On March 25, a landmark paper titled "AI Poses Risks to Democratic and Social Systems" was published by a group of 25 top scholars including 2018 Turing Award winner Yoshua Bengio, Berkeley's Stuart Russell, and Taiwan's Digital Minister Audrey Tang. Unlike mainstream AI safety research that focuses on model-level issues (hallucinations, toxic outputs, or AI takeover), this paper argues that a whole class of systemic risks has been overlooked—damage to democratic institutions and governance at scale.
Seven Failure Modes
The paper maps seven failure modes (T1–T7) along the governance feedback loop. At the public belief stage: Belief Homogenization (T1) occurs when most people use similarly trained models, compressing viewpoint diversity due to post-training methods like RLHF. Belief Intensification (T2) sees personalized AI assistants reinforce users' existing views through long-term memory, creating closed loops. Studies cited show GPT-4's persuasion rate increases by over 80% when it has access to users' sociodemographic data.
At the institutional processing stage: Bureaucratic Congestion (T3) allows anyone to generate vast volumes of plausible public comments at near-zero cost, overwhelming agencies. Information Flooding (T4) makes verification cost far exceed generation cost, burying the information ecosystem. At the accountability stage: Unreviewable Authority (T5)—AI's opacity, scale, and access barriers smash existing oversight. Normative Centralization (T6)—when governments adopt advanced AI models, developers' values get embedded in public infrastructure, shifting normative power from elected officials to a handful of developers. Power Concentration (T7) cuts across all stages, as AI substitutes human labor in economic, ideological, political, and military domains, eroding citizens' leverage for checks and balances.
Audrey Tang's Contributions: Taiwan's Experiments
Tang contributed key sections proposing participatory governance as a countermeasure. For T3, she advocates structured deliberation platforms that use dimensionality reduction to surface consensus rather than volume, combined with sortition (randomly selected citizen juries) to structurally prevent large-scale impersonation. For T4, she cites Taiwan's COVID-era tactic of "humor beats rumors"—authorities producing verified content within minutes to outrun misinformation. For T6, she points to emerging research on collective constitutional AI, where representative citizen panels draft AI constitutions via deliberation; models trained on these constitutions show comparable safety metrics with less bias, and the process should be federal to allow different polities different normative priorities.
The paper's most concrete case appears in recommendation R7 (invest in deliberative infrastructure for AI governance): In 2024, deepfake ads impersonating public figures flooded Taiwan's social media. The Ministry of Digital Affairs convened 447 randomly selected citizens across 44 virtual deliberation rooms, with an AI dialogue engine synthesizing proposals within a day. The resulting citizens' assembly focused on regulating actors and actions—platform liability for unauthorized deepfakes, mandatory labeling of unverified ads, and throttling non-compliant services—rather than content censorship. The bill passed with cross-party support; impersonation ads dropped 94% in a year.
Seven Recommendations and Core Findings
- R1 Develop multi-agent simulations to stress-test institutional resilience under large-scale AI participation.
- R2 Train models to support cognitive health—honest dissent and epistemic humility beyond mere harm avoidance.
- R3 Limit AI autonomy in governance scenarios; keep human accountability.
- R4 Establish Institutional Safety Levels (ISL) calibrated to AI capability.
- R5 Require AI systems in governance to keep decision logs and verify public participants.
- R6 Public AI procurement mandates interoperability and multi-vendor strategies to prevent normative monopoly by a single model family.
- R7 Invest in deliberative governance infrastructure to make democratic participation channels more manipulation-resistant.
The paper directly counters two common objections: that society will adapt naturally (but AI centralizes economic rents while eroding the political and organizational capacity for self-correction), and that AI alignment alone suffices (but failure modes like cost-asymmetric flooding attacks and citizen leverage dilution persist even with perfect alignment). The conclusion notes institutional resilience need not be built from scratch—current civic tech initiatives already prove structured deliberation works at national scale—but equipping these tools for AI governance remains a wide-open research challenge.