Home / Library / Guides

GuideElection Readiness Series

Ready for AI Risks

Election risk scenarios and the AI threat dynamic

Risk and security Policy and governance

From

In partnership with

AI and Elections Clinic, Arizona State University

Share Copy link Updated June 2026

Executive Summary

This primer offers a framework for thinking about artificial intelligence and the security of American elections. It is meant to anchor future conversations about AI in election administration with a common vocabulary and a defensible analytical structure that election officials, technology providers, policymakers and the public can use together.

Organized around the election lifecycle

The framework follows the operational lifecycle of an election, from the moment a voter registers through the certification of final results. AI changes the risk dynamic at each stage, and it serves as a tool for adversaries and administrators alike.

Assessed through five lenses

AI risk to elections is assessed here through five lenses. Four are drawn from the NIST AI Risk Management Framework (AI RMF 1.0): Govern, Map, Measure and Manage. To these we add a fifth lens specific to election administration: Public Trust.

Public Trust is an Operational Variable

A correctly functioning election that is widely believed to have been manipulated produces democratic harm similar to a genuinely compromised one. Perception is therefore operational, and AI capabilities allow perception to be attacked at scale.

Grounded in the Alliance Standards

The framework draws on the U.S. Alliance for Election Excellence Standards as the operational grounding for what good election administration looks like in practice. The Standards, developed by practitioners for practitioners, describe what an excellent office does across Office Operations, Voting Operations, Poll Workers, Ballots and Audits, Communications, and Emergencies and Security. Each lifecycle phase opens with the Standards most relevant to that phase.

Key Findings

Election risks are lifecycle risksDependencies across the process mean consequences do not respect phase boundaries. A compromised voter registration database amplifies risk in every later phase.
AI changes the calculus in both directionsThe same capabilities that let adversaries scale disinformation, automate attacks and dox workers also let administrators detect anomalies, improve list accuracy and extend language access.
Disparate impact is a legal riskAI systems deployed in election administration are subject to the same laws as the rest of an office's work, not merely an ethical concern.
Digital tools can incite physical consequencesAI that enables doxing and personalized threats puts workers at risk, stresses the system and worsens the challenge of resourcing elections.
Perception of integrity is operationalAI can shape how the integrity of an election is perceived, creating real operational challenges even when the systems work.
The information environment is the sharpest surfaceThe post-election results window, when results are still being tabulated and certified, is among the periods when AI-generated false content can do the most damage.

The Framework: Five Lenses

The framework adopts the NIST AI Risk Management Framework as its structural backbone. NIST provides four core functions – Govern, Map, Measure and Manage – that together comprise a complete approach to AI risk. The rationale is practical: the NIST framework is technology-neutral, process-oriented, voluntary and rights-preserving, all of which suit the decentralized reality of American elections. It does not tie an office to any product or vendor. It focuses on how decisions get made.

To the four NIST functions we add a fifth lens specific to election administration: Public Trust. NIST treats trustworthiness as a property of the AI system. In most domains that framing is enough. Elections do not work that way. A jurisdiction can deploy AI that is valid, fair, secure and transparent and still suffer democratic harm if the public believes the AI is manipulating the outcome. The Public Trust lens asks the question the four functions do not naturally ask: regardless of whether the system performs correctly, will the public perceive it as performing correctly, and what does the office need to do to make accurate perception possible?

NIST · Cross-cutting

Govern

Build a culture of AI risk management across the office. Set policy, accountability, AI inventories, vendor standards and bipartisan oversight. This is where binding legal obligations such as the Voting Rights Act and HAVA enter the framework.

NIST Function

Map

Identify, categorize and contextualize AI risks across the lifecycle. Enumerate use cases, threat scenarios, affected stakeholders and the local narrative environments the jurisdiction operates in.

NIST Function

Measure

Assess the likelihood and impact of risks with methods that fit the evidence. Set key risk indicators, monitoring baselines and disparate impact protocols. Document what cannot be measured.

NIST Function

Manage

Put mitigations into practice: incident response, residual risk acceptance, vendor contract requirements, fallback options and resilience planning.

Election-Specific

Public Trust

Actively manage the relationship between system reality and public perception of that reality. Trustworthy operation is necessary but not sufficient. Accurate public perception is itself part of the operational outcome.

How the Lenses Relate

The four NIST functions describe activities. Public Trust describes an objective those activities serve, alongside operational integrity. A publicly disclosed bias test is both a Measure activity and a Public Trust activity. Work that serves only operational integrity, and not public perception, is incomplete in the election context.

Four Election Risk Domains

Risks are organized across four election-specific domains. Each addresses a distinct category of AI-enabled harm that is unique to or significantly elevated in elections.

Domain

Information & Narrative

AI content built to deceive voters, suppress turnout or undermine confidence in results. Includes deepfakes, synthetic candidate disinformation, suppression messaging and coordinated inauthentic behavior.

Domain

Cybersecurity

AI-assisted attacks on systems, infrastructure and personnel. Includes automated exploitation of voter registration databases, phishing of workers and supply chain compromise.

Domain

Physical Security

AI-enabled targeting of workers and physical infrastructure. Includes doxing, personalized threat generation, drone surveillance and coordinated polling place intimidation.

Domain

Legitimacy

AI that undermines the fairness and trust elections depend on through bias, opacity or partisan perception. Includes algorithmic disparate impact, opaque vendor deployment and undisclosed AI use.

Risk Scoring Methodology

Each risk is assessed with a likelihood-impact approach. A risk receives a likelihood score and an impact score, each on a one-to-five scale. The two are multiplied to produce an overall band: Critical, High, Medium or Low.

Likelihood and Impact (1–5)

Certain / SevereExpected during a cycle. Catastrophic effect on operations, integrity, access or public trust.

Likely / MajorHigh probability, seen in comparable environments. Significant disruption or reputational harm needing substantial intervention.

ModerateMay occur in some cycles. Noticeable disruption that can be managed with added effort.

Unlikely / MinorLow but credible probability. Limited effect, addressable within existing processes.

Rare / MinimalHighly unlikely under normal conditions. Negligible effect, handled by routine procedures.

Bands

Critical

20–25. Treat as a top priority for mitigation and oversight.

High

15–19. Requires a defined plan and assigned ownership.

Medium

10–14. Monitor and address as resources allow.

Low

Below 10. Manage through routine practice.

Decision Support, Not Measurement

Numeric scores support consistent comparison and prioritization, but they are decision-support tools rather than precise measurements. Treat the band as the primary judgment and use the underlying scores to support discussion, calibration and periodic reassessment in your own context.

The Election Lifecycle Risk Register

This register traces the operational lifecycle of an American election across four phases. For each phase it identifies the relevant Alliance Standards, offers a short context primer, and enumerates the primary risk scenarios with their NIST function assignments. Each scenario shows how AI amplifies the threat and how AI can support the defense.

Each scenario starts unscored. Set a likelihood and an impact for your jurisdiction directly on the row, and the scenario takes on its band. Your scores collect in a printable register near the end of this guide.

Phase 1 · Voter Registration

Primer · Expand for context How Voter Registration Works

Expand

Relevant Alliance Standards Voter Registration Rolls (Office Operations); Cybersecurity (Emergencies and Security); Voter Communications, Plain Language, Language Access and Community Relationships (Communications).

A state function under a federal framework

Registration is a state function run largely under the National Voter Registration Act (NVRA) and the Help America Vote Act (HAVA), with states retaining wide discretion. Voters register in person, online, by mail, through motor-voter and public assistance transactions, and through automatic voter registration in about 24 states. Same-day registration is available in about 22 states.

The statewide database

HAVA Section 303 requires each state to maintain a single, official, centralized, computerized statewide voter registration list, the VRDB. It is the master record of eligibility and the foundation for every later phase, coordinating with DMV records, Social Security death records, the SAVE citizenship system and state correctional records.

List maintenance

NVRA Section 8 governs list maintenance. It imposes both affirmative obligations and strict limits: no bulk removals within 90 days of a federal election, uniform and non-discriminatory removals, and the Voting Rights Act throughout. ERIC and bilateral data-sharing support accuracy. The now-defunct Interstate Crosscheck produced high false-positive rates that fell hardest on voters with common surnames. Most states offer public lookup tools, and HAVA guarantees a provisional ballot when eligibility cannot be confirmed at check-in.

REG-01 Erroneous Purges

Map · Measure · Manage Self-Score

Address-match errors, data quality failures or unverified third-party challenge lists remove eligible voters, concentrated among renters, minority voters and voters with name variations. AI amplifies through automated challenge-list generation that manufactures mass purge requests appearing legitimate and individually researched. AI defends through matching that reduces false positives and cross-checks proposed removals against multiple sources before any action.

REG-02 VRDB Cybersecurity Attack

Govern · Map · Measure · Manage Self-Score

AI automates discovery and exploitation of VRDB vulnerabilities for mass data theft or quiet record manipulation. AI amplifies through automated scanning that shortens exploitation windows against legacy systems and lowers the skill needed to chain misconfigurations into an intrusion. AI defends through anomaly detection on access logs and alert triage, backed by baseline controls such as multifactor authentication.

REG-03 Synthetic Registration Fraud at Scale

Map · Measure · Manage Self-Score

AI-generated synthetic identities enable coordinated fake registrations that overwhelm manual review and invite costly post-election cleanup. AI amplifies through low-cost synthetic identity data and volume timed to overwhelm verification just before deadlines. AI defends through pattern detection that flags anomalous IP clustering and implausible data for review, raising the cost of bulk submission.

REG-04 DMV / Agency Feed Transmission Failure

Measure · Manage Self-Score

Automatic and motor-voter flows depend on reliable handoffs between DMV systems and the VRDB. Transmission errors, format mismatches or downtime silently block registrations. AI amplifies through supply chain attacks on integration points and third-party services. AI defends through data quality monitoring and automated reconciliation of record counts and checksums against the source on each transfer.

REG-05 Deepfake Voter Registration Confusion

Govern · Map · Manage Self-Score

AI-generated audio or video of officials announces false deadlines, closures or requirement changes, targeted at specific communities during the registration window. AI amplifies through voice cloning and synthetic video that now require minimal skill and can be micro-targeted by language and platform. AI defends through C2PA provenance on official communications, paired with sustained public education on verifying information through official channels.

REG-06 Spear-Phishing of Registration Staff

Govern · Map · Measure · Manage Self-Score

AI-generated, position-specific phishing targets staff with VRDB credentials to raise success rates above generic phishing. AI amplifies through messages that are more convincing, contextually accurate and scalable than manual attempts. AI defends through email security tooling backed by phishing-resistant multifactor authentication and verification rules for any sensitive request.

REG-07 AI Voter-Assistant Information Errors

Map · Measure · Manage Self-Score

Voters increasingly ask AI chatbots and AI-enabled search how, when and where to register. These tools can return confident but incorrect answers on deadlines, eligibility rules or ID requirements. AI amplifies because reliance is rising and errors are stated authoritatively, hard to trace and shift as models change. AI defends when offices publish authoritative, machine-readable information from .gov domains and proactively flag errors for remediation.

REG-08 AI Information Environment Poisoning

Map · Measure · Manage Self-Score

Adversaries seed and boost inauthentic content so AI assistants and AI-enabled search return planted answers about deadlines, eligibility or ID rules, targeting the retrieval layer where local information is thin. AI amplifies because a few coordinated sites can surface as a model's source on narrow local queries and data voids invite low-quality content. AI defends by filling the void with authoritative .gov content, monitoring assistant outputs and coordinating to deprioritize known inauthentic domains.

Phase 2 · Absentee & Mail Ballot

Primer · Expand for context How Mail Voting Works

Expand

Relevant Alliance Standards Mail Ballots (Voting Operations); Ballot Creation and Secure Ballot Management (Ballots and Audits); Voter Communications and Plain Language (Communications); Physical Security (Emergencies and Security).

A large and growing share of ballots

Mail voting, both no-excuse absentee and all-mail systems, accounts for a substantial and growing share of American ballots. In 2024, states sent about 67 million ballots, roughly 72 percent were returned, and about 1.2 million were rejected, primarily for missing or mismatched signatures. States as ideologically diverse as Utah, California and Washington rely heavily on all-mail systems.

From request to receipt

A voter requests a ballot, or receives one automatically, and the office verifies eligibility against the VRDB before generating a packet with the correct ballot style, envelopes and instructions. High-speed printers and inserters produce packets at scale, a point where wrong-style insertion is a serious risk. USPS handles delivery with Intelligent Mail Barcode tracking, and delays in rural areas and at postal bottlenecks can push voters against return deadlines.

Verification and cure

Voters return ballots by mail, by hand, or through secure drop boxes where authorized. Before counting, the signature on the return envelope is verified against the VRDB. Signature mismatch is the leading reason for rejection, and cure processes vary widely, including by county within the same state.

MAIL-01 Signature Verification Disparate Impact

Govern · Map · Measure · Manage Self-Score

AI signature matching produces higher rejection rates for younger voters with inconsistent signatures, elderly voters with age-related variation and minority voters, creating a systemic Voting Rights Act concern at scale. AI amplifies by systematizing disparities that once depended on which individual adjudicator happened to review a ballot. AI defends when well-validated tools reduce human bias, flag close calls for multi-person review and publish aggregate cure and rejection rates so disparities surface early.

MAIL-02 Drop Box Physical Interference

Map · Manage Self-Score

AI-driven surveillance and predictive modeling identify high-use drop box locations for coordinated disruption. AI amplifies through public-data analysis that enables adversarial targeting, often paired with narratives that exaggerate a single incident to suppress turnout. AI defends by combining weather, event and traffic data to position staff and harden locations before problems escalate.

MAIL-03 Ballot Tracking Data Exploitation

Map · Manage Self-Score

Public USPS tracking creates visibility into return patterns that adversaries can aggregate to find timing vulnerabilities or interception choke points. AI amplifies because mass aggregation and analysis of tracking data is trivial with AI tools. AI defends through anomaly detection on return data and by minimizing and aggregating published tracking so granular patterns are not exposed.

MAIL-04 Wrong Ballot Style Insertion

Manage Self-Score

Automated inserter misconfiguration or software error delivers a voter the wrong ballot style. AI amplifies through supply chain attacks on ballot management configuration files. AI defends through machine-vision quality control that checks insertions in real time and cryptographic signing of ballot-definition files so any tampered configuration fails validation before production.

MAIL-05 AI-Generated False Deadline Disinformation

Govern · Map · Manage Self-Score · indicative medium to high

Targeted disinformation about return deadlines, signature requirements or drop box locations suppresses mail return among targeted communities. AI amplifies because personalized, multilingual suppression messaging is now producible at industrial scale with large language models. AI defends through content provenance standards and real-time platform monitoring.

MAIL-06 USPS Processing Delay Exploitation

Map · Measure · Manage Self-Score · indicative low to medium

Coordinated disruption or predictive timing of ballot interception exploits the known vulnerability of ballots in transit. AI amplifies through predictive modeling of USPS routing vulnerabilities. AI defends through data analysis that supports proactive voter outreach when delivery failures are predicted before deadlines.

Phase 3 · In-Person Voting

Primer · Expand for context The In-Person Voting Ecosystem

Expand

Relevant Alliance Standards In-Person Voting (Voting Operations); Poll Worker Recruitment, Training, Management and Retention (Poll Workers); Voting Resources Allocation and Ballot Creation (Ballots and Audits); Physical Security and Emergency Planning (Emergencies and Security).

Still the dominant mode

In-person voting, early and on Election Day, remains the dominant mode in most jurisdictions even as mail voting grows. Only about 34 percent of voters cast ballots on Election Day itself in 2024, reflecting the long-term move toward early and alternative voting. This phase holds the tightest intersection of technology, people, law and logistics in the lifecycle.

Check-in and equipment

Electronic pollbooks, tablet or laptop check-in systems tied to the VRDB, have largely replaced paper poll books. They add dependencies on connectivity, configuration accuracy and battery life, and they are not subject to the independent certification required of voting machines. Synchronization failures across a jurisdiction can produce inconsistent records and provisional ballot surges.

Voting systems

Most jurisdictions use optical scan tabulators that read hand-marked paper ballots, with the paper serving as the authoritative record for recounts. Ballot marking devices provide accessible voting and a voter-verifiable paper record. Direct recording electronic systems without a paper trail are inconsistent with modern security practice. The Election Management System, often air-gapped, programs ballots and aggregates results. Logic and accuracy testing is an essential pre-election control.

IPV-01 EPollbook Sync Failure

Manage Self-Score

Connectivity or synchronization failures produce inconsistent check-in records across precincts. AI amplifies through targeted network interference or crafted inputs that deliberately trigger sync failures at scale. AI defends through real-time anomaly detection on EPollbook traffic, with tested paper poll books as a non-technical backup.

IPV-02 Wrong Ballot Style Assignment

Manage Self-Score

An EPollbook configuration error or operator mistake assigns a voter the wrong ballot style. AI amplifies through configuration tampering in ballot style assignment tables. AI defends through automated pre-election ballot style verification paired with robust logic and accuracy testing.

IPV-03 Voting Equipment Failure on Election Day

Manage Self-Score

Scanner jams, memory card failures, battery depletion or software crashes, most common in jurisdictions with aging equipment. AI amplifies through adversarial discovery of equipment failure conditions. AI defends through predictive maintenance models, backed by sufficient spare equipment and emergency paper ballots, since prediction reduces but never eliminates failure.

IPV-04 AI-Optimized Polling Place Disruption

Map · Measure · Manage Self-Score

Predictive modeling on public precinct demographics, registration data and historical turnout identifies which polling places, if disrupted, would maximize disenfranchisement in specific communities. AI amplifies because public data makes targeting analysis accessible, and it compounds when physical disruption is timed with AI-generated false information. AI defends when administrators use the same tools to harden and pre-position supplies and signage at the highest-risk locations.

IPV-05 Real-Time Voter Misdirection on Election Day

Map · Manage Self-Score

AI-generated posts and text messages spread false information about polling locations, hours, equipment failures or procedure changes. AI amplifies because targeted content can be delivered at scale with micro-targeting precision during the highest-impact window. AI defends through real-time social monitoring, though the response window is extremely short.

IPV-06 Poll Worker Training Gaps and Confusion Exploitation

Manage Self-Score

Turnover among election workers leaves more jurisdictions with less experienced staff. AI amplifies through social engineering enhanced with official-sounding directives and voice impersonation of supervisors. AI defends through AI-assisted training tools and real-time decision-support apps.

Phase 4 · Results, Audits & Certification

Primer · Expand for context The Results and Certification Process

Expand

Relevant Alliance Standards Results Reporting (Voting Operations); Tabulation Audits and Secure Ballot Management (Ballots and Audits); Voter Communications, Media Relationships and Community Relationships (Communications); Cybersecurity and Physical Security (Emergencies and Security).

Where complexity meets consequence

Results and certification is where operational complexity meets political and legal consequence. It is the phase most exposed to AI-enabled information warfare, because the window between unofficial election night results and official certification, which can run days to weeks, is precisely when public trust is most fragile and adversaries have the most to gain from narrative disruption.

Tabulation and reporting

Optical scan tabulators generate results that aggregate through the Election Management System, with memory cards or physically controlled connections moving results to central count. Election Night Reporting pushes unofficial results to public portals. States vary widely in reporting speed by law, not by failure, and the normal appearance of shifts as mail ballots are counted is consistently exploited as false evidence of fraud.

Canvass and chain of custody

The canvass reconciles all ballot types into a complete official result, adjudicating provisional ballots and resolving discrepancies between precinct totals and check-in records. Chain of custody documentation, tracking ballots, media and equipment from logic and accuracy testing through post-election storage, is a foundational control. Any documented break is a reportable event that must be investigated, even absent evidence of tampering.

RES-01 AI-Fabricated Results Dashboards

Govern · Map · Manage Self-Score

AI-generated fake tallies, county result maps and websites that mimic official reporting publish in real time on election night to undermine trust before the canvass is complete. AI amplifies through generative tools that produce convincing look-alikes and graphics instantly. AI defends through content provenance standards on official reporting outputs.

RES-02 Deepfake Officials Post-Election

Govern · Map · Manage Self-Score

Synthetic audio or video of officials, candidates or credentialed observers makes false post-election statements, including false concessions, fraud allegations or results announcements. AI amplifies through voice cloning and synthetic video. AI defends through media authentication and detection tools and a pre-established response protocol that limits the spread of inaccurate information.

RES-03 Exploiting Legitimate Counting Delays

Govern · Map · Manage Self-Score

Disinformation exploits the normal, legal and expected pattern of mail ballot late-counting producing shifts in results, manufacturing false fraud narratives during the uncertainty window. AI amplifies through coordinated inauthentic networks that now use generative AI to run numerous human-like personas that evade simple detection. AI defends through pre-election communication about expected timelines, authenticated, widely distributed and carried by trusted partners.

RES-04 AI-Generated Synthetic Evidence

Map · Measure · Manage Self-Score

AI produces synthetic documentary evidence such as fabricated chain of custody records, manipulated surveillance footage and fake poll worker statements. This material can land anywhere from public debate to misuse in post-election litigation or legislative proceedings. AI amplifies because convincing forgery tools lower the barrier for bad-faith contests. AI defends through an unbroken, well-documented chain of custody.

RES-05 EMS / ENR Cybersecurity Attack

Govern · Map · Measure · Manage Self-Score

A targeted cyberattack hits the Election Management System or the reporting system during the results window. AI amplifies through AI-assisted vulnerability discovery and automated reconnaissance. AI defends through anomaly detection on access logs, strict access governance, signed software updates and monitoring of the reporting layer, which is the higher-probability vector since the management system is typically air-gapped.

RES-06 Certification Pressure and Subversion

Govern · Map · Manage Self-Score

AI-enhanced doxing, targeted threat campaigns and disinformation about individual canvassing board members aim to pressure officials into refusing certification or certifying false results. AI amplifies because personalized pressure campaigns can be launched at scale with minimal resources. AI defends through the same protections that safeguard election workers generally.

AI Changes the Calculus in Both Directions

The capabilities that make AI a threat to elections are, in many cases, the same capabilities that make it useful to administrators. The question for an office is not whether to engage with AI, but how to govern it so the defensive value is captured and the adversarial value is contained. Three areas show the pattern.

The information environment

AI as Threat

Suppression and confusion at scale. Voice cloning, synthetic video and large language models produce personalized, multilingual disinformation about deadlines, locations and results, delivered with apparent official authority.

AI as Defense

Provenance and rapid response. Content provenance standards authenticate official communications, and monitoring tools surface emerging false narratives early enough for a pre-authorized response.

Systems and infrastructure

AI as Threat

Faster, cheaper intrusion. AI accelerates vulnerability discovery against legacy systems and powers contextually accurate spear-phishing of staff with database credentials.

AI as Defense

Anomaly detection and triage. AI monitors access logs for unusual query volumes and off-hours exports, helping limited security resources focus on the highest-risk exposures.

Operations and access

AI as Threat

Systematized disparate impact. Poorly validated signature matching or list maintenance can scale bias that once depended on individual reviewers, turning an ethical concern into a legal one.

AI as Defense

Accuracy and access gains. Validated matching reduces false positives for voters with name variations, and translation tools can extend language access beyond statutory minimums when independently checked.

The Upside Deserves a Mention

Many of the tasks election officials shoulder are repetitive, exacting and unforgiving of small slips. Those are exactly the conditions where a well-governed AI system can excel.

The Critical Window

Why the post-election period is the sharpest risk surface

The window between polls closing and certification is the highest-risk information environment in the lifecycle. It combines maximum public attention, maximum uncertainty about final results because of the legally necessary multi-day count, and maximum adversarial motivation to disrupt the legitimacy of the outcome. Generative content, synthetic media and automated amplification are tailor-made for this environment. It is here, more than anywhere else, that the Public Trust lens earns its place: the technical correctness of the count is necessary but not sufficient.

Why the Count Shifts

Stanford research on the 2024 election documents the predictable vulnerability. Persistent state-level differences in counting speed create multi-day gaps. About 1.7 million provisional ballots were cast, with 74 percent ultimately counted, so meaningful tallies continue to shift after election night. Rising official turnover leaves many jurisdictions with less experienced administrators.

Publish the timeline early

Before Election Day, publish expected counting timelines, authenticated and widely distributed, so the multi-day count is understood as normal before it can be reframed as suspicious.
Pre-stage authenticated content

Prepare plain-language explainers on the canvass, the audit process and the certification timeline, ready for rapid deployment when counter-narratives emerge.
Stand up round-the-clock monitoring

Establish, in advance, a 24/7 monitoring and rapid-response capability with pre-authorized messaging and clear decision authority. Relationships with media and community partners must exist before the window, not be built during it.
Sign official outputs

Apply content provenance to official reporting outputs so fabricated dashboards and look-alike sites can be distinguished from authentic results.
Confirm the count and report

Use audits and an unbroken chain of custody to confirm tabulation accuracy, then commit publicly, before the election, to a post-certification transparency report.

A Standing Concern: Election Worker Safety

A cross-cutting risk in every phase

The safety of the people who run elections is fundamental to election administration. AI tools that enable doxing and personalized threat generation put workers at risk, stress the system and worsen the challenge of resourcing elections. This concern does not belong to any single phase. It runs from the registration window through certification, when individual canvassing board members can become targets of pressure campaigns.

The Threat Is Already Active

The 2024 election saw more than 227 bomb threats against polling places and election offices, many targeting drop box locations. This is not a hypothetical vector.

Authenticate official messagingApply content provenance such as C2PA to official communications so impersonation and deepfakes of officials are easier to flag and refute.
Pre-establish response protocolsDecide in advance who responds, how quickly and with what authority when threats or synthetic statements appear.
Coordinate with partners earlyBuild standing relationships with EI-ISAC, CISA and the FBI Election Crimes Branch before an incident, not during one.
Monitor for doxing and threatsMaintain monitoring for personalized threat campaigns, with the capacity to push accurate information quickly.
Use trusted local voicesLean on Community Relationships so credible local messengers can carry accurate information where the office cannot reach directly.
Protect the certification functionExtend worker protections to canvassing board members targeted by pressure campaigns designed to subvert certification.

Appendix

Reference · Expand to view Glossary of Key Terms

Expand

Alliance Standards

The U.S. Alliance for Election Excellence Standards for Election Excellence, a practitioner-developed framework describing what excellent election administration looks like across six domains: Office Operations, Voting Operations, Poll Workers, Ballots and Audits, Communications, and Emergencies and Security.

Air-gapping

Keeping a system physically disconnected from the internet and other networks, so it cannot be reached remotely.

BMD (Ballot Marking Device)

An accessible voting system that produces a voter-verifiable paper ballot record. Required for ADA-accessible voting.

C2PA (Coalition for Content Provenance and Authenticity)

An open technical standard for attaching tamper-evident provenance data to images, video and audio, recording where a file originated and how it was edited. It gives officials a way to verify authentic media and flag content that has been altered or stripped of its history.

Canvass

The official post-election process of reconciling all ballot types into a complete official result, prior to formal certification.

Deepfake

Synthetic or manipulated audio, video or images generated with AI to convincingly impersonate a real person or event.

EPollbook

An electronic voter check-in system, typically tablet or laptop-based, that replaces paper poll books and connects to the statewide voter registration database.

ERIC (Electronic Registration Information Center)

A nonprofit data-sharing consortium used by member states to improve voter list accuracy.

HAVA (Help America Vote Act of 2002)

Established the EAC, required statewide voter registration databases, set minimum voting system standards and created the provisional ballot right.

IMb (Intelligent Mail Barcode)

A USPS barcode applied to mail ballots that enables tracking through the postal system.

Information Environment Poisoning (LLM Grooming)

Deliberate seeding and boosting of inauthentic content so AI assistants and AI-enabled search return it as a legitimate answer. Retrieval-layer poisoning targets the live sources a model cites; training-data poisoning targets the model itself. It is most effective where authoritative information is scarce.

L&A Testing (Logic and Accuracy Testing)

Pre-election testing of voting equipment to verify accurate recording and counting.

NCOA (National Change of Address)

A USPS program providing address-change data used by jurisdictions for voter list maintenance under the NVRA safe harbor provision.

NVRA (National Voter Registration Act of 1993)

The 'motor voter' law, establishing federal registration requirements and list maintenance rules.

Public Trust Lens

The fifth lens added to NIST's four functions in this framework. It captures the election-specific reality that public perception of system integrity is itself an operational variable, not a downstream consequence of operational reality.

UOCAVA

The Uniformed and Overseas Citizens Absentee Voting Act, providing absentee voting rights for military and overseas citizens, with a mandated 45-day ballot transmission window.

VRDB (Voter Registration Database)

The statewide computerized voter registration list required by HAVA.

VRA (Voting Rights Act of 1965)

Prohibits discriminatory voting practices.

VVSG (Voluntary Voting System Guidelines)

EAC guidelines establishing testing and certification requirements for voting systems. VVSG 2.0 is the current version.

Reference · Expand to view Reference Standards and Resources

Expand

NIST frameworks and publications

NIST AI Risk Management Framework (AI RMF 1.0), January 2023. NIST Special Publication 800-53 Rev. 5, Security and Privacy Controls. NIST SP 1270, Towards a Standard for Identifying and Managing Bias in AI. NIST Cybersecurity Framework (CSF 2.0).

Federal election security guidance

CISA Election Security Resources. CISA, Risk in Focus: Generative A.I. and the 2024 Election Cycle, January 2024. CISA, FBI and ODNI, Securing Election Infrastructure Against the Tactics of Foreign Malign Influence Operations, April 2024. EAC Voluntary Voting System Guidelines (VVSG 2.0) and Election Security Preparedness guidance.

Standards bodies and research

U.S. Alliance for Election Excellence, Standards for Election Excellence. Brennan Center for Justice, Artificial Intelligence and Election Security, 2023. Coalition for Content Provenance and Authenticity (C2PA). Cloudflare Project Galileo.

Election law references

National Voter Registration Act (NVRA), 52 U.S.C. ch. 205. Help America Vote Act (HAVA), 52 U.S.C. § 20901 et seq. Voting Rights Act (VRA), 52 U.S.C. § 10101 et seq., Sections 2, 5 and 203. Americans with Disabilities Act (ADA), 42 U.S.C. § 12101 et seq. UOCAVA / MOVE Act. Federal election crime statutes, 18 U.S.C. §§ 594–610.

Key programs and organizations

Elections Infrastructure Information Sharing and Analysis Center (EI-ISAC). Electronic Registration Information Center (ERIC). CISA Protective Security Advisor Program. FBI Election Crimes Branch. DOJ Voting Section. NCSL Artificial Intelligence Legislation Database.

Supporting resources

A short set of primary sources behind this framework. Each links to the issuing body.

AI Risk Management Framework (AI RMF 1.0)

NIST

The technology-neutral backbone of this primer: the Govern, Map, Measure and Manage functions.
View Resource →
Risk in Focus: Generative A.I. and the 2024 Election Cycle

CISA

Federal guidance on generative AI threats across the election cycle.
View Resource →
Standards for Election Excellence

U.S. Alliance for Election Excellence

The practitioner-developed operational grounding each phase of this framework opens with.
View Resource →
Artificial Intelligence and Election Security

Brennan Center for Justice

Research on AI risks to election infrastructure and the information environment.
View Resource →
Elections Infrastructure ISAC (EI-ISAC)

Center for Internet Security

Threat sharing and incident support tailored to election offices.
View Resource →
The Election Security Exchange

Practitioner Network

Practitioner-led threat sharing and hands-on support for election offices.
View Resource →
Artificial Intelligence Legislation Database

National Conference of State Legislatures

Tracks state AI legislation relevant to government and election use.
View Resource →

Work with us

Need training or support for your jurisdiction or organization?

We offer custom training, workshops, and consulting tailored to your needs.

Explore our services

About This Series

The guides in this series help election officials connect communications strategy with operational practice before, during, and after elections. Each guide addresses a specific moment in the election cycle and offers language, frameworks, and decision tools that are nonpartisan, procedurally grounded, and transferable across jurisdictions.

Readers are encouraged to adapt language and examples to reflect the specific laws, procedures, and terminology of their state and jurisdiction.

The AI and Elections Practice supports the AI + Elections Clinic at the Arizona State University Mechanics of Democracy Laboratory.

Part of the Election Readiness Series. Adapt language and examples to reflect the laws and procedures of your state and jurisdiction. Last Updated June 2026.
A Ready for Tuesday guide · Privacy Policy · Terms and Conditions of Use