This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
Amazon's Alexa voice assistant generated politically biased responses regarding the 2024 US presidential candidates, appearing to endorse Kamala Harris while refusing to provide reasons to vote for Donald Trump. The issue was linked to a large language model component called 'Info LLM' intended to improve accuracy. Amazon characterized the behavior as an error and implemented manual blocks to prevent further biased responses.
Amazon's Alexa was found to provide politically biased responses when asked about the 2024 presidential candidates. It refused to give reasons to vote for Donald Trump, citing neutrality, while offering detailed endorsements for Kamala Harris. Amazon labeled the discrepancy an "error" and reportedly corrected it.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
Alexa also provided outdated or incorrect political information, such as listing RFK Jr. as an active candidate after he dropped out.
The voice assistant failed to perform reliably and neutrally under varying user queries, generating biased political endorsements due to an underlying software error.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The biased responses were caused by an unexpected error in Alexa's Info LLM software component after it was deployed to users.
Amazon's Alexa voice assistant generated politically biased and inaccurate responses regarding the 2024 US presidential election due to an unintentional software error. While the incident fueled public controversy and accusations of election interference, it represents a minor national security threat that was quickly mitigated through corporate intervention.
The report characterizes the AI system as Amazon's digital assistant or voice assistant Alexa, which responds to user queries to provide information and requires human interaction.
Risk Level 3: Limited Risk. The system is a chatbot/voice assistant (Alexa) interacting with users, which falls under Limited Risk requiring transparency obligations.
Alexa is a voice assistant designed to answer user questions and provide information.
The reports provide clear details about the nature of the AI's error, the specific software component responsible (Info LLM), Amazon's response, and the public reaction. There is little ambiguity about what occurred.
None
Technology
Info LLM Integration
New Info LLM component introduced unpredictable behavior in political queries.
Unpredictable LLM Outputs
Underlying LLM technology can produce erratic and biased responses.
Inadequate Automated Filters
Automated guardrails failed to block politically biased responses.
Data Inputs
Outdated Training Data
Model relied on outdated political information, like active candidates.
Lack of Neutral Data Tuning
Data inputs or tuning led to asymmetric responses for different candidates.
Human Factors
Complex User Queries
Voters ask highly varied and high-stakes questions during election cycles.
Process and Methods
Deficient Pre-Release Testing
Testing protocols failed to detect the biased responses before deployment.
Reactive Error Detection
System relied on social media reports to identify and fix critical errors.
Management
Insufficient Proactive Auditing
Management failed to thoroughly audit political queries before live deployment.
Inadequate Risk Management
Underestimated the impact of LLM hallucinations during election season.
Regulatory Environment
Lack of AI Election Mandates
No clear regulatory guidelines exist for AI assistants during elections.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
