Draft Guidelines on the implementation of the transparency obligations for certain AI systems
- Categories Case Studies
- Date May 10, 2026
The EU AI Act Transparency Obligations: A Case Study for M&E Professionals
Sector: Global Health & Development
Transparency Obligations · Article 50
1. Background: Why AI Transparency Matters for M&E
Under the EU AI Act (Regulation (EU) 2024/1689), Article 50 introduces binding transparency obligations for providers and deployers of certain AI systems. For Monitoring & Evaluation professionals working with EU-funded programmes — or whose AI-generated outputs (reports, dashboards, synthetic data) reach EU-based stakeholders — compliance is mandatory by 2 August 2026. This case study follows “HealthAccess International,” a medium-sized INGO that uses AI (Claude, ChatGPT, Power BI) across its M&E workflows and successfully adapted to the new legal landscape.
📌 Key fact from the Draft Guidelines (point 10):
“A third country provider of a generative AI system may be subject to the obligation laid down in Article 50(2) AI Act, even if the system is only marketed outside the EU, if the system’s outputs are intended to be used in the Union.”
2. Mapping Article 50 to Real-World M&E Tasks
The organization audited 12 AI-assisted M&E processes against the four transparency obligations.
| Provision | Obligation | M&E Example | Compliance action |
|---|---|---|---|
| Art. 50(1) | Inform about interaction with AI system | Chatbot for health workers, automated SMS surveys | Persistent label “AI Assistant — human oversight” |
| Art. 50(2) | Marking & detection of AI-generated content | Donor reports, synthetic community summaries | Visible footer + metadata watermark (C2PA) |
| Art. 50(4) a | Disclose deep fakes (image/audio/video) | AI-manipulated training photos (background edits) | No disclosure if minor editing (point 109) |
| Art. 50(4) b | Disclose AI-generated text on matters of public interest | AI-written policy briefs, public health alerts | Clear disclosure on first page + machine-readable label |
3. Article 50(1): Interactive AI Systems in M&E
HealthAccess used a custom Claude-powered assistant to guide enumerators through survey protocols. Initially, no disclosure was made that the assistant was AI. According to point 29-30 of the Guidelines, the provider must ensure notification at the latest at the first interaction.
4. Article 50(2): Marking AI-Generated Content in Donor Dashboards
The M&E team used ChatGPT to draft quarterly indicator performance narratives inside Power BI. Under the Guidelines (point 65-66), both machine-readable marking AND detectability are required. The NGO combined visible disclaimers with embedded metadata.
Footer on each report page: “AI-generated analysis — verified by M&E team on [date].”
Implemented C2PA content credentials + metadata watermarking for all final PDFs and images.
Guidance from point 74:
Technical solutions must be effective, interoperable, robust, reliable. Single technique is rarely sufficient; combination of labels + metadata is current state-of-the-art.
5. Article 50(4): Deep Fakes and AI Text on Public Interest
The NGO publishes AI-generated summaries of community feedback on its public website. Under point 123, such text “informs the public on matters of public interest”, therefore it must be labelled. Additionally, AI-manipulated images (background edits in training photos) were reviewed: because alterations were minor (cropping, lighting), they fell outside the deep fake definition (point 109).
6. Exception: Human Review & Editorial Control (Art. 50(4), subpara.2)
The NGO publishes a weekly newsletter on maternal health policy, AI-generated but reviewed by a senior editor. Because substantive human review + named editorial responsibility was documented, the content qualified for the exception (no disclosure required). However, superficial spell-checking would not suffice (point 126-127).
7. Timeline to Compliance (2 August 2026) and Penalties
Point 141: Article 50 applies from 2 August 2026. Non-compliance can lead to fines up to €15 million or 3% of worldwide annual turnover (point 140). HealthAccess started internal audits in January 2025, completed transparency playbook by Q1 2026, and will be fully compliant by July 2026.
| Milestone | Action |
|---|---|
| Q2 2025 | Inventory of all AI tools & outputs reaching EU stakeholders |
| Q3 2025 | Classification under Art. 50(1),(2),(4) |
| Q1 2026 | Implementation of marking solutions (visible + machine-readable) |
| 2 August 2026 | Full compliance deadline |
8. Key Lessons for the EvalCommunity
Add disclosures at first interaction or exposure – not buried in terms.
Both are required under Art. 50(2).
But material alteration of meaning triggers labelling.
Named person + meaningful review, not cosmetic.
9. Frequently Asked Questions for M&E Teams
10. Interplay with DSA, GDPR and Sector Standards
Point 88-92 highlight that marking of AI content must be compliant with data protection principles (data minimisation, security). Deployers should also respect existing media and consumer protection laws. For M&E, this means consent forms should mention AI-generated analyses, and personal data within AI outputs must be anonymised.
11. Codes of Practice as a Compliance Pathway
Under point 135-136, adhering to a code of practice deemed adequate by the AI Office will be the simplest way to demonstrate compliance. HealthAccess joined a pre-standardization working group for the development sector. Signatories benefit from increased trust and reduced supervisory burden.
📢 Final takeaway for evaluators:
The EU AI Act’s transparency obligations are not optional for organizations using AI in M&E with EU connections. Implementing visible, machine-readable markings, reviewing deep fake definitions, and documenting human oversight are practical steps that also improve donor confidence and methodological rigour.
Master AI transparency & compliance for M&E
Explore templates, checklists, and accredited training on responsible AI in evaluation.
Case study based on the Draft Guidelines on the implementation of the transparency obligations under Article 50 of Regulation (EU) 2024/1689 (AI Act). For educational purposes. Always refer to official legal texts and professional legal advice.
The courses and articles are developed by a team of experienced evaluators, collaborators, authors, and software developers, guided by Fation Luli. EvalCommunity Academy combines practical expertise in Monitoring & Evaluation and International Development with the latest advances in AI to create high-quality, accessible, and practical learning experiences for professionals worldwide.