The Verification Crisis: Expert Perceptions of GenAI Disinformation and the Case for Reproducible Provenance

The growth of Generative Artificial Intelligence (GenAI) has shifted disinformation production from manual fabrication to automated, large-scale manipulation. This article presents findings from the first wave of a longitudinal expert perception survey (N=21) involving AI researchers, policymakers, and disinformation specialists. It examines the perceived severity of multimodal threats – text, image, audio, and video – and evaluates current mitigation strategies. Results indicate that while deepfake video presents immediate “shock” value, large-scale text generation poses a systemic risk of “epistemic fragmentation” and “synthetic consensus,” particularly in the political domain. The survey reveals skepticism about technical detection tools, with experts favoring provenance standards and regulatory frameworks despite implementation barriers. GenAI disinformation research requires reproducible methods. The current challenge is measurement: without standardized benchmarks and reproducibility checklists, tracking or countering synthetic media remains difficult. We propose treating information integrity as an infrastructure with rigor in data provenance and methodological reproducibility.

💡 Research Summary

The paper “The Verification Crisis: Expert Perceptions of GenAI Disinformation and the Case for Reproducible Provenance” investigates how generative artificial intelligence (GenAI) is reshaping the disinformation ecosystem and how experts perceive the associated risks and mitigation options. Using a purposive snowball sampling strategy, the authors recruited 21 senior experts—including AI researchers, policymakers (including a European Commission member), journalists, and industry executives—between July and December 2025. The survey measured perceived threat severity for four modalities (text, image, audio, video) on a 7‑point Likert scale and evaluated the effectiveness of five mitigation strategies (media literacy, C2PA provenance standards, regulation, platform interventions, and technical detection).

Key Findings on Threat Perception

• Deep‑fake video received the highest mean threat rating (6.4/7), reflecting its “shock value” and the strong human reliance on visual evidence. Experts noted that while a single convincing video can destabilize public trust, it is often more amenable to debunking once identified.
• AI‑generated text scored a close second (6.1/7) and was highlighted as the most systemic danger. Respondents described a phenomenon they term “synthetic consensus” or “astroturfing at scale,” where massive volumes of near‑human‑like text flood online discourse, eroding the baseline of factual discernment and fragmenting shared epistemic foundations.
• Voice cloning (audio) (5.9/7) was flagged for its immediate fraud potential, especially in voice‑based payment scams and impersonation attacks.
• AI‑generated images (5.8/7) were seen as less threatening than text or video but still capable of calibrating public expectations about visual manipulation.

Assessment of Mitigation Strategies
Technical detection tools were rated poorly (mean effectiveness ≈ 3.4/7). Experts attributed this to the “black‑box” nature of commercial detectors: lack of transparency about model architecture, training data, and evaluation protocols prevents reproducibility and thus trust. In contrast, provenance standards such as C2PA and regulatory frameworks received relatively high scores (≈ 5.8/7 and 5.6/7 respectively), indicating a preference for transparent, standards‑based approaches that can be audited and enforced across platforms. Media literacy was viewed as useful but insufficient on its own, while platform‑level interventions were seen as variable depending on corporate willingness.

Methodological Rigor and Reproducibility Emphasis
The authors embed their empirical work within a broader argument that the disinformation field suffers from a reproducibility crisis. They adopt the Momeni‑Khan reproducibility checklist and Bleier’s Methods Hub concept to propose a “R2CASS” (Reproducible Resistance to the Crisis of Authenticity and Synthetic Sources) infrastructure. This includes:

1. Publishing anonymized CSV datasets, survey instruments, and analysis scripts with exact library versions and random seeds.
2. Containerizing detection pipelines so that performance can be re‑executed on new data.
3. Linking provenance metadata to a structured knowledge graph (e.g., ClaimsKG, TeleScope) to track narratives across platforms and languages.

By providing a reproducibility package that follows the checklist, the authors demonstrate how future research on GenAI‑driven disinformation can be validated, compared, and built upon, thereby supporting evidence‑based policy making.

Critical Reflections
The study’s sample is heavily weighted toward European experts, which may bias the perception of regulatory effectiveness and under‑represent challenges faced in the Global South. The small N (21) limits statistical generalizability; however, the authors explicitly prioritize depth of insight over inferential statistics. The reliance on self‑reported effectiveness of detection tools may not capture actual system performance, suggesting a need for empirical benchmarking in future work.

Implications
The paper argues that without transparent provenance mechanisms and reproducible research pipelines, policy interventions risk being speculative and potentially counter‑productive. It calls for an “information integrity infrastructure” that treats provenance as a public good, akin to internet routing standards, and for coordinated international efforts to standardize benchmarks, share datasets, and enforce provenance metadata. In doing so, the authors position reproducibility not merely as an academic virtue but as a strategic defense against the accelerating threat of GenAI‑generated disinformation.