Towards trustworthy misinformation detection

Abstract

The detection of fake news and artificial intelligence (AI)-generated synthetic content has become one of the central challenges of contemporary information science. While deep learning approaches—particularly transformer-based language models and convolutional networks—have achieved strong performance on established benchmarks, their reliance on centralised training pipelines renders them ill-suited to the institutional, regulatory, and adversarial realities of real-world deployment. Simultaneously, the emergence of AI content watermarking as a regulatory instrument—culminating in the EU AI Act Article 50 Code of Practice published in December 2025—establishes a new technical landscape in which provenance signals should be primary rather than supplementary detection inputs. This chapter examines the convergence of three paradigms that together address the principal limitations of existing detection approaches: Federated Learning (FL), which enables privacy-preserving collaborative model training across distributed institutional participants without raw data exchange; blockchain technology, which provides verifiable, tamper-proof aggregation and audit infrastructure replacing the trusted central server assumption inherent in conventional FL; and multi-layer AI watermark verification, which integrates Coalition for Content Provenance and Authenticity (C2PA) Content Credentials, neural watermarking schemes such as Google SynthID, and cryptographic embedding approaches such as MetaSeal as first-class classification signals alongside content-based behavioural features. We systematically survey the state of the art across these three domains, analyse their theoretical and practical complementarities, and propose a unified architectural design that jointly addresses privacy preservation, Byzantine resilience, and standards-compliant provenance verification. A comprehensive compliance analysis against the EU AI Act Article 50 Code of Practice demonstrates that the proposed design satisfies all applicable technical obligations ahead of the August 2026 enforcement deadline. We further discuss the fundamental limitations and open research problems at the intersection of these paradigms, including the watermark trilemma, GDPR tensions with blockchain immutability, and the detection gap for unmarked synthetic content.