Alignment with the EU AI Act
D2.3 maps each PoliRural+ AI principle to an AI Act provision and to the tool that delivers it. Read it as a self-check, not a legal verdict.
How does PoliRural+ relate its AI principles to the EU AI Act, and what kind of document is that mapping?
D2.3's alignment map pairs each principle with AI Act provisions it reads as matching (for example transparency with Article 13, ethical use with human oversight and accuracy in Articles 14 and 15) and names the project mechanism behind it. It is the project's own self-assessment against a risk-based Regulation (EU) 2024/1689, not a legal conformity assessment.
- Regulation (EU) 2024/1689 entered into force on 1 August 2024 and applies in phases
- Risk-based: prohibited practices, high-risk systems, transparency obligations, minimal risk, plus rules for general-purpose AI models
- D2.3 maps principles to Articles 9, 13, 14 and 15 and to concrete tools
- The map is a self-assessment; article numbers here follow the adopted Regulation
Holding the principles against the EU AI Act
main tellingdepth: standard3 min markdown
The EU AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024 and its obligations apply in phases. It follows a risk-based approach: some practices are prohibited; high-risk systems must meet specific requirements; other systems carry transparency obligations, such as telling people that they are interacting with an AI system (Art. 50); minimal-risk uses carry no specific obligations. Separate rules cover general-purpose AI models. Which dates apply to which obligation is set in the Regulation and can change through later acts, so check the official text on EUR-Lex rather than a summary.
The alignment map (D2.3, Table 64)
| Principle | AI Act provision as cited in D2.3 | PoliRural+ mechanism |
|---|---|---|
| 1 Transparency & Explainability | Art. 13: users can interpret outputs and decide in an informed way | Advisor outputs are explainable; JackDaw references its sources |
| 2 Stakeholder Co-Creation | Art. 9: risk management and human oversight that involve stakeholder perspectives | RAPs co-designed through the Advisor; stakeholder foresight interviews |
| 3 Multi-Actor Involvement | a recital on involving civil society and public actors (as cited in D2.3) | MAAT supports quadruple-helix participation |
| 4 Iterative Feedback | Art. 9: risk management as a continuous, iterative process | VULTURE feedback loops and work-awareness journaling |
| 5 Ethical & Responsible Use | Arts. 14 and 15: human oversight, robustness, accuracy | Bias and hallucination risks flagged in Advisor and JackDaw |
| 6 Capacity Building | training users (D2.3 cites Art. 29 of the 2021 proposal; in the adopted text: AI literacy, Art. 4, and deployers’ duties, Art. 26) | Workshops on MAAT, JackDaw and the Advisor |
| 7 Inclusive Policy Design | Art. 5(1)(c): the ban on social scoring that leads to unjustified detrimental treatment | Inclusive foresight and RAP drafting |
| 8 Gender Equality & Social Justice | none mapped | Workshop and sub-project reporting |
How to read it
- It is a self-assessment. D2.3’s “Notes on Compliance Level” states that the Advisor and MAAT align with most requirements on risk management (Art. 9), data governance (Art. 10), transparency (Art. 13), human oversight (Art. 14) and accuracy and robustness (Art. 15). That is the project’s reading. Whether a particular deployment counts as high-risk depends on its use case as the Regulation defines it.
- Article numbers checked against the adopted Regulation. Articles 5(1)(c), 9, 10, 13, 14, 15 and 50 match the adopted text. Where D2.3 follows the numbering of the 2021 proposal (its Art. 29 on users), the table gives the adopted articles. Recital numbers cited in D2.3 were not re-checked and are left out.
- The eighth principle has no AI Act counterpart. It comes from the project’s Just Transition agenda, not from the Regulation.
- Planned steps are still plans. D2.3 lists as actions a public endorsement of the voluntary General-Purpose AI Code of Practice and a white paper for discussion with DG AGRI, DG CONNECT and the AI Office. It also contains a draft press release with the date left blank. This book does not confirm that these steps took place.
The useful part for a region is the right-hand column: every provision is matched to something a pilot can show, ask about, or fix. The AI ethics dashboard turns that column into a form.
Check yourself: An AI assistant to pre-screen local funding applications
A rural municipality runs a small grant scheme for local associations. Last year 120 applications arrived and two staff members spent weeks on eligibility checks. The mayor proposes an AI assistant to pre-screen this year’s applications before the funding committee meets.
A) Let the assistant reject clearly ineligible applications automatically, so the committee sees only the rest.
B) Let it only flag issues, each with its reason and source; keep every decision with staff and the committee, tell applicants that AI supports the screening, protect their data and train the staff first.
C) Buy a commercial tool whose vendor guarantees AI Act compliance, and rely on that guarantee.
D) Use no AI until every AI Act obligation is fully in force.
More than one answer can be defended. Choose one, then read where it works and where it breaks.
Feedback on each option
- A) This is where it breaks. A rejection decided by the tool removes the human oversight that D4.2 §10.2 names as a responsible-AI safeguard and that D2.3's alignment map links to Article 14 of the AI Act. The project's own open calls combined automated and manual eligibility checks with expert panels and consensus review, and beneficiaries described that process as transparent and fair (D6.3).
- B) The sound default. People still decide, applicants are told that AI supports the screening, and every flag shows its reason and source, as principles 1 and 5 and D4.2 §6.4 ask. Applicant data stay under a declared access level with a sensitive-data flag (D4.2 §4.4), and staff are trained before launch (principle 6). Whether this use counts as high-risk depends on how the AI Act defines its use cases, so check the official text before launch.
- C) A vendor's assurance does not answer the questions the ethics dashboard asks of the deployer: how decisions are explained, how bias and hallucination are handled, who was involved. Depending on one proprietary service without a fallback is a risk D4.2 scores at 6 of 9. Buying can be part of the answer, but not instead of your own checks.
- D) Defensible if the municipality cannot staff a human review or protect applicant data: then not deploying is the responsible choice. As a general rule it overshoots. The AI Act is risk-based rather than a ban, and D2.3 Annex B shows AI giving useful structure when people check it. Waiting also gives staff no practice in questioning AI output.
Related
- Builds on: Participatory AI Principles
Sources
- D2.3 Rural-urban data analysis, §2.4 PoliRuralPlus AI Principles vs. EU AI Act mapping (Table 64, Notes on Compliance Level); §3.2 Plan to strategically position PoliRuralPlus; §2.5–2.6 press release drafts — PDF
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), official text on EUR-Lex — EUR-Lex