Research synthesis · AI, democracy & civic resilience

AI and democracy: separating the vulnerabilities, naming the harms, governing the media

Vajjhala’s work on AI and democracy treats the exposure of a democracy as a property of its institutions rather than of its technology, separates five distinct categories of unethical AI use so that each can be matched to its own regulatory response, and shows that the research literature on synthetic media has so far built in almost no consent or provenance safeguards.

A short narrative review of research by Narasimha Rao Vajjhala and co-authors in this area. Every claim carries an APA citation that links to the publication’s own page (abstract, key findings, DOI); the full references are listed at the end. Updated .

Synthesis

Narasimha Rao Vajjhala’s work on artificial intelligence and democracy starts from a distinction that public debate usually collapses: not every democracy is exposed in the same way, and not every harm is the same harm. The volume he edited with Jacob Udo-Udo Jacob sets out a framework its editors call AI-Enhanced Reflexive Control, and distinguishes underdeveloped, strained and stressed democracies as separate vulnerability types rather than degrees of a single weakness (Jacob & Vajjhala, 2025). The argument is carried by case material from Central and Eastern Europe and beyond — the annulment of Romania’s 2024 election, disinformation networks operating in Bulgaria, and democratic innovations in Kenya and the Philippines.

The same volume separates five categories of unethical AI use — disinformation, electoral interference, manipulation of human rights, cultural exploitation and invasion of privacy — on the reasoning that a regulatory response fitted to one will not fit the others. Electoral integrity is treated across three of its chapters: regulatory responses to AI in modern elections, AI solutions for election integrity in developing democracies, and algorithmic fairness and bias mitigation in the use of social media for political campaigning. Its applied chapters extend to computer vision in public policy and to AI-driven tools as democratic equalisers for access to justice. Vajjhala co-wrote the introduction and, with Kenneth David Strang, the chapter on unethical AI in democratic systems.

Where that volume works at the level of institutions, the synthetic-media research works at the level of the artefact. Reviewing 92 talking face generation studies published between 2016 and 2025, Salahudeen et al. (2026) found that only one system embedded any technical safeguard — a post hoc watermark that was not C2PA-compliant — and that none implemented consent or provenance, although the EU AI Act’s Article 50 transparency obligations for synthetic content have applied since 2 August 2026. They proposed the RTFG-STS Framework, combining cryptographically signed C2PA provenance with risk-tiered consent, a deepfake detectability audit and human oversight. The finding matters for electoral integrity because provenance, not detection, is what survives contact with a fast-moving information environment.

A methodological thread connects this work to the security research. Vajjhala et al. (2015) showed what open data on contested events can and cannot support: correspondence analysis of 125,087 global terrorism records related nine attack methods to thirteen world regions, with the first two dimensions capturing 94.5% of total inertia, while the overall association between attack type and region proved low. The caution generalises — open incident and platform data will carry a strong descriptive structure and a weak inferential one, and claims about information operations need to respect that difference.

The practice side runs alongside the publications rather than inside them: Vajjhala was Principal Investigator of DemPro, a civic-participation platform built at the American University in Bulgaria’s Center for Information, Democracy and Citizenship for young people in Bulgaria and Eastern Europe, and spoke on the poisoning of large language models at the Pravda Disinformation Network roundtable in Sofia. Both are described on the projects page and under invited talks.

Key claims with sources

  1. Automated manipulation of democracies is framed as reflexive control enhanced by AI rather than as disinformation alone, with underdeveloped, strained and stressed democracies treated as distinct vulnerability types (Jacob & Vajjhala, 2025).
  2. Five categories of unethical AI use are separated — disinformation, electoral interference, manipulation of human rights, cultural exploitation and invasion of privacy — each matched to its own regulatory response (Jacob & Vajjhala, 2025).
  3. Of 92 talking face generation studies published in 2016–2025, only one embedded any technical safeguard and none implemented consent or provenance, although EU AI Act Article 50 transparency obligations for synthetic content have applied since 2 August 2026 (Salahudeen et al., 2026).

The publications and what each contributes

PublicationMethodCore finding or contribution
Salahudeen et al. (2026)
Systems
Systematic literature review (Kitchenham and Charters; reported to PRISMA 2020) of 92 primary studies published 2016–2025, screened from 18,343 records in six repositories over four search iterations, with a five-criterion quality assessment and qualitative narrative synthesis across six research questions.Across 92 primary studies (2016–2025), Diffusion Transformers and Gaussian Splatting account for about 36% of studies dated 2024 or later while GAN use has fallen below 3%; training data are demographically skewed and under-documented, and the dominant metrics (PSNR, SSIM, SyncNet) do not measure what deployment requires.
Jacob and Vajjhala (2025)
Chapman and Hall/CRC (edited book)
Edited volume of 11 chapters by international contributors; chapter methods range from conceptual and regulatory analysis to country case studies and applied technical treatments.The volume gathers eleven chapters that set out a framework its editors call AI-Enhanced Reflexive Control, distinguish underdeveloped, strained and stressed democracies as distinct vulnerability types, and work through five categories of unethical AI use: disinformation, electoral interference, manipulation of human rights, cultural exploitation and invasion of privacy.
Vajjhala et al. (2015)
2015 3rd International Conference on Future Internet of Things and Cloud (FiCloud), Rome, Italy
Positivist exploratory case study applying simple correspondence analysis in SPSS 21 to an open big dataset of 125,087 global terrorism records (1970–2013), relating nine attack methods to 13 world regions.Simple correspondence analysis in standard statistical software, applied to 125,087 global terrorism records (1970–2013), related nine attack methods to 13 world regions with two dimensions capturing 94.5% of inertia — demonstrating an accessible method for visualizing open big data.

Key terms in this area

Talking face generation (TFG)
The synthesis of photorealistic, temporally coherent video in which a face speaks in synchrony with a given audio signal, typically from a single portrait image and an audio clip. — as used in Salahudeen et al. (2026)
Diffusion Transformer (DiT)
A diffusion model in which the usual U-Net denoising network is replaced by a transformer operating over latent tokens, so that output quality scales with model size and compute. — as used in Salahudeen et al. (2026)
Gaussian Splatting
A 3D rendering approach that represents a head or scene as learned 3D Gaussians and draws them by fast differentiable rasterisation instead of volume rendering, which makes real-time talking-head synthesis possible. — as used in Salahudeen et al. (2026)
Content provenance (C2PA)
A cryptographically signed manifest, specified by the Coalition for Content Provenance and Authenticity (C2PA v2.0), recording how a media asset was produced and bound to that asset, so that the authenticity and integrity of the claim can be verified. — as used in Salahudeen et al. (2026)
Deepfake Detectability Rate (DDR)
A metric proposed in this review: the proportion of generated frames correctly classified as synthetic by an ensemble of state-of-the-art deepfake detectors, intended to operationalise the machine-detectability required by EU AI Act Article 50(2). — as used in Salahudeen et al. (2026)
Systemic embeddedness
The sixth principle the RTFG-STS Framework adds to the five AI4People principles: the design and governance of a generative system should reference the concrete socio-technical context in which it is deployed, not the model in isolation. — as used in Salahudeen et al. (2026)

All key terms across the publications →

Standards, datasets and methods named in these studies

References

How to cite this synthesis

Please cite the original publications above for specific findings. To cite this overview itself:

Vajjhala, N. R. (2026, October 21). AI and democracy: separating the vulnerabilities, naming the harms, governing the media. Narasimha Rao Vajjhala. https://www.narasimharao.net/research/focus-areas/ai-democracy-civic-resilience/

Markdown version