The Algorithmic Leviathan: Artificial Intelligence, State Power, Corporate Surveillance, and the Erosion of Democratic Accountability

Full report

On this page

Executive Overview: The Structural Reconfiguration of Institutional Power

The rapid advancement and deployment of artificial intelligence (AI) systems—ranging from automated decision-making (ADM) algorithms and biometric identification networks to large-scale generative models—represent a fundamental reconfiguration of how institutional power is organized, exercised, and contested1. This transformation occurs against a backdrop of global economic volatility, labor market restructuring, and declining public trust in democratic governance4. Rather than acting as a neutral tool for administrative optimization, advanced AI technologies alter the balance of power between individual citizens, private corporate monopolies, and the state2. Historically, state authority operated through human bureaucracies governed by statutory procedural rules, judicial oversight, and administrative discretion1. Similarly, corporate power was constrained by contractual relationships, physical management structures, market competition, and labor organization2. The integration of AI into public administration and private enterprise introduces automated computational systems that centralize control, obfuscate decision-making logic, and scale surveillance to pervasive levels2. This concentration of power operates through three structural mechanisms:

1. Asymmetric Visibility: State agencies and technology platforms gain real-time insight into individual behavior, financial transactions, communications, and physical movements, while remaining opaque regarding their own internal algorithmic logic, pre-training data, and operational parameters4.

2. Automated Administrative Enforcement: Delegating legal, regulatory, and managerial decisions to algorithmic frameworks replaces administrative discretion and human empathy with rigid, low-cost compliance rules that disproportionately penalize vulnerable populations2.

3. Infrastructure Consolidation: A small cartel of technology hyper-scalers controls the computational infrastructure, specialized hardware, data repositories, and foundation models required to deploy modern AI, transforming private corporations into indispensable operational partners—and unchecked gatekeepers—of state functions7.

As a result, traditional institutional safeguards designed to ensure democratic accountability—such as judicial review, notice and hearing rights, collective bargaining, and public petitioning—face structural degradation1. Rebalancing this institutional asymmetry requires examining the operational reality of AI systems across law enforcement, public administration, labor oversight, and information governance, while establishing actionable legal and technical safeguards1.

State Surveillance and Automated Administrative Governance

Biometric Surveillance, Facial Recognition, and Predictive Law Enforcement

The deployment of biometric surveillance and predictive analytics by law enforcement represents a major expansion of state investigative capacity11. Modern police operations increasingly rely on facial recognition technology (FRT) fed by massive, unconsented web scraping, alongside automated license plate readers (ALPR), acoustic gunshot detection networks, and predictive policing algorithms13. A primary driver of this transformation is the commercialization of biometric databases by private entities13. Vendors like Clearview AI compiled billions of images scraped from public websites and social media platforms, providing law enforcement agencies with biometric identification capabilities that circumvent traditional warrant requirements and legislative oversight13. The systemic risks of these technologies are demonstrated by documented false-positive matches and wrongful arrests16. In City of Detroit, law enforcement relied on a low-quality surveillance still processed through facial recognition software, resulting in the wrongful arrest of Robert Williams16. Similarly, in private retail environments, loss-prevention workflows utilizing facial recognition have misidentified innocent patrons, leading to unwarranted police interventions and civil rights litigation16. Beyond post-hoc identification, predictive policing systems utilize historical arrest records to allocate police patrols, calculate recidivism risk, and flag individuals for pre-arrest intervention11. Algorithmic risk tools like COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) score criminal defendants on their likelihood of reoffending11. However, because historical arrest data reflects historical over-policing and systemic socio-economic disparities, these algorithms reinforce institutional biases under the guise of mathematical objectivity4. When deployed to monitor public assemblies or demonstrations, real-time facial recognition creates systemic chilling effects on constitutional guarantees of free speech and association, enabling pre-emptive state intervention against political dissidents5.

Automated Administrative Decision-Making and the Rupture of Due Process

While biometric tools alter physical policing, automated administrative decision-making (ADM) systems silently alter the social safety net and regulatory state1. Modern public administration delegating eligibility determinations, tax audits, child welfare assessments, and fraud detection to algorithmic frameworks frequently replaces procedural due process with automated efficiency2. When improperly designed and deployed without human oversight, these systems generate widespread administrative injustice1. The legal and institutional failures of automated administrative governance are illustrated by three high-profile international case studies:

• Australia's "Robodebt" Scheme (Online Compliance Intervention): Implemented by the Department of Human Services, Robodebt utilized an automated algorithm to calculate welfare overpayments by averaging historical annual tax data across biweekly pay periods1. The system inverted the traditional legal burden of proof, automatically issuing computerized debt notices and demanding that welfare recipients prove they owed nothing1. The scheme raised over \$1.7 billion in alleged debts against hundreds of thousands of vulnerable citizens, leading to extreme financial distress1. Subsequent federal class-action litigation resulted in consent orders declaring the income-averaging methodology irrational and unlawful, culminating in a Royal Commission that exposed severe systemic failures in administrative culture, executive accountability, and legal compliance1.
• The Netherlands' System Risk Indication (SyRI): Designed by the Dutch government to detect social benefits, taxation, and labor fraud, SyRI linked and analyzed data across multiple public authorities4. SyRI did not automatically issue binding decisions; instead, it used secret risk-profiling algorithms to generate watchlist leads for administrative investigators, systematically targeting low-income neighborhoods4. In 2020, the District Court of The Hague invalidated the governing legislation under Article 8 of the European Convention on Human Rights (right to respect for private and family life)4. The court ruled that SyRI lacked basic algorithmic transparency, preventing affected citizens from understanding or challenging their risk scores, and generated disproportionate interferences with privacy that failed the test of democratic necessity4.
• Michigan’s Integrated Data Automated System (MiDAS): Designed to automate unemployment fraud determinations, MiDAS processed data without mandatory human oversight, issuing thousands of automated fraud accusations with error rates exceeding 90 percent16. The system automatically imposed aggressive monetary penalties and initiated wage garnishments without providing claimants adequate notice or an opportunity to contest the determination16. In Bauserman v. Unemployment Insurance Agency, state courts addressed systemic due process violations, highlighting how delegating adjudicative authority to unverified code strips citizens of fundamental procedural protections16.

These operational failures extend into the private sector, where administrative AI systems act with pseudo-governmental authority2. Class action litigation filed in late 2023 alleged that major health insurers utilized predictive algorithms (such as UnitedHealth's nH Predict) to systematically override human physician evaluations and deny post-acute care to elderly patients, relying on error-prone statistical models to maximize cost savings under the blanket authority of automated processing2.

Corporate Surveillance, Algorithmic Management, and the Future of Labor

Workplace Monitoring, Continuous Metrics, and Algorithmic Control

The workplace has become a primary testing ground for pervasive AI surveillance and automated decision-making6. Algorithmic management—the practice of using algorithms, sensors, and data analytics to supervise, evaluate, allocate tasks, and discipline workers—has expanded from gig-economy platforms into traditional corporate offices, logistics centers, healthcare facilities, and remote working environments6. Modern algorithmic management relies on continuous monitoring systems, including keystroke logging, real-time webcam tracking, desktop sentiment analysis, wearable GPS trackers, telematics in logistics fleets, and AI-driven task allocation engines10. These technologies generate real-time performance scores based on customer ratings, idle-time indicators, and keystroke frequencies6. This dynamic shifts the fundamental nature of managerial authority6. Traditional management relied on human supervisors capable of context-aware discretion, negotiation, and flexibility1. In contrast, algorithmic management enforces rigid metrics that treat human workers as mechanical units within an optimized software pipeline6. Micro-surveillance increases physical strain, psychological stress, and workplace precarity6. Automated performance evaluation engines routinely issue punitive warnings or reduce shift allocations based on subtle metric deviations, without providing workers access to the underlying data or an opportunity to offer contextual explanations6.

Automated Deactivation and the EU Platform Work Directive

The most extreme manifestation of algorithmic management is "automated deactivation"—the unilateral termination of a worker's account or employment access by an algorithm without human intervention10. In gig economy platforms, algorithms automatically disconnect workers whose acceptance rates drop below arbitrary thresholds or who receive negative customer ratings, effectively stripping them of their livelihood without notice, clear reasons, or procedural appeal6. In response to these systemic abuses, the European Union adopted Directive (EU) 2024/2831 on improving working conditions in platform work (the EU Platform Work Directive)6. The Directive introduces major legislative safeguards designed to reframe algorithmic management within labor markets6:

• Legal Presumption of Employment: The Directive establishes a rebuttable legal presumption of an employment relationship when a platform exercises direction and control over work performance, shifting the legal burden onto platforms to prove that workers are genuinely self-employed22.
• Algorithmic Transparency: Platforms must inform workers about the existence and operation of automated monitoring and decision-making systems on their first working day, including descriptions of the categories of actions tracked and the parameters governing task allocation, rating systems, and promotion or demotion pipelines10.
• Mandatory Human Oversight and Right to Appeal: The Directive prohibits algorithms from making critical decisions—such as account suspension, termination, or contractual changes—without mandatory human review by qualified staff10. Workers are granted a statutory right to receive written explanations for automated decisions and to challenge those decisions before a human manager10.
• Strict Privacy and Data Enclosures: Platforms are explicitly barred from processing personal data regarding a worker's emotional or psychological state, private communications, political or union activities, or any data collected while the worker is not actively performing work10.

The Directive's two-year transposition window (ending December 2, 2026\) highlights a major global regulatory shift: recognizing that workers' fundamental rights cannot be preserved without hard legal limits on algorithmic surveillance and management21.

Synthetic Media, Generative Persuasion, and the Political Information Environment

Generative AI, Targeted Persuasion, and Information Pollution

The proliferation of generative AI models capable of creating realistic text, high-resolution synthetic imagery, hyper-realistic voice clones, and deepfake video content significantly degrades public discourse and political information integrity27. While information manipulation predates digital computation, generative AI alters the economics of synthetic content production27. Producing targeted, high-impact disinformation campaigns no longer requires specialized technical expertise or extensive budgets; advanced open-weight models and voice-cloning APIs allow actors to execute targeted persuasion at minimal cost27. Generative AI impacts political information environments through four primary mechanisms:

1. Automated Micro-Targeted Persuasion: Combining large language models (LLMs) with behavioral profiling data allows for the automated generation of tailored political messages designed to exploit individual psychological vulnerabilities, cognitive biases, or local grievances at scale27.

2. Identity Impersonation and Voice Cloning: Voice-cloning technologies can replicate the vocal characteristics of political figures, election administrators, or journalists using short audio samples, creating synthetic recordings designed to suppress voter turnout or provoke social instability27.

3. The "Liar's Dividend": As the public becomes increasingly aware that audio, video, and photographic evidence can be generated by AI, bad actors can routinely dismiss authentic, damning evidence of political corruption or state crimes as synthetic fabrications28.

4. Information Flooding and Epistemic Fatigue: Mass-generating low-cost synthetic news articles and social media posts degrades the public's ability to discern verified information, fostering cynicism, apathy, and withdrawal from civic participation27.

Documented Disinformation Incidents versus Hypothetical Capabilities

Evaluating the threat of AI-driven election manipulation requires distinguishing empirical, documented incidents from speculative scenarios28. While early public discourse often predicted catastrophic AI-generated political chaos, documented real-world deployments highlight specific operational tactical profiles:

Slovakia Parliamentary Election (September 2023)

Two days before the Slovakian parliamentary election, a deepfake audio recording was posted to social media, purporting to depict Michal Šimečka, leader of the pro-European Progressive Slovakia party, discussing ballot manipulation and tax increases with investigative journalist Monika Tódová27. The timing was deliberately calculated to exploit Slovakia's mandatory 48-hour pre-election "silence period"—a period during which political parties and media outlets were legally prohibited from issuing political statements or debunking claims27. Although fact-checkers subsequently identified the audio as an AI-generated synthesis, the clip went viral across encrypted messaging networks (such as Telegram) and social platforms, exploiting a low-trust information environment27. Although Progressive Slovakia subsequently placed second behind the populist Smer party led by Robert Fico, the incident demonstrated how synthetic media can exploit legal vulnerabilities in electoral silence windows27.

New Hampshire Democratic Primary (January 2024)

Thousands of New Hampshire voters received an automated robocall featuring a cloned voice of U.S. President Joe Biden28. The call instructed voters to stay home during the primary election and save their votes for the November general election28. The actor responsible acknowledged that creating the voice clone took under 20 minutes and cost less than \$1 using commercially available voice-synthesis software29. The rapid identification of the source led the Federal Communications Commission (FCC) to issue a legal ruling affirming that AI-generated voice cloning in robocalls violates the Telephone Consumer Protection Act (TCPA), establishing a key regulatory precedent against synthetic telephonic voter suppression29.

Infrastructure Concentration, Compute Oligopolies, and Asymmetric Power

The Compute and Data Infrastructure Oligopoly

The transformative power of contemporary AI relies on massive computational capacity and expansive data infrastructure7. Modern foundation models demand specialized computational hardware (such as advanced GPU clusters), ultra-high-speed networking, petabyte-scale data pipelines, and gigawatt-scale power installations8. Consequently, technical capabilities are highly concentrated within a small cluster of public technology monopolies and cloud hyper-scalers7. This concentration creates a structural barrier to entry, consolidating AI development within an oligopoly consisting of dominant cloud infrastructure providers and heavily capitalized research labs7. Public institutions, academic research centers, civil society organizations, and small-to-medium enterprises are increasingly forced to operate as clients dependent on closed commercial APIs8. This structural reality yields profound socio-political implications:

• Monopolization of Foundational Architecture: The basic design choices, safety alignment guardrails, political neutrality boundaries, and commercial terms of AI systems are dictated by private corporate executives rather than public democratic processes7.
• Data Enclosures and Epistemic Asymmetry: Technology hyper-scalers absorb vast swaths of public digital discourse, scientific literature, and creative works into proprietary pre-training datasets8. Once processed into proprietary model weights, this collective human knowledge is monetized and rented back to the public8.
• Regulatory Capture and Dependence: State agencies—including intelligence, defense, law enforcement, and welfare departments—frequently lack the internal infrastructure and technical expertise needed to build, maintain, or audit advanced AI1. As a result, public governance functions are routinely outsourced to private contractors, granting commercial vendors unaccountable leverage over public policy and state execution1.

Asymmetric Visibility and Democratic Degradation

The imbalance in computational power and data access creates an institutional power dynamic characterized by asymmetric visibility4. Sovereign state authorities and multinational technology firms deploy AI to render individual citizens, consumers, and workers entirely visible2. Personal communications, biometric identities, purchasing histories, location trajectories, and psychological profiles are logged, structured, and queried in real time5. Conversely, the operations of these institutions remain completely opaque1. State intelligence algorithms, predictive law enforcement models, corporate credit-scoring systems, and platform recommendation engines operate behind proprietary code, trade-secret protections, and institutional nondisclosure agreements2. When an individual is denied a visa, flagged for welfare fraud, terminated by an algorithm, or assigned a high recidivism score, they encounter a "black box" that defies meaningful judicial review or democratic accountability1. This opacity erodes democratic citizenship, reducing individuals from autonomous legal agents into passive data subjects managed by automated systems1.

Comparative Governance Regimes: United States, European Union, and China

Democratic accountability and technological regulation are shaped by national political systems7. Globally, three primary regulatory approaches have emerged: the market-driven sectoral approach of the United States, the rights-based omnibus framework of the European Union, and the state-centric security framework of China7.

The United States: Sectoral Patchwork and Enforcement Innovation

The United States lacks comprehensive federal legislation governing artificial intelligence, relying instead on a patchwork of sector-specific statutes, executive orders, state-level privacy acts, and administrative agency enforcement7. The primary federal enforcement mechanism for AI governance in the U.S. rests with the Federal Trade Commission (FTC) under Section 5 of the FTC Act, which prohibits "unfair or deceptive acts or practices"16. The FTC has pioneered an enforcement tool known as algorithmic disgorgement20. Where an entity collects data or trains an AI model through deceptive or unlawful means, the FTC has required companies to destroy not only the unlawfully collected data, but also the underlying algorithms, weights, and software models trained on that data20.

• In Everalbum, the FTC ordered the destruction of facial recognition models trained on illegally harvested user photos20.
• In Rite Aid, the FTC prohibited the company from deploying facial recognition technologies in retail locations for five years after identifying systemic false-positive misidentifications that disproportionately impacted women and racial minorities16.

At the state level, statutory frameworks like Illinois's Biometric Information Privacy Act (BIPA) enforce strict consent requirements and private rights of action for biometric harvesting, resulting in substantial class-action settlements against major technology vendors like Clearview AI14. Simultaneously, states like Texas have enacted comprehensive public-sector AI frameworks (e.g., Texas SB 1964\) mandating state agency AI inventories and risk assessments14.

The European Union: Rights-Based Omnibus Regulation

The European Union has established a comprehensive, rights-based legal framework anchored by the landmark EU Artificial Intelligence Act (Regulation (EU) 2024/1689), complemented by the General Data Protection Regulation (GDPR) and the EU Platform Work Directive2. Entering into force in August 2024, the EU AI Act classifies AI systems according to a four-tiered risk hierarchy35:

1. Unacceptable Risk (Prohibited Practices): Systems that pose an intolerable threat to citizens' rights are strictly banned37. This includes subliminal manipulation, social scoring by public authorities, real-time remote biometric identification in publicly accessible spaces for law enforcement (subject to narrow exceptions), biometric categorization to infer protected attributes, and emotion recognition in workplaces and educational institutions10.

2. High Risk: AI deployed in critical infrastructure, educational admissions, employment/worker management, essential public and private services (e.g., welfare and credit scoring), law enforcement, migration/border control, and administration of justice2. High-risk systems must undergo mandatory conformity assessments, implement continuous risk management systems, maintain audit-ready logging infrastructure, ensure human oversight, and satisfy stringent data governance standards1.

3. Specific Transparency / General-Purpose AI (GPAI): Foundation models and generative systems must satisfy transparency obligations, including technical documentation, compliance with EU copyright law, and detailed disclosures regarding training data content8.

4. Minimal Risk: Video games, spam filters, and non-sensitive applications, which operate under voluntary codes of conduct37.

China: State-Centric Security and Targeted Algorithmic Control

China’s AI governance model is state-led, prioritizing national security, social stability, and state control over information dissemination, while aggressively fostering industrial AI development9. Rather than enacting a single omnibus law, the Cyberspace Administration of China (CAC) has systematically promulgated targeted, scenario-specific regulations9:

• Provisions on Algorithmic Recommendation (2022): Mandates that recommendation algorithms must not be used to engage in unfair competition, manipulate public opinion, or generate economic disruption, and requires algorithm filing with the CAC for systems with "public opinion attributes or social mobilization capabilities"32.
• Provisions on Deep Synthesis (2023): Regulates deepfakes and generative synthetic media, requiring real-name user identity verification, explicit user consent to alter biometric features, and mandatory watermarking3.
• Interim Measures for Generative AI Services (2023): Applies to public-facing generative AI, mandating that generated content must adhere to "socialist core values," avoid content that subverts state power, pass pre-market security assessments, and utilize verified, lawful training data8.
• Mandatory AI Content Labeling Standards (GB 45438-2025): Effective late 2025/2026, this binding standard requires both explicit (visible) and implicit (embedded metadata) labels on all synthetic content, aggressively enforced by the CAC through app store removals and service suspensions3.
• Agentic AI & Companion Frameworks (2026): In mid-2026, Chinese authorities issued targeted frameworks specifically governing autonomous AI agents and interactive companion AI systems, imposing strict boundaries on autonomous decision execution3.

Economic Insecurity, Public Protest, and Technocratic State Responses

Labor Dislocation and the Precaritization of Work

The expansion of AI occurs alongside economic volatility, structural inflation, and labor market disruption5. Advances in generative AI and administrative automation expand technological substitution beyond traditional manual assembly into white-collar, creative, technical, and middle-management occupations6. This dynamic exacerbates income inequality, depresses real wage growth, and expands the "gigified" workforce6. As traditional employment contracts are unbundled, workers are increasingly pushed into platform-mediated, task-based arrangements6. In this environment, economic precarity functions as a mechanism of compliance: workers under intense financial pressure are less able to refuse algorithmic micro-surveillance, unsafe working speeds, or arbitrary pay adjustments dictated by management systems6.

Protest Surveillance, Predictive Management, and Civil Suppression

Economic insecurity frequently triggers civil unrest, mass public demonstrations, labor strikes, and social movements42. In response, state security agencies increasingly leverage surveillance technologies to monitor, manage, and neutralize political protest5. State agencies deploy predictive protest analytics, automated social media monitoring engines, ALPR networks, and drone surveillance equipped with real-time facial recognition13. These tools allow police authorities to identify political organizers, map activist social networks, track movement patterns prior to demonstrations, and deploy pre-emptive policing strategies11. This creates a serious chilling effect on democratic assembly5. When citizens know that attending a peaceful protest will result in permanent biometric indexing, facial logging, and potential inclusion on state risk watchlists (similar to the Dutch SyRI framework or police predictive databases), civic participation drops sharply4. AI-driven surveillance enables states to suppress political dissent without relying on overt physical force, neutralizing opposition through automated administrative pressure and legal harassment4.

Potential Democratic Benefits: Public Administration, Transparency, and Civic Governance

While advanced AI presents severe institutional risks, the underlying computational techniques can yield positive democratic outcomes if deployed within transparent, privacy-preserving, and accountable governance frameworks1. Modernizing public administration through ethically aligned AI can strengthen public capacity, optimize municipal resource distribution, and improve civic access1. Key areas of democratic benefit include:

• Enhanced Fraud Detection and Tax Enforcement: When bound by strict due process safeguards and human review, machine learning algorithms can analyze high-volume financial transactions to identify corporate tax evasion, cross-border trade fraud, and money laundering far more effectively than manual auditing1. Unlike automated welfare debt schemes that target vulnerable citizens, AI-assisted corporate tax enforcement can recover lost public revenues from sophisticated financial actors2.
• Administrative Efficiency and Service Accessibility: Natural language processing models can simplify complex regulatory filings, translate public agency services into dozens of minority languages, and assist citizens with disability accommodations in applying for municipal benefits1.
• Participatory Budgeting and Civic Analysis: Generative AI tools can summarize thousands of public policy submissions, analyze civic feedback during city planning consultations, and assist non-expert citizens in analyzing complex public budgets, lowering technical barriers to democratic participation1.
• Regulatory Consistency and Legislative Auditing: AI can evaluate proposed administrative regulations to ensure compliance with statutory mandates, detect conflicting statutory provisions, and identify regulatory loopholes exploited by corporate lobbies1.

Realizing these benefits requires an institutional framework where public administration uses AI exclusively to assist human discretion, expand public access, and audit powerful actors, rather than automate punishment against individuals1.

Civil-Liberties Considerations and Institutional Oversight Mechanisms

Risks to Due Process, Privacy, and Freedom of Expression

The expansion of AI into public and private governance generates systemic risks across core civil liberties:

• Due Process and Procedural Fairness: Automated decisions that lack explainability deprive individuals of the fundamental right to contest adverse agency actions, reversing legal burdens of proof and eliminating meaningful administrative appeals1.
• Equal Protection and Non-Discrimination: Profiling algorithms trained on historically biased data perpetuate racial, socioeconomic, and gender disparities in hiring, healthcare, credit allocation, and criminal sentencing2.
• Privacy and Personhood: Pervasive biometric harvesting and continuous workspace surveillance eradicate reasonable expectations of privacy, transforming public spaces and private workplaces into continuous monitoring zones5.
• Freedom of Expression and Information Integrity: Generative deepfakes and targeted persuasion manipulate political debate, while aggressive automated moderation and political censorship engines suppress constitutionally protected speech27.

Institutional Oversight Mechanisms and Procedural Safeguards

Mitigating these systemic risks requires actionable procedural mechanisms to re-establish accountability across automated legal and administrative frameworks1:

• The Warning-First Default: Automated administrative enforcement systems must issue non-punitive warnings and compliance guidance before executing monetary penalties, account suspensions, or benefit terminations2.
• The Compliance Parity Principle: State regulators and corporate platforms must face the same evidentiary burdens and procedural timelines when deploying automated enforcement tools as private citizens face when defending against automated actions2.
• The Right to Be Forgiven and Anti-Exploit Protocols: Automated systems tracking criminal records, credit scores, or workplace performance must incorporate statutory expiration dates and "right to be forgotten" protocols that erase historical risk indicators, preventing lifelong algorithmic penalization2.
• Reinvestment Rules and Severity Ladders: Administrative penalties generated through automated detection must be constrained by statutory severity caps, with collected fines dedicated strictly to public defense funds and civil liberties oversight rather than agency enrichment2.

Conclusion: Safeguards and Design Choices for Democratic Accountability

Preserving democratic accountability during a period of expanding AI capabilities requires combining rigorous institutional safeguards with privacy-preserving technical design choices1. Abstract ethical guidelines are insufficient; oversight requires enforceable legal structures and verifiable code architectures1.

Essential Institutional Safeguards

1. Mandatory Human Oversight and Non-Delegable Adjudication: Public agencies and commercial entities must be statutorily prohibited from delegating binding legal, administrative, or employment decisions to automated systems without qualified human review1. Human supervisors must possess the training, operational authority, and explicit mandate to override algorithmic outputs1.

2. Expansion of Algorithmic Disgorgement Remedies: Enforcement agencies must routinely deploy algorithmic disgorgement to mandate the complete destruction of AI models, weights, and pre-training datasets derived from unlawful data collection or deceptive practices16. Removing the commercial value of illicitly trained models provides an effective deterrent against corporate overreach20.

3. Statutory Pre-Market Assessments and Independent Auditing: High-risk AI systems deployed in public administration, law enforcement, or workplace management must undergo mandatory, independent pre-market conformity assessments1. Audits must evaluate not only abstract model accuracy, but also systemic bias, false-positive distributions across protected classes, data security, and operational reliability under real-world conditions5.

4. Enforceable Labor Safeguards: Codifying global frameworks similar to the EU Platform Work Directive ensures that workers retain statutory rights to human management, protection against emotional tracking, and collective access to the parameters governing algorithmic task allocation10.

Essential Technical Design Choices

1. Immutable Audit-Ready Logging Infrastructure: High-risk AI systems must be architected with native, immutable logging pipelines10. Every decision output, system override, data input update, and human supervisor intervention must be logged in tamper-evident data stores to enable post-hoc judicial discovery and regulatory auditing1.

2. Cryptographic Provenance and Multi-Layered Content Verification: Addressing synthetic media threats requires moving beyond easily circumvented, post-hoc watermarks28. Technical standards should adopt cryptographic provenance frameworks (such as C2PA) that embed cryptographically signed metadata at the hardware level during content creation3. Combining visible user labels with embedded metadata ensures verifiable tracking across political information environments3.

3. Privacy-Preserving Computation and Open-Weight Ecosystems: To counter infrastructure oligopolies, governments should fund open-weight foundation models and public compute infrastructure8. Integrating privacy-preserving techniques—such as zero-knowledge proofs (ZKPs), differential privacy, and secure multi-party computation (SMPC)—allows state agencies to perform public verification and fraud detection without aggregating raw, identifiable citizen data into centralized surveillance repositories2.

By binding technological development within enforceable legal boundaries and privacy-preserving system design, democratic societies can prevent the emergence of an unchecked algorithmic state, ensuring that automated systems remain subordinated to the rule of law, civil liberties, and public accountability1.

Works cited

1. Administrative Law and "Artificial Intelligence"—A Match Made in, https\://digitalcommons.law.lsu.edu/cgi/viewcontent.cgi?article=7107\&context=lalrev

2. Designing Empathy and Leniency into AI Moderated Legal, https\://acdroi.org/101987/1000483812/20260708192013

3. AI Regulation in China: Laws, Status and Penalties, https\://artificialintelligenceregulations.com/jurisdictions/china.html

4. ROYAL COMMISSION INTO THE ROBODEBT SCHEME, https\://espace.library.uq.edu.au/view/UQ:c408050/ANO999900010042.pdf

5. Preventing discrimination caused by the use of artificial intelligence, https\://www\.eerstekamer.nl/bijlage/20201105/preventing\_discrimination\_caused/document3/f=/vldiexl467t0.pdf

6. The EU Platform Work Directive | etui \- European Trade Union Institute, https\://www\.etui.org/publications/eu-platform-work-directive

7. COMPARATIVE ANALYSIS OF AI REGULATIONS IN THE US, THE, https\://brooklynworks.brooklaw.edu/cgi/viewcontent.cgi?article=2053\&context=bjil

8. AI Regulation in China: 2026 Update \- VLO Law Firm, https\://vlolawfirm.com/trackers/ai-regulation-china

9. Navigating China's regulatory approach to generative artificial, https\://pdfs.semanticscholar.org/9d62/e6c70c151c435b27b652bf8f6723f04c0d22.pdf

10. A deep dive into the EU Platform Work Directive | Arthur Cox LLP, https\://www\.arthurcox.com/knowledge/a-deep-dive-into-the-eu-platform-work-directive/

11. “Part Man, Part Machine, All Cop”: Automation in Policing \- PMC, https\://pmc.ncbi.nlm.nih.gov/articles/PMC8269447/

12. An AI Taxonomy for Criminal Justice, https\://counciloncj.org/wp-content/uploads/2026/05/AI-Taxonomy.pdf

13. Racial Recognition \- Cardozo Law Review, https\://www\.cardozolawreview.com/racial-recognition/

14. Emerging Technologies in Law Enforcement \- Right On Crime, https\://rightoncrime.com/wp-content/uploads/2025/08/2025-08-ROC-Emerging-Technologies-in-Law-Enforcement-Cochran.pdf

15. Law Enforcement in the Age of AI \- ACLU of Colorado, https\://www\.aclu-co.org/app/uploads/2025/11/USE\_ACLU\_2025SmartJusticeMini-report\_FINAL.pdf

16. AI Incident Law, https\://aiincidentlaw.org/

17. AI Decision Explanation Rights in Hiring, Welfare | Pebblous, https\://blog.pebblous.ai/blog/ai-decision-explanation-right/en/

18. Retrofitters, pragmatists and activists \- arXiv, https\://arxiv.org/pdf/2511.03211

19. Report on Artificial Intelligence and Civil Liability, https\://www\.bcli.org/wp-content/uploads/Report-AI-and-civil-liability-final.pdf

20. When the Model Must Be Destroyed \- by Alexandra Car, https\://alexandracar.substack.com/p/when-the-model-must-be-destroyed

21. EU Platform Work Directive: which countries have implemented?, https\://iuslaboris.com/insights/eu-platform-work-directive-which-countries-have-implemented/

22. EU Directive 2024/2831: Platform Work Guide & Tracker \- CXC Global, https\://www\.cxcglobal.com/blog/cxc-comply/eu-directive-platform-work-guide/

23. New EU Platform Work Directive Impacts Freelancers and Gig, https\://www\.fisherphillips.com/en/insights/insights/new-eu-platform-work-directive-impacts-freelancers-and-gig-economy

24. Platform Work Directive in France (2026): status and rules | teamed., https\://www\.teamed.global/eu-platform-work/france

25. Draft legislative proposal for the Platform Work Act open for, https\://knowledge.dlapiper.com/dlapiperknowledge/globalemploymentlatestdevelopments/2026/draft-legislative-proposal-for-the-platform-work-act-open-for-consultation

26. EU Platform Work Directive: Uncertainty Ahead of the, https\://www\.littler.com/news-analysis/asap/eu-platform-work-directive-uncertainty-ahead-implementation-deadline

27. Deepfake Recordings Allegedly Influence Slovakian Election, https\://incidentdatabase.ai/cite/573/

28. Election 2024: The Deepfake Threat to the 2024 Election, https\://www\.cfr.org/articles/election-2024-deepfake-threat-2024-election

29. How to Identify and Investigate AI Audio Deepfakes, a Major 2024, https\://gijn.org/resource/tipsheet-investigating-ai-audio-deepfakes/

30. Beyond the deepfake hype: AI, democracy, and “the Slovak case”, https\://misinforeview.hks.harvard.edu/article/beyond-the-deepfake-hype-ai-democracy-and-the-slovak-case/

31. AI-Generated Deepfakes Target Slovak President and Election, https\://oecd.ai/en/incidents/2023-08-06-9afc

32. China's New AI Regulations \- Latham & Watkins LLP, https\://www\.lw.com/admin/upload/SiteAttachments/Chinas-New-AI-Regulations.pdf

33. China: Regulation of Artificial Intelligence \- Loc, https\://tile.loc.gov/storage-services/service/ll/llglrd/2023555933/2023555933.pdf

34. AI Regulation Compared: EU, US, UK, China (2026) \- Legalithm, https\://www\.legalithm.com/en/blog/ai-regulation-comparison-eu-us-uk-china-global

35. An American's Guide to the EU AI Act, https\://scholar.law.colorado.edu/cgi/viewcontent.cgi?article=3104\&context=faculty-articles

36. Artificial intelligence policy worldwide: a comparative analysis, https\://royalsocietypublishing.org/rsos/article/13/2/242234/480264/Artificial-intelligence-policy-worldwide-a

37. EU AI Act Compliance Guide: Categories & Obligations \- Metricstream, https\://www\.metricstream.com/learn/what-is-the-eu-ai-act-guide-to-compliance-categories-obligations.html

38. International AI Law Compared to U.S. Frameworks: EU AI Act and, https\://ailegalauthority.com/international-ai-law-us-comparison/

39. China AI Regulation Overview \- Regulations.ai, https\://regulations.ai/regulations/RAI-CN-NA-SUMMARY-2026

40. Law / proposed law in China \- AI Laws of the World, https\://intelligence.dlapiper.com/artificial-intelligence/?t=01-law\&c=CN

41. Overview of Draft Measures on Generative AI \- China Law Translate —, https\://www\.chinalawtranslate.com/en/overview-of-draft-measures-on-generative-ai/

42. Slovakia's election deep fakes show how AI could be a danger to, https\://thefulcrum.us/slovakias-election-deep-fakes-show-how-ai-could-be-a-danger-to-us-elections

43. Slovakia | Reuters Institute for the Study of Journalism, https\://reutersinstitute.politics.ox.ac.uk/digital-news-report/2024/slovakia

44. Preventing discrimination caused by the use of artificial intelligence, https\://pace.coe.int/files/28715/html

45. draft-das-eu-ai-act-execution-enforcement-02 \- IETF Datatracker, https\://datatracker.ietf.org/doc/draft-das-eu-ai-act-execution-enforcement/