⚡ Quick Summary
Published by the Forum on Information and Democracy, this policy framework sets out a roadmap for democratic control of AI in the information and communication space. It is the output of a 14-member Working Group launched on 28 September 2023 and co-chaired by Laura Schertel Mendes and Jonathan Stray, whose rapporteurs consulted more than 150 people worldwide.
The framework is organised in four chapters: developing and deploying safe and responsible AI systems; liability and accountability regimes; incentivising ethical AI; and AI governance and oversight. It frames AI as a public good and combines a risk-based approach with a rights-based one, protecting the right to be informed, to an explanation, to challenge a machine-generated outcome and to non-discrimination.
Key recommendations include participatory processes to set rules for dataset provenance and curation, training-data transparency through a searchable database, bias and language impact assessments before deployment, systemic risk assessment and third-party conformity assessment for medium- and high-risk systems, a “right of recommendability” tied to cryptographic signatures, provenance and authenticity standards, fault-based and strict liability regimes with a reversed burden of proof, an AI Ombudsman, tiered transparency for regulators and vetted researchers, and a tax on AI companies to fund public-interest alternatives.
🧩 What’s Covered
The extraction covers the front matter and key recommendations, the glossary and introduction, Chapters 1 and 2 in full, and the opening of Chapter 3.
- Front matter and key recommendations: forewords by co-chairs Laura Schertel Mendes and Jonathan Stray and by the Forum's executive director Michael Bąk; a consolidated recommendation list grouped around AI companies and entities, regulation, trust in the information space, accountability for harms, and independent oversight.
- Working group, glossary and introduction: the 14 Working Group members and the rapporteur team; around forty definitions from accountability sandbox to watermarking; a survey of AI and information creation, dissemination and consumption and of the regulatory landscape, from the US Executive Order and Bletchley Declaration to Brazil's Bill 2338 and Canada's AIDA.
- Chapter 1 – safe and responsible AI systems: training datasets (inclusive curation, access to high-quality data, bias, low-resource languages, privacy, media compensation), human labeling, content moderation and ranking, optimization objectives, content authenticity and provenance, then red-teaming, pre-release risk assessments and post-release monitoring.
- Chapter 1 tables: Table 1.1 maps attacks on the information space, the methods and possible defences; Table 1.2 lists systemic harms and the factors for assessing their likelihood and severity.
- Chapter 2 – liability and accountability: duties of actors in the AI value chain, multilevel transparency, contractual liability, and liability regimes by function in Table 2.1; platforms hosting AI-generated content; synthetic content and entities; generative AI in politics (Table 2.2); complaint handling and redress following the UN Guiding Principles on Business and Human Rights.
- Chapter 3 – incentivising ethical AI (partial): opening sections on codes of conduct and on certifications and ratings, including the Fair Trade analogy; the extraction stops mid-chapter, so the remaining sections and Chapter 4 are not covered.
💡 Why it matters?
The framework is aimed at people who must turn AI principles into enforceable practice. Chapter 1 supplies design-level measures — participatory dataset rules, bias and language impact assessments, tiered training-data disclosure, broader red-teaming, model cards, output moderation — that engineering and trust-and-safety teams can adopt. Chapter 2 gives regulators and legal teams a function-based method for assigning fault-based or strict liability and for reversing the burden of proof. Chapter 3 covers non-regulatory levers such as codes of conduct, certification and public procurement. The recommendations reference the DSA, the draft EU AI Act and the G7 Hiroshima Process, so they can be read alongside those instruments.
❓ What’s Missing
The input is partial: the document runs to 141 pages, and the closing sections of Chapter 3, all of Chapter 4 on governance and oversight, the acknowledgements and the bibliography are not in the extraction, so the oversight, funding and AI-literacy proposals could not be reviewed. The framework itself states that it is not a negotiated consensus — it reflects “the rapporteur team's best efforts” and unanimity was not sought — and several questions are explicitly left open, including whether strict liability should apply to fundamental-rights violations and how surveillance-based advertising should be legislated. Much of the liability analysis is anchored in the draft EU AI Act and the DSA as they stood in early 2024.
👥 Best For
Policy advisers and legislators drafting AI, election or platform rules; legal and compliance teams allocating liability across the AI value chain; trust-and-safety and engineering teams designing data curation, red-teaming, provenance and output-moderation controls; and civil society, journalism and research organisations arguing for oversight, transparency and compensation rights.
📄 Source Details
AI as a Public Good: Ensuring Democratic Control of AI in the Information Space — Policy Framework, published by the Forum on Information and Democracy, February 2024, 141 pages, English. Authorship is credited to a rapporteur team led by Viviana Padelli and to a 14-member Working Group co-chaired by Laura Schertel Mendes and Jonathan Stray. The input was a partial text extraction of 80 pages, covering roughly pages 1–80 and ending mid-way through Chapter 3; Chapters 4, the acknowledgements and the bibliography were unavailable.
