and warning banners for systems that infer emotions or sensitive biometric categories. The transparency obligation is closely related to, but distinct from, explainability (how a particular decision was reached), traceability (recording lifecycle artefacts), human oversight, and documentation practices (model cards, data sheets). International norms reinforce these expectations: the OECD AI Principles call on AI actors to provide meaningful, context‑appropriate information to enable understanding and challenge, and ISO/IEC 22989:2022 defines system transparency as making appropriate information available to relevant stakeholders. (OECD AI Principles, Principle 1.3; ISO/IEC 22989:2022, cl. 3.5.15; NIST AI RMF 1.0.)