CASRAI Dictionary

Tag: ai act code of practice

  • AI Act Code of Practice Timeline: 2026 Compliance Guide

    The AI Act code of practice covering transparency of AI-generated content moved from a draft published 8 May 2026, through a stakeholder consultation that closed 3 June 2026, to a final text published 10 June 2026 — all ahead of 2 August 2026, when Article 50’s transparency obligations become legally binding across the EU. The AI Act code of practice on transparency is a voluntary, European Commission-facilitated compliance tool, distinct from the earlier General-Purpose AI Code of Practice, that helps providers and deployers of generative AI systems meet the marking, detection and labelling duties set out in Regulation (EU) 2024/1689.

    For research offices, publishers and institutional communications teams, this is not an abstract EU process. Article 50 reaches any organisation whose AI-generated text, audio, image or video content reaches people in the EU — including AI-assisted research summaries and funder communications. This guide walks through the timeline as a compliance-planning tool.

    What is the AI Act code of practice on transparency?

    The Code of Practice on Transparency of AI-Generated Content is a non-binding framework drafted by independent experts, generative AI providers, deployer associations, civil society bodies and academics, facilitated by the EU AI Office. It gives organisations a recognised way to demonstrate compliance with Article 50 of the AI Act without waiting for harmonised technical standards to be finalised.

    This is a separate instrument from the General-Purpose AI (GPAI) Code of Practice under Article 56, published in final form on 10 July 2025 and applying to GPAI model providers since 2 August 2025. Confusing the two is a common error in searches for “AI Act code of practice” — the table below sets out the difference.

    Feature GPAI Code of Practice (Article 56) Transparency Code of Practice (Article 50)
    Legal basis Article 56, Regulation (EU) 2024/1689 Article 50, Regulation (EU) 2024/1689
    Audience Providers of general-purpose AI models Providers and deployers of generative AI systems
    Final text published 10 July 2025 10 June 2026
    Obligations apply from 2 August 2025 2 August 2026
    Core focus Safety, copyright, transparency documentation for models Marking, detection and labelling of AI-generated content

    The transparency code is organised around two working groups mirroring Article 50’s structure: Working Group 1 covers providers’ obligations to mark AI-generated audio, image, video and text in a machine-readable, detectable format; Working Group 2 covers deployers’ obligations to label deepfakes and AI-generated text on matters of public interest.

    What changed in the May 2026 draft guidelines?

    On 8 May 2026, the European Commission published draft implementation guidelines on Article 50 alongside the near-final Code of Practice text. These guidelines are the Commission’s own interpretive document — distinct from the stakeholder-drafted Code — clarifying how the transparency obligations apply in practice.

    The May draft addressed several points that had been ambiguous through the drafting rounds that ran from November 2025 to March 2026:

    • How “AI system” is scoped for the purposes of the human-interaction disclosure duty in Article 50(1);
    • The deepfake definition, including where content depicting real persons, places or events would falsely appear authentic;
    • The editorial-responsibility carve-out, under which AI-generated text on matters of public interest need not be labelled if it has undergone human review and is subject to editorial responsibility;
    • Expectations that marking techniques be interoperable, robust and reflect the “generally acknowledged state of the art” rather than a single mandated technology.

    What did the consultation closing 3 June 2026 cover?

    The Commission’s consultation on the May draft guidelines closed on 3 June 2026, giving providers, deployers, standards bodies and civil society a final window to flag practical gaps before the text was locked. In parallel, the multi-stakeholder drafting process for the Code of Practice itself held its closing plenary, and the AI Office published the final Code of Practice on Transparency of AI-Generated Content on 10 June 2026.

    This timing is deliberate: the guidelines interpret what Article 50 legally requires, while the Code offers voluntary methods — marking formats, labelling icons, detection mechanisms — for meeting those requirements. Signing the Code is optional; complying with Article 50 by 2 August 2026 is not.

    What applies from 2 August 2026 — and where is there a grace period?

    From 2 August 2026, Article 50 becomes legally applicable across all EU member states. Providers must ensure outputs of generative AI systems are marked in a machine-readable format detectable as artificially generated or manipulated. Deployers must disclose deepfakes and label AI-generated or manipulated text published on matters of public interest, unless a human has reviewed the content and taken editorial responsibility for it.

    One practical relief applies to systems already in the market. Legal trackers monitoring the rollout report that generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to retrofit the machine-readable marking requirement under Article 50(2) — a four-month bridge for legacy tooling rather than a change to the core application date.

    How should research offices prepare?

    Research administration, publisher and funder-communications teams should treat 2 August 2026 as a hard planning date, not a distant EU milestone. The obligations bite wherever AI-generated text, images or audio reach an EU audience — including institutional websites, funder newsletters, and AI-assisted drafting workflows.

    • Inventory every generative AI tool used to produce public-facing text, images, audio or video, and confirm whether outputs are already machine-readably marked;
    • Map authorship and editorial-review workflows against the human-review carve-out, so genuinely human-edited content is not mislabelled as AI-generated;
    • Align AI-use disclosure practices in manuscripts and grant narratives with existing publisher policies (for example, ICMJE and COPE guidance on declaring generative AI assistance), since Article 50 labelling and authorship disclosure are converging expectations;
    • Confirm with vendors supplying AI writing, transcription or media tools whether their systems will meet the marking requirement by 2 August 2026 or fall under the 2 December 2026 legacy window;
    • Assign clear internal ownership — communications, legal/compliance, and research integrity offices each hold part of this obligation and need a shared owner before August.

    Answer-first questions on the AI Act code of practice

    What is the EU AI Act code of practice?

    The EU AI Act code of practice on transparency is a voluntary framework, facilitated by the AI Office, that helps providers and deployers of generative AI systems meet Article 50’s marking, detection and labelling duties. It was finalised on 10 June 2026, ahead of the 2 August 2026 application date, and sits alongside a separate GPAI Code of Practice covering model-level obligations under Article 56.

    Is there a UK equivalent to the AI Act code of practice?

    No. The UK has no AI-specific legislation equivalent to the EU AI Act; AI is instead regulated through existing sector frameworks. UK research institutions, publishers and vendors that publish AI-generated content reaching EU audiences, or that operate EU subsidiaries, must still meet Article 50’s transparency obligations from 2 August 2026.

    How does the transparency code relate to the AI Act’s risk categories?

    The AI Act classifies systems into four risk tiers — unacceptable, high, limited and minimal risk. Article 50’s transparency duties sit within the “limited risk” tier and apply horizontally to generative and interactive systems regardless of their risk classification elsewhere, which is why the transparency code applies more broadly than the high-risk rules.

    Implications and outlook

    The 2 August 2026 application date closes a year-long drafting process that began in September 2025 and ran through three formal drafting rounds before the May 2026 draft and June 2026 consultation. For research-adjacent organisations, the practical implication is less about the Code of Practice itself — which remains voluntary — and more about Article 50, which is not. Institutions that already maintain authorship-disclosure and editorial-review workflows for generative AI have a head start.

    Expect further guidance around the 2 December 2026 legacy-marking deadline, and continued convergence between AI Act transparency labelling and research-integrity disclosure norms from bodies such as ICMJE and COPE. Organisations tracking both processes together, rather than as separate compliance tracks, will be better placed for the obligations that follow.

    See CASRAI’s related coverage of research administration compliance workflows and authorship transparency disclosures for how generative AI disclosure expectations intersect with existing research-integrity practice.

  • NIST AI Risk Management Framework Playbook Guide for Research Offices

    The NIST AI Risk Management Framework Playbook is a voluntary, non-mandatory companion resource that translates the four functions of the NIST AI RMF — Govern, Map, Measure, Manage — into suggested actions research offices can adapt into a working AI-tool review checklist, without adopting the document as a rigid audit standard.

    The AI RMF Playbook is a reference companion to NIST AI RMF 1.0, published by the U.S. National Institute of Standards and Technology on 26 January 2023, that maps suggested implementation actions to each subcategory in the framework’s Core (Tables 1–4). Research administration offices evaluating AI writing tools, manuscript-screening systems, grant-matching algorithms, or peer-review assistants are increasingly being asked — by faculty, ethics committees, or funders — to show some structured basis for that review. The Playbook is the most directly usable NIST artefact for building one, but most explainers stop at describing the four functions rather than showing how a research office turns them into an actual intake form. This walkthrough does that conversion.

    What is the NIST AI RMF Playbook?

    The NIST AI Risk Management Framework Playbook is a living, voluntary implementation resource published alongside AI RMF 1.0. NIST’s AI Resource Center states plainly that the Playbook “is neither a checklist nor set of steps to be followed in its entirety” — organisations are meant to borrow “as many – or as few” suggestions as fit their use case.

    Each suggestion in the Playbook is tied to a specific subcategory under one of the framework’s four functions:

    • Govern — establishes the culture, policies, and accountability structures for managing AI risk across the organisation.
    • Map — establishes context: what the AI system does, who it affects, and what could go wrong.
    • Measure — analyses, benchmarks, and tracks identified risks using quantitative and qualitative methods.
    • Manage — allocates resources to risks by priority and monitors the system after deployment.

    NIST distributes the Playbook as a PDF, CSV, Excel workbook, and JSON file via the AI Resource Center, and it is updated approximately twice per year as AI technology and community feedback evolve. That release cadence matters operationally: a review checklist built from the Playbook should be version-dated and re-checked against each update rather than treated as a one-time policy document.

    Turning the four functions into a research-office AI-tool checklist

    The Playbook’s value for a research office is not the four function names — it is the subcategory-level actions underneath them, which read almost like intake-form questions once relabelled. Below is a working mapping from AI RMF 1.0 Core subcategories to the questions a research administration office can ask when a faculty member or department proposes adopting an AI tool (a manuscript screener, grant-matching assistant, or peer-review support system).

    AI RMF subcategory (paraphrased) Research-office checklist question
    GOVERN 1.1 — legal and regulatory requirements are understood and documented Does this tool trigger institutional research-ethics, data-protection, or funder AI-disclosure obligations?
    GOVERN 2.1 — roles and responsibilities for AI risk are assigned Who in the office owns ongoing oversight of this tool once it is approved?
    MAP 1.1 — intended purpose and deployment context are understood What specific research-administration task is this tool being used for, and by whom?
    MAP 5.1 — likelihood and magnitude of impacts are assessed What happens to a manuscript, grant application, or reviewer assignment if the tool errs?
    MEASURE 2.6 / 2.7 — safety, security, and resilience are evaluated Has the vendor supplied evidence of testing for bias, data leakage, or hallucinated citations?
    MANAGE 1.1 — determination of whether the system meets its objectives Did a pilot period show the tool actually improves the workflow it was bought for?
    MANAGE 4.1 — post-deployment monitoring plans are implemented Who re-reviews this tool annually, and what triggers an early re-review?

    Built this way, the checklist stays traceable to a named NIST subcategory for every question an ethics committee or auditor might ask “why do you check this?” — which is the practical benefit of using the Playbook rather than writing a bespoke policy from scratch.

    NIST AI RMF Playbook vs ISO 42001 vs the EU AI Act Code of Practice

    Research offices operating internationally increasingly need to know how the voluntary US framework relates to certifiable and regulatory instruments used elsewhere. None of the three is a substitute for the others; they serve different purposes and audiences.

    Framework Status Best fit for a research office
    NIST AI RMF 1.0 + Playbook Voluntary US guidance, published January 2023 Building an internal AI-tool review process and shared vocabulary
    ISO/IEC 42001:2023 Certifiable international AI management-system standard Institutions seeking third-party certification of their AI governance programme
    EU AI Act General-Purpose AI Code of Practice Regulatory compliance mechanism under Regulation (EU) 2024/1689, applying to GPAI providers from August 2025 Institutions in, or contracting with, the EU that procure general-purpose AI models

    A practical pattern for a research office with European partners: use the AI RMF Playbook’s subcategories to build the internal checklist, use ISO 42001’s clause structure if formal certification is the goal, and treat the EU AI Act Code of Practice as a due-diligence question to put to any GPAI vendor — “can you show your Code of Practice commitments?” — rather than as a framework the research office itself must implement.

    Answer-first questions about the AI RMF Playbook

    What is the NIST AI RMF Playbook?

    The NIST AI RMF Playbook is a voluntary companion resource to NIST AI RMF 1.0 that provides suggested actions for each subcategory across the framework’s four functions — Govern, Map, Measure, and Manage. It is not a checklist to complete in full; organisations select the suggestions relevant to their own AI use case.

    What are the two main parts of the NIST AI RMF?

    NIST AI RMF 1.0 is structured in two parts: Part 1 sets out foundational context — the framing of AI risks and the characteristics of trustworthy AI — and Part 2 contains the Core, organised into the four functions, their categories, and subcategories that the Playbook then operationalises.

    Is NIST AI RMF compliance mandatory?

    No. The AI RMF and its Playbook are voluntary for private organisations and most research institutions; there is no certification body. Some US federal agencies reference it in AI-procurement guidance, and funders or partner institutions may ask an office to show alignment as a matter of due diligence rather than legal obligation.

    What are the seven steps of the NIST Risk Management Framework?

    The seven-step RMF — Prepare, Categorize, Select, Implement, Assess, Authorize, Monitor — comes from NIST SP 800-37, a separate cybersecurity authorisation framework for federal information systems. It is distinct from the AI RMF’s four functions; research offices should not conflate the two when a vendor or auditor cites “the NIST framework.”

    Implications for research offices

    Research administration is adopting AI tools faster than most offices have built governance for: manuscript-screening assistants, grant-matching engines, and reviewer-recommendation systems are already in use across publishers and institutions. Building the intake checklist directly from AI RMF Playbook subcategories gives a research office a defensible answer when asked how a tool was vetted, without waiting for a mandatory US or UK regulatory framework to arrive.

    Because NIST revises the Playbook roughly twice yearly, and because the EU AI Act’s GPAI obligations are still being phased in through 2025–2026, offices that adopt this checklist approach should treat it as a living document, re-checked at each Playbook release rather than filed away after a single review cycle.

  • GPAI Code of Practice Signatories: Who Signed and What It Means for Research Tool Vendors

    As of mid-2026, Amazon, Anthropic, Google, IBM, Microsoft, OpenAI, Mistral AI and Aleph Alpha have signed the EU’s General-Purpose AI Code of Practice in full, xAI has signed only its Safety and Security chapter, and Meta has declined to sign at all. For research offices and publishers procuring AI-enabled tools, a vendor’s foundation-model supplier and that supplier’s gpai code of practice signatories status is now a material, checkable compliance signal.

    The General-Purpose AI Code of Practice (GPAI CoP) is a voluntary compliance framework, published by the European Commission on 10 July 2025, that lets providers of general-purpose AI models demonstrate adherence to the transparency, copyright and safety obligations of the EU AI Act’s Articles 53 and 55.

    Which AI labs have signed the Code of Practice?

    The European Commission maintains and continuously updates a public list of signatories on its digital-strategy portal. Signatories were first published on 1 August 2025, one day before the AI Act’s GPAI obligations took effect on 2 August 2025. The largest foundation-model providers active in academic and publishing tooling have signed all three chapters.

    Provider Signature status Chapters covered
    OpenAI Full signatory Transparency, Copyright, Safety and Security
    Microsoft Full signatory Transparency, Copyright, Safety and Security
    Google Full signatory Transparency, Copyright, Safety and Security
    Amazon Full signatory Transparency, Copyright, Safety and Security
    Anthropic Full signatory Transparency, Copyright, Safety and Security
    IBM Full signatory Transparency, Copyright, Safety and Security
    Mistral AI Full signatory Transparency, Copyright, Safety and Security
    Aleph Alpha Full signatory Transparency, Copyright, Safety and Security
    xAI Partial signatory Safety and Security only
    Meta Non-signatory None

    A Signatory Taskforce, chaired by the EU AI Office, was established to help signing providers implement the Code consistently and to keep the commitments current as models are updated. Institutions should check the Commission’s live list before relying on any third-party summary, including this one, since new signatories are added on a rolling basis.

    Which major providers have not signed, and why?

    Meta is the most significant non-signatory. In July 2025, Meta’s Chief Global Affairs Officer Joel Kaplan stated that the Code introduces legal uncertainties and obligations that “go far beyond the scope of the AI Act,” and confirmed Meta would not sign. xAI took a narrower position, signing only the Safety and Security chapter while rejecting the Transparency and Copyright chapters as, in the company’s view, potentially harmful to innovation.

    • Meta — declined to sign any chapter
    • xAI — signed Safety and Security only; declined Transparency and Copyright
    • Alibaba, Baidu and DeepSeek — no public commitment to sign as of early 2026

    Declining to sign does not exempt a provider from the AI Act itself. The Code is a voluntary route to demonstrating compliance; the underlying legal obligations in Articles 53 and 55 remain binding on any GPAI provider placing a model on the EU market, signatory or not.

    What does signing actually commit a provider to?

    The Code is organised into three chapters, each addressing a distinct obligation under the AI Act. Signing the full Code commits a provider to detailed documentation, copyright policy and, for the largest models, systemic-risk management.

    • Transparency — model documentation covering capabilities, limitations and training-data summaries, shared with downstream providers on request within 14 calendar days.
    • Copyright — a policy aligned with EU copyright law, including respecting rights-holder opt-outs and mitigating infringing outputs.
    • Safety and Security — applies only to models classified as carrying systemic risk (broadly, those trained above 1025 floating-point operations, per Article 55); requires independent external evaluation, incident reporting and a documented risk-management framework.

    Non-signatories that provide GPAI models must still satisfy Articles 53 and 55 through other means and face closer supervisory scrutiny from the AI Office. Under Article 101 of the AI Act, the Commission can fine GPAI providers up to €15 million or 3% of total worldwide annual turnover, whichever is higher, for breaches of these obligations — the same penalty tier applies regardless of Code signature status.

    What this means for research-tool vendor risk assessments

    Research offices, publishers and institutional procurement teams rarely contract directly with foundation-model developers. They contract with the AI-enabled research tools — plagiarism and integrity checkers, peer-review triage systems, writing and translation assistants, literature-discovery platforms — built on top of those models. The signatory status of the underlying model provider is a proxy for how much documentation, incident transparency and risk evidence a research-tool vendor can realistically pass through to an institutional buyer.

    A vendor built on a full Code signatory’s model can typically point to a standardised Model Documentation Form, published training-data summaries, and (for systemic-risk models) an externally evaluated Safety and Security Model Report. A vendor built on a non-signatory model has none of this by default; it must obtain equivalent assurances directly from its model supplier or demonstrate compliance through bespoke documentation, which is harder for a research office to verify at procurement stage.

    • Ask vendors which foundation model(s) power their product and whether that provider is a full, partial or non-signatory.
    • For partial signatories such as xAI, confirm whether the tool relies on capabilities covered only by the unsigned Transparency or Copyright chapters.
    • Where a vendor relies on a non-signatory model, request the provider’s own AI Act compliance documentation directly, rather than accepting the vendor’s assurance alone.
    • Track the Commission’s signatory list periodically — a vendor’s compliance posture can change as its underlying model supplier’s status changes.

    This procurement lens is distinct from the legal-compliance framing most coverage of the Code takes: research administration offices are not GPAI providers themselves, but they inherit downstream documentation risk every time they adopt an AI-enabled tool, a consideration that belongs alongside existing due-diligence practice for research administration vendor reviews.

    Common questions on GPAI Code of Practice signatories

    What is a GPAI system?

    A general-purpose AI (GPAI) model is a foundation model, such as those underpinning ChatGPT, Gemini or Claude, capable of performing a wide range of tasks without being built for one specific use. Under the AI Act, providers of GPAI models placed on the EU market carry distinct transparency and, above certain compute thresholds, systemic-risk obligations.

    What happens if a provider does not sign the Code of Practice?

    A non-signatory is not exempt from the AI Act. It must demonstrate compliance with Articles 53 and 55 through alternative means, and the EU AI Office has indicated it will apply closer regulatory scrutiny to non-signatories, increasing enforcement uncertainty relative to signatories.

    What are the penalties for GPAI Act non-compliance?

    Under Article 101 of the AI Act, the Commission can fine a GPAI provider up to €15 million or 3% of total worldwide annual turnover, whichever is higher, for breaches of the transparency, copyright or systemic-risk obligations, independent of whether the provider signed the Code.

    Can a provider sign only part of the Code of Practice?

    Yes. xAI signed only the Safety and Security chapter of the Code, declining the Transparency and Copyright chapters. Partial signature means a provider gains reduced administrative burden for the chapters it signed, while still needing to demonstrate compliance with the others through other evidence.

    Outlook for research administration

    The signatory list will keep shifting as new models cross the compute thresholds in Articles 53 and 55, and as the Signatory Taskforce publishes further implementation guidance. Research offices building AI-tool procurement checklists should treat Code of Practice status as one input alongside existing vendor due-diligence questions on data provenance, licensing terms and institutional data protection — not as a substitute for direct verification against the Commission’s live signatory list.

  • EU AI Office: Enforcement for Research Bodies

    The EU AI Office does not enforce most of the AI Act. It is a European Commission unit, inside the Directorate-General for Communications Networks, Content and Technology (DG CNECT), with exclusive competence over general-purpose AI (GPAI) models. Day-to-day enforcement against high-risk AI systems — the category covering most tools used in universities, funders and public research bodies — falls to each Member State’s national market surveillance authority, not the AI Office.

    The EU AI Office is the Commission’s central coordinating body for Regulation (EU) 2024/1689 (the AI Act), responsible for supervising GPAI models, chairing the technical governance structure and preparing Commission guidance — while national authorities retain enforcement power over almost everything else.

    What is the EU AI Office?

    The AI Office was established by a European Commission decision in January 2024, alongside political agreement on the AI Act. It sits within DG CNECT rather than as a stand-alone agency, and functions legally as part of the Commission — so references to “the AI Office” in the Act’s text are references to the Commission acting through that unit.

    Its headquarters are in Brussels. Wikipedia’s infobox for the European Artificial Intelligence Office records around 60 staff at 2024 launch, projected above 140, under Director Lucilla Sioli. The Office also acts as Secretariat to the European AI Board, the forum of one representative per Member State coordinating national implementation.

    • Supervises GPAI model providers under AI Act Chapter V
    • Drafts codes of practice, guidelines and implementing acts for the Commission
    • Coordinates joint investigations across Member States on cross-border AI risk
    • Runs the AI Act Service Desk and single information platform
    • Chairs the scientific panel of independent experts monitoring systemic-risk models

    Who actually enforces the AI Act — the AI Office or national authorities?

    Enforcement is split by system type, not centralised in one body. The AI Office’s remit is narrow but powerful: only GPAI models and systems — the foundation models underpinning many downstream research tools. Everything else, including the high-risk systems a university, funder or public research agency is far more likely to deploy directly, is enforced nationally.

    Each Member State designates one or more market surveillance authorities (MSAs) under Article 74, alongside a “notifying authority” overseeing conformity-assessment bodies. Because States may designate sector-specific bodies rather than one regulator, the map is fragmented: CMS Law’s 2025 enforcement analysis notes that, once sectoral designations are counted, several thousand bodies across the EU can hold market-surveillance-authority status, with AI systems now added to their remit.

    A separate rule applies to the EU’s own institutions. Under Article 74(9), the European Data Protection Supervisor (EDPS) is the market surveillance authority for AI systems used by EU institutions, bodies, offices and agencies — relevant to EU-funded research infrastructures and executive agencies, as distinct from national universities and funders.

    Body Enforces Covers Key power
    EU AI Office GPAI model obligations (Chapter V) Foundation-model providers, EU-wide Model evaluations, mitigation orders, market withdrawal
    National market surveillance authority High-risk and other AI system obligations Deployers/providers within one Member State, incl. universities and public bodies Inspections, corrective orders, fines
    European Data Protection Supervisor All AI Act obligations EU institutions, bodies, offices and agencies Fines against EU public administration
    European AI Board Coordination, not direct enforcement All 27 Member States (via national reps) Consistency, joint-investigation coordination

    Does the research exemption apply to universities and public bodies?

    Partly, and the boundary matters more than most explainers acknowledge. Article 2(8) states that obligations do not apply to research, testing or development activity on an AI system before it is placed on the market or put into service. Article 2(6) separately exempts systems developed and used for the sole purpose of scientific research and development.

    Neither carve-out protects a university once it moves from research into operational use. Annex III(3) classifies AI systems used to evaluate exam answers, determine admission or assess applicants as high-risk. A plagiarism-detection or admissions-scoring tool a university actually deploys against students is therefore fully in scope — and, because most universities and funders are “bodies governed by public law”, Article 27 requires a fundamental rights impact assessment (FRIA) before deployment.

    How can research institutions and public bodies seek guidance?

    Three channels exist, and institutions frequently default to the wrong one. The AI Act Service Desk (ai-act-service-desk.ec.europa.eu) is the Commission’s central portal where any stakeholder, including a university legal office or funder’s compliance team, can submit a question and get an answer from a Commission-coordinated expert team; it is the right first stop for interpretive questions on scope, classification or the research exemptions above.

    For enforcement-specific queries — “is our deployed system high-risk, and what must we file?” — the correct contact is the national market surveillance authority in the institution’s own Member State, not the AI Office, which has no jurisdiction over nationally-deployed high-risk systems. EU-affiliated bodies should instead approach the EDPS. National governments must separately establish AI regulatory sandboxes, giving public research bodies a supervised route to trial new systems before full-scale deployment.

    What are the penalties for AI Act non-compliance?

    Article 99 sets three fine tiers, using the higher figure for large organisations and the lower for SMEs and start-ups:

    • Up to €35 million or 7% of global annual turnover for breaching prohibited AI practices (Article 5)
    • Up to €15 million or 3% of global annual turnover for breaching most other provider or deployer obligations
    • Up to €7.5 million or 1% of global annual turnover for supplying incorrect, incomplete or misleading information to authorities or notified bodies

    Article 101 gives the Commission a separate fining power against GPAI model providers, up to 3% of worldwide annual turnover or €15 million, whichever is higher, for infringements the AI Office identifies through model evaluation. Public-sector bodies are not exempt from Article 99 fines, though Member States retain some discretion over how penalties apply to public administration.

    Providers can reduce GPAI exposure by signing the General-Purpose AI Code of Practice, published by the AI Office in 2025 with independent experts across transparency, copyright and safety/security chapters. Adherence is voluntary but, pending harmonised standards, creates a presumption of conformity — worth knowing for institutions procuring GPAI tools from signatory vendors.

    Answer-first questions on the EU AI Office

    Where is the EU AI Office?

    The EU AI Office is headquartered in Brussels, inside the European Commission’s Directorate-General for Communications Networks, Content and Technology (DG CNECT). It is not a separate legal agency; it operates as a Commission unit with its own director, staff and published mandate under the AI Act’s governance provisions.

    Who is the head of the EU AI Office?

    The EU AI Office is led by Director Lucilla Sioli, who reports within DG CNECT’s management structure. The director’s mandate covers GPAI supervision, Secretariat duties for the European AI Board, and coordination of the scientific panel of independent experts that monitors systemic-risk models.

    What is a market surveillance authority?

    A market surveillance authority is the national body a Member State designates to monitor, inspect and take corrective or punitive action against non-compliant products — including, under the AI Act, high-risk AI systems deployed within that country’s territory, such as university admissions or assessment tools.

    What is post-market monitoring under the AI Act?

    Post-market monitoring is the ongoing obligation on providers and deployers of high-risk AI to actively collect and analyse performance data after deployment. It feeds directly into market surveillance authority oversight, giving regulators evidence to investigate serious incidents or systemic risk once a system is in real-world use.

    Implications for research administrators

    The practical takeaway is that “who do we ask” and “who can fine us” are different questions with different answers. The AI Office is the right destination for interpretive guidance on GPAI; the national market surveillance authority holds actual enforcement jurisdiction over a deployed high-risk system inside a research institution.

    As GPAI-based tools proliferate across grant review, plagiarism screening and admissions, institutions that conflate the AI Office’s central mandate with national enforcement risk misdirecting queries and missing the FRIA obligations Article 27 attaches to public bodies. Building this literacy now, ahead of the Act’s staged 2025–2027 application timeline, is cheaper than resolving a misdirected enforcement dispute later. For related governance context, see CASRAI’s research administration resources.

  • AI Act Article 50 Transparency in Research Tools

    The EU AI Act’s Article 50 transparency obligations apply from 2 August 2026 and require any research chatbot, literature-review assistant or generative lab tool that interacts with people or produces synthetic content to disclose that it is AI. This duty is separate from the machine-readable watermarking requirement in Article 50(2), which the AI Omnibus provisional agreement has pushed back to 2 December 2026 for generative systems already on the market — so procurement teams cannot treat the whole article as delayed.

    Article 50 of Regulation (EU) 2024/1689 is the EU AI Act provision that sets disclosure duties for AI systems that interact directly with people, generate synthetic content, perform emotion recognition or biometric categorisation, or produce deepfakes and AI-written public-interest text. For research offices, this covers a growing shelf of everyday tools: AI-assisted literature-review platforms, participant-facing chatbots, lab-based generative-image tools and AI drafting assistants used in public engagement.

    What does Article 50 actually require?

    Article 50 sets out four distinct disclosure duties, and a research tool can trigger more than one at once. Under Article 50(1), providers of systems intended to interact directly with people — chatbots, virtual assistants, conversational research interfaces — must design them so users are told they are dealing with AI, at the latest by the first interaction. Draft Commission Guidelines published on 8 May 2026 confirm that AI agents fall within this duty, and that disclosure must be repeated where a single notice at the outset would not be obvious later in the exchange.

    Article 50(2) requires providers of generative AI to mark synthetic audio, image, video or text outputs in a machine-readable, detectable format. Article 50(3) requires deployers of emotion-recognition or biometric-categorisation systems to inform exposed individuals, alongside GDPR Articles 12–14. Article 50(4) requires deployers to label deepfakes and disclose AI-generated public-interest text, unless a named person holds editorial responsibility after substantive human review.

    What applies from August 2026, regardless of the watermarking delay?

    The full transparency regime enters into force on 2 August 2026. Only one obligation has been pushed back: the AI Omnibus provisional agreement of May 2026 gives generative AI systems already placed on the market before that date until 2 December 2026 to meet the machine-readable marking requirement in Article 50(2). Nothing else in Article 50 moves.

    That means the disclosure duties most relevant to research offices are unaffected by the delay:

    • Article 50(1) chatbot and virtual-assistant disclosure applies from 2 August 2026 in full.
    • Article 50(3) emotion-recognition and biometric-categorisation disclosure applies from 2 August 2026 in full.
    • Article 50(4) deepfake and public-interest text labelling applies from 2 August 2026 in full.
    • Only the technical marking format under Article 50(2) has a grace period, and only for pre-existing systems.

    A research office that assumes “the watermarking clause is delayed, so we have more time” is conflating one narrow technical carve-out with the whole article. The European Commission’s own Compliance Checker data indicates transparency obligations are the second most common compliance trigger after AI literacy, affecting around 33% of organisations that have assessed themselves against the Act.

    Which research AI tools are caught by Article 50?

    Most research-facing AI tools map cleanly onto one or more Article 50 provisions. The table below sets out the mapping research offices should use when auditing procured and internally built tools.

    Research AI tool type Article 50 provision Who must act Deadline
    Literature-review or systematic-review chatbot assistant 50(1) — AI interaction disclosure Deployer (institution), if using a third-party tool; provider, if built in-house 2 August 2026
    Participant recruitment or survey chatbot 50(1) — AI interaction disclosure Deployer 2 August 2026
    Lab tool generating synthetic images, audio or text (e.g. synthetic dataset generation) 50(2) — machine-readable marking Provider 2 August 2026; pre-existing systems to 2 December 2026 for marking only
    Emotion-recognition or biometric-categorisation research instrument 50(3) — disclosure to exposed individuals Deployer 2 August 2026
    AI-drafted public engagement or press content 50(4) — public-interest text labelling Deployer, unless human-reviewed with editorial responsibility 2 August 2026

    Note that the same tool often needs two audits: a chatbot that also produces AI-written summary text for publication can trigger both 50(1) and 50(4).

    Does the research exemption cover the tools you actually use?

    Recital 25 of the AI Act exempts AI systems and models developed and put into service for the sole purpose of scientific research and development. This exemption is narrower than it sounds. It covers AI built as the object of research — a novel model a lab is developing and testing — not commercial or off-the-shelf tools that a research team merely uses to do research.

    A university deploying a commercially available literature-review assistant, a general-purpose chatbot, or a vendor’s lab-imaging tool does not benefit from the Recital 25 carve-out for that deployment. The institution acts as a deployer under Article 50 the moment that tool interacts with people or generates in-scope content, regardless of the research being scientific in nature. Procurement teams should not assume “we’re a research organisation” is itself an exemption — the exemption attaches to the AI system’s development purpose, not the purpose of the team using it.

    Common questions on Article 50 and research AI

    What are the transparency obligations under Article 50 of the AI Act?

    Article 50 sets four disclosure duties: providers of interactive AI (chatbots, assistants) must flag their AI nature at first contact; providers of generative AI must mark synthetic outputs; deployers using emotion-recognition or biometric tools must inform exposed individuals; and deployers publishing deepfakes or AI-written public-interest text must label it as such, unless human-reviewed.

    What is the EU Code of Practice on AI-generated content?

    It is a voluntary Commission-coordinated framework covering Article 50(2) and 50(4), setting a standardised EU visual label for AI content, a taxonomy separating fully AI-generated from AI-assisted material, and modality-specific labelling guidance. A second draft was published in March 2026, with a final version expected by June 2026.

    Why has Article 50’s transparency obligation been criticised as insufficient?

    Academic analysis, including work published via the University of Glasgow’s repository, argues that a simple AI-interaction notice does not stop users from over-trusting confidently worded but unverified chatbot output — disclosure alone does not compel verification behaviour, which matters directly for research assistants summarising literature.

    Are AI systems built only for research exempt from Article 50?

    Only narrowly. Recital 25 exempts AI developed solely for scientific research and development as the object of study. It does not exempt a research office’s use of commercial, off-the-shelf AI tools — those deployments remain subject to Article 50 in the same way as any other organisation’s use.

    What this means for research offices

    Research administration teams procuring or building AI tools should treat 2 August 2026 as the operative date for every disclosure duty except machine-readable marking of pre-existing generative systems. Practical steps:

    • Inventory every AI tool used in research workflows — literature review, participant engagement, lab generation, public communications — and tag each against Articles 50(1)–50(4).
    • Confirm vendor contracts assign responsibility: does the vendor act as provider, leaving the institution as deployer with its own disclosure duties?
    • Check chatbot and assistant interfaces disclose AI involvement clearly at first use, not buried in terms and conditions.
    • Do not treat the December 2026 marking grace period as covering anything beyond Article 50(2) technical marking of pre-existing systems.
    • Review public-facing AI-drafted content (news releases, dissemination summaries) for the human-review and editorial-responsibility carve-out under Article 50(4).

    Institutions with dedicated research administration functions are well placed to run this audit alongside existing research-integrity and data-governance processes, since the same tool inventory typically maps onto GDPR and funder AI-use disclosure requirements already in place.

    What happens next

    The Commission’s final Guidelines and the finalised Code of Practice are both due before 2 August 2026, and both will refine — not delay — the duties above. Offices waiting for the Code before acting on Article 50(1), (3) and (4) will miss the window, since the Code covers marking and labelling detail only, not the underlying legal duty to disclose. Institutions best placed by August will have already mapped their AI tool inventory against Article 50, rather than treating the whole article as paused.

  • AI Act Watermarking Obligations Delay: December 2026

    The AI Act watermarking obligations delay pushes Article 50(2) of the EU AI Act — the machine-readable marking duty for synthetic content — from 2 August 2026 to 2 December 2026 for AI systems already on the market before that date. This is a narrow, four-month transitional concession agreed in the EU’s Digital Omnibus trilogue on 7 May 2026. It does not touch Article 50(1), the separate duty to disclose that a person is interacting with an AI system, which still applies from 2 August 2026 as originally scheduled.

    Article 50 of Regulation (EU) 2024/1689 (the AI Act) is the transparency article governing four distinct duties: disclosure of AI interaction, machine-readable marking of synthetic content, deployer labelling of deepfakes, and labelling of AI-generated text on matters of public interest. Confusing these four sub-obligations — or confusing this watermarking delay with the separate, much longer postponement of high-risk AI system rules — is the most common compliance-timeline error research offices, publishers and institutional AI-governance teams are currently making.

    What actually changed in the Digital Omnibus trilogue

    The Council of the European Union and the European Parliament reached a provisional political agreement on the AI-related Digital Omnibus on 7 May 2026, after a nine-hour trilogue session held under the Cypriot Council Presidency. The text still requires formal endorsement by both institutions and legal-linguistic revision before it is published in the Official Journal, but its substance on watermarking is settled.

    The European Commission’s original November 2025 Digital Omnibus proposal sought a six-month postponement of the Article 50(2) marking obligation. The European Parliament’s negotiating mandate, adopted on 26 March 2026, pushed back for a shorter, three-month postponement. The trilogue compromise landed on four months, moving the application date for existing systems from 2 August 2026 to 2 December 2026.

    This is a narrow, technical fix, not a policy reversal. The stated rationale is operational: the AI Office’s Code of Practice defining how to meet the marking duty is still being finalised, and providers argued they could not build machine-readable marking, metadata and detector tooling against guidance that had not yet stabilised.

    Article 50(2) watermarking vs Article 50(1) disclosure: the nuance

    This is the distinction research administrators need to track separately, because press coverage frequently blurs it. Article 50(1) and Article 50(2) are different obligations with different deadlines, and only one of them moved.

    Provision What it requires Who it binds Application date Delayed?
    Article 50(1) Inform natural persons they are interacting with an AI system (e.g. chatbots) Providers 2 August 2026 No — unchanged
    Article 50(2) Machine-readable marking of synthetic audio, image, video or text output, detectable as artificially generated Providers 2 December 2026 (existing systems) Yes — 4-month delay
    Article 50(3) Label deepfake image, audio or video content shown to the public Deployers 2 August 2026 No — unchanged
    Article 50(4) Label AI-generated text published to inform the public on matters of public interest Deployers 2 August 2026 No — unchanged

    In other words, the disclosure and labelling duties that sit closest to end-user and reader-facing transparency — telling a person they are talking to a bot, or flagging that an image is a deepfake — proceed on the original 2 August 2026 timetable. Only the upstream, provider-side technical marking duty in Article 50(2) has moved.

    Who is affected, and from what date

    The four-month extension operates as a transitional grace period, not a blanket new deadline. It applies specifically to generative AI systems already placed on the EU market before 2 August 2026. Providers bringing a new generative AI system to the EU market on or after 2 August 2026 must comply with Article 50(2) marking from the point of placement, with no transitional window.

    • Existing systems (on the EU market before 2 August 2026): Article 50(2) marking applies from 2 December 2026.
    • New systems (placed on the market from 2 August 2026 onward): Article 50(2) marking applies immediately from placement.
    • Article 50(1), 50(3) and 50(4) duties: unaffected, all apply from 2 August 2026 for every system in scope.

    The same Digital Omnibus package also postpones application of the AI Act’s high-risk system requirements — Annex III stand-alone systems now apply from 2 December 2027, and Annex I product-embedded systems from 2 August 2028. These are separate rules on an entirely separate track from Article 50 transparency, and conflating the two — as some commentary has done — materially understates how soon the watermarking duty actually bites.

    The Code of Practice on Transparency of AI-Generated Content

    Article 50(2) compliance is operationalised through the AI Office’s Code of Practice on Transparency of AI-Generated Content. A first draft was published in December 2025, with a further draft circulated in May 2026 as the trilogue concluded. The European Commission’s Digital Strategy portal lists the Code among its active transparency-obligation guidance as of June 2026.

    The technical benchmark most frequently cited in industry guidance for machine-readable marking is C2PA Content Credentials, a provenance specification backed by major generative-AI and platform providers. Whichever technical route a provider chooses, the compressed runway between a finalised Code of Practice and the 2 December 2026 application date means marking, metadata-embedding and detector-tooling work needs to start now rather than after final guidance lands.

    Answer-first questions

    Has the AI Act watermarking deadline been delayed?

    Yes. Article 50(2) of the EU AI Act, which requires machine-readable marking of AI-generated synthetic content, moves from 2 August 2026 to 2 December 2026 for systems already on the market, under the Digital Omnibus trilogue agreement reached 7 May 2026.

    What is Article 50 of the AI Act?

    Article 50 is the AI Act’s transparency article. It sets four separate obligations: disclosing AI interaction, marking synthetic content, labelling deepfakes, and labelling AI-generated public-interest text — each with its own scope and, now, its own timetable.

    Does the delay affect the AI chatbot disclosure rule?

    No. Article 50(1), which requires providers to inform users they are interacting with an AI system such as a chatbot, is not delayed and continues to apply from 2 August 2026, unchanged by the Digital Omnibus.

    What is the Code of Practice on Transparency of AI-Generated Content?

    It is the AI Office’s guidance document operationalising Article 50 compliance, first drafted in December 2025 with further drafts through mid-2026. It is the practical reference providers use to meet the machine-readable marking requirement ahead of the 2 December 2026 deadline.

    Implications for research offices and publishers

    Institutions running AI-governance or research-integrity functions should treat this as a compliance-tracking, not a compliance-relief, event. Two separate dates now sit on the same calendar entry that many trackers previously listed as a single 2 August 2026 milestone. Research administration teams responsible for institutional AI-use policies, and publishers building AI-content-disclosure workflows alongside existing authorship-disclosure practices, need to record both dates and both scopes distinctly rather than treating “the AI Act deadline” as one event.

    • Update institutional compliance calendars to show 2 August 2026 (disclosure/labelling duties) and 2 December 2026 (marking duty for existing systems) as separate entries.
    • Distinguish the Article 50(2) watermarking delay from the much longer high-risk system postponement (2027/2028) when briefing leadership — the two are unrelated in scope and timing.
    • Track the AI Office’s Code of Practice finalisation, since the technical detail of “machine-readable” marking is defined there, not in the Regulation’s text.

    For institutions already documenting AI-content-disclosure alongside research-administration compliance tracking, the practical task is unchanged in substance and compressed in time: providers and deployers still need working marking and labelling capability, just against a marginally later date for one specific obligation.

    What happens next

    The Digital Omnibus text still requires formal endorsement and legal-linguistic revision before Official Journal publication, expected ahead of the original 2 August 2026 application date for the AI Act’s high-risk obligations. Once published, the 2 December 2026 date for Article 50(2) becomes fixed law rather than a trilogue compromise. Research offices, publishers and AI providers should treat the current text as the operative planning baseline, while watching for the AI Office’s final Code of Practice, which will determine exactly what “machine-readable” marking must look like in practice.