AI Models Reportedly Found to Provide Misinformation on Election Processes in Spanish

An analysis reportedly found that multiple AI models provided inaccurate responses to election-related questions, with 52% of Spanish-language answers and 43% of English-language answers containing misinformation or omissions. Errors included misidentifying voting processes and providing information about foreign elections.

Risk classification

• Primary risk domain: 1 Discrimination & Toxicity
• Primary risk subdomain: 1.3 Unequal performance across groups

The AI models exhibited a significant disparity in accuracy rates between English (43% inaccurate) and Spanish (52% inaccurate) queries, resulting in unequal performance for Spanish-speaking users seeking election information.

Additional risk subdomains

• 3.1 False or misleading information: The models frequently generated false, misleading, or incomplete information regarding U.S. election processes, candidates, and voting security.

Causal factors

• Entity: AI
• Intent: Unintentional
• Timing: Post-deployment

The inaccuracies and language disparities were produced by deployed AI models responding to user queries, representing an unintended failure mode of the systems.

EU AI Act risk tier

• Risk tier: 3 Limited Risk

Limited Risk: The systems evaluated are general-purpose chatbots, which are classified as limited risk under the EU AI Act, requiring transparency so users know they are interacting with AI.

AI system and alleged parties

• AI system: GPT-4, Gemini 1.5 Pro, Llama 3, Mixtral 8x7B v0.1 (Google, Meta, Mistral AI, OpenAI)
• AI purpose: Chatbot; Question Answering
• Behaviour type: Assistant
• Alleged developer: OpenAI, Mistral, Meta, Google, Anthropic
• Alleged deployer: OpenAI, Mistral, Meta, Google, Anthropic
• Alleged harmed parties: U.S. electorate, Spanish-speaking American voters, Spanish speakers, Electoral integrity, Democracy

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Minor. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Negligible, indirect Negligible
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Minor, indirect Minor
• Civil rights: direct Negligible, indirect Minor
• Democracy: direct Minor, indirect Minor
• Privacy: direct Negligible, indirect Negligible
• Psychological: direct Negligible, indirect Negligible
• Epistemic: direct Minor, indirect Substantial
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Differential treatment

Reported: The report explicitly describes differential treatment in the quality of information provided to Spanish-speaking users compared to English-speaking users.

Directly caused: Spanish-language queries resulted in a 52 percent inaccuracy rate compared to 43 percent for English queries across five leading AI models.

Indirectly caused: Spanish-speaking voters receive lower-quality, potentially misleading information about voting processes, which could affect their ability to participate equally in elections.

Inferred additional harm: Spanish-speaking communities may face systemic disadvantages in accessing accurate voting information, potentially leading to lower voter turnout or disenfranchisement.

Civil rights

Reported: The report explicitly describes issues affecting the ability of Spanish-speaking citizens to access accurate voting information, which impacts their civil right to vote.

Directly caused: AI models provided incorrect or incomplete information on voting requirements, such as 'federal only voters' in Arizona, and directed Spanish queries to irrelevant foreign election processes.

Indirectly caused: Inaccurate information about voter registration and mail-in voting security could discourage or prevent eligible voters from exercising their right to vote.

Inferred additional harm: Potential disenfranchisement of Spanish-speaking voters who rely on these AI systems for election information.

Democracy

Reported: The report explicitly describes the risk of AI models providing inaccurate information about elections, candidates, and voting processes, which threatens democratic norms.

Directly caused: AI models generated misleading information about U.S. elections, including false claims about mail-in voting security and incorrect details about candidates like Donald Trump and Project 2025.

Indirectly caused: The dissemination of inaccurate election information undermines trust in democratic institutions, election integrity, and the voting process.

Inferred additional harm: Widespread reliance on these models during elections could lead to broader erosion of public trust in democratic processes and election outcomes.

Epistemic

Reported: The report explicitly describes extensive generation of false, misleading, and incomplete information by AI models regarding election topics.

Directly caused: 48 percent of AI model responses to election questions in English and Spanish contained incorrect or incomplete information, including hallucinated facts and broken links.

Indirectly caused: Users seeking reliable voting information are misled by false claims, such as incorrect definitions of 'federal only voters' or false claims about ballot printing companies.

Inferred additional harm: The proliferation of inaccurate AI-generated election information contributes to the overall pollution of the information ecosystem, making it harder for voters to find verified facts.

People affected

• Occurrences reported: 1

Potential causes

Management

• Premature Model Release: Releasing models for public use before verifying election-related safety.
• Neglecting Developer Guidelines: Meta released Llama 3 without including election safety in developer guides.
• Broken Safety Pledges: Management failed to enforce commitments to redirect election queries.

Technology

• Language Bias in Model Training: Models like Claude 3 were trained primarily on English language data.
• Failure to Detect Spanish Queries: Systems failed to recognize election queries when asked in Spanish.
• Algorithmic Hallucination of Links: Models generated broken links or English links in Spanish responses.

Data Inputs

• Lack of Spanish Election Data: Lack of Spanish data led to referencing Latin American elections.
• Vulnerability to Online Misinfo: Models absorbed unverified claims where truthful online data was scarce.
• Incomplete Context in Training Data: Omission of crucial localized voting facts like AZ live video feed rules.

Human Factors

• Developer Neglect of Non-English: Developers prioritized English performance, neglecting safety in Spanish.
• User Trust in Chatbot Accuracy: Users rely on chatbots for critical voting info despite high error rates.

Process and Methods

• Inadequate Multi-language Testing: Companies failed to test safety guardrails across multiple languages.
• Failed Redirection Mechanisms: Pledges to redirect users to TurboVote or CanIVote were not executed.
• Poor Quality Control on Outputs: Lack of validation allowed models to output contradictory safety claims.

Regulatory Environment

• Lack of Standardized AI Mandates: No regulatory mandates forced companies to ensure bilingual election safety.

Information quality

• Classification confidence: High
• Reason for confidence: The report is based on a structured, empirical study conducted by the AI Democracy Projects and Factchequeado, providing clear statistics (e.g., 52% vs 43% inaccuracy rates) and specific examples of model failures across multiple languages. The roles of the AI systems and the nature of the failures are explicitly documented.
• Ambiguities identified: The report does not specify the exact number of real-world users who have been misled by these specific model outputs, focusing instead on controlled testing.
• Alternative interpretations: The incident could be viewed primarily as a general capability limitation (7.3) rather than unequal performance (1.3), but the significant statistical disparity between English and Spanish performance strongly supports the discrimination/unequal performance classification.

An empirical study revealed that five leading AI models frequently provide inaccurate election information, with a higher error rate for Spanish-language queries than English. This systemic failure poses a substantial threat to sovereignty and societal stability by potentially disenfranchising Spanish-speaking voters and undermining public trust in democratic processes.

• Overall national security impact: Substantial
• Scope: Single nation
• Primary target: United States
• Alleged perpetrator: Unknown

Threat characteristics

• Imminence: Long-term. This represents an ongoing, systemic capability and alignment challenge for AI developers rather than an active, fast-moving crisis.
• Autonomy: Full autonomy. The AI models generate and deliver responses directly to users in real-time without any human-in-the-loop filtering or verification by the developers.
• Novelty: Evolved capability. AI hallucinations and translation issues are established problems, but their manifestation at scale across five leading models during a major election represents an evolved threat.

Impact by dimension

• Physical security: Negligible. The incident involves informational errors by text-based AI models and does not present any threat to physical systems, critical infrastructure, or human safety.
• Information security: Minor. While not a coordinated campaign, the high error rates and hallucinated facts pollute the information ecosystem, presenting a vulnerability that could be exploited for information warfare.
• Sovereignty: Substantial. Providing inaccurate information on voting processes and election integrity directly threatens core democratic processes and public trust in government institutions, particularly concerning U.S. election administration.
• Economic security: Negligible. There are no direct implications for financial systems, strategic technology theft, or critical supply chains arising from these model inaccuracies.
• Societal stability: Substantial. The significant performance disparity (52 percent inaccuracy in Spanish vs 43 percent in English) risks systemic disenfranchisement of the 42 million Spanish speakers in the U.S. seeking voting information.