Full Report
Artificial intelligence chatbots are becoming a source for voters to ask questions about upcoming elections, but a new study determined nearly a third of AI-generated replies could be inaccurate or outdated. In a report released Thursday, researchers with the Institute for Strategic Dialogue (ISD) tested 15 different generic prompts across six AI chatbots and found 29…
Analysis Summary
# Research: Chatbots and the Ballot Box: Evaluating Accuracy, Sourcing and Language Gaps in AI Answers to Election Questions
## Metadata
- **Authors:** Researchers at the Institute for Strategic Dialogue (ISD)
- **Institution:** Institute for Strategic Dialogue (ISD)
- **Publication:** ISD Global Report (Summary via Threat Beat/The Hill)
- **Date:** September 9, 2026 (Report released Thursday prior)
## Abstract
This research evaluates the reliability of Large Language Model (LLM) chatbots as primary information sources for voters. By testing six major AI chatbots with standardized election-related queries, the study found that nearly one-third (29%) of generated responses contained inaccurate, outdated, or incomplete information regarding critical democratic processes such as voter registration, deadlines, and mail-in voting procedures.
## Research Objective
The study aims to quantify the "hallucination" and misinformation rate of AI chatbots when presented with factual, time-sensitive questions about upcoming elections, specifically focusing on whether these tools are safe for public use as civic information resources.
## Methodology
### Approach
The researchers employed a red-teaming/probing methodology, submitting a set of standardized "generic prompts" to multiple AI platforms to observe the consistency and accuracy of the output against official government election data.
### Dataset/Environment
- **Test Prompts:** 15 different generic prompts covering civic topics.
- **Topics:** Voter registration, mail-in voting requirements, important dates (Election Day), and local deadlines.
- **Scope:** Six major AI chatbots (commercial LLMs).
### Tools & Technologies
- Six unidentified generic AI chatbots (likely including industry leaders such as OpenAI, Google, and Anthropic, though the specific names for all six are indexed in the full ISD report).
## Key Findings
### Primary Results
1. **High Error Rate:** 29% of all AI-generated replies were found to be deficient (inaccurate, incomplete, or outdated).
2. **Temporal Errors:** A significant portion of errors involved misidentifying the specific date of Election Day.
3. **Procedural Misinformation:** Chatbots frequently provided outdated requirements for voter registration and mail-in ballots, failing to account for recent legislative changes.
### Supporting Evidence
- Empirical testing across 15 prompts revealed that nearly one in three interactions resulted in a failure to provide correct civic information.
### Novel Contributions
- The study highlights a "language and sourcing gap," suggesting that AI models struggle with hyper-local and time-sensitive civic data compared to general knowledge.
## Technical Details
While specific architectural failures aren't detailed in the summary, the findings point toward **Knowledge Cutoff Issues** (the training data precedes current election cycles) and **Retrieval-Augmented Generation (RAG) Failures** (the AI's inability to successfully pull the most recent facts from verified government domains in real-time).
## Practical Implications
### For Security Practitioners
- **Misinformation Risk:** AI tools can inadvertently become "automated misinformation engines," scaling the distribution of incorrect voting data without malicious intent from the user.
### For Defenders
- **Monitoring Requirements:** Social media platforms and election officials must monitor for AI-generated civic misinformation that might discourage voting or lead to disenfranchisement via incorrect deadline information.
- **Official Channel Promotion:** There is a critical need to steer users away from LLMs for factual civic data and toward `.gov` or verified institutional sources.
### For Researchers
- **Audit Frameworks:** The study emphasizes the need for standardized auditing frameworks for LLMs specifically regarding "High-Stakes" information (Health, Finance, Elections).
## Limitations
- The study focused on "generic" prompts; highly specific or localized prompts might yield different (potentially worse) error rates.
- AI models are updated frequently; the "snapshot" nature of the study may not reflect improvements made by developers immediately following the research.
## Comparison to Prior Work
This research reinforces previous findings on AI "hallucinations" but applies them specifically to the 2024/2026 election cycles, showing that despite technical advancements, the accuracy of civic data in AI remains stagnant and unreliable compared to traditional search engines.
## Real-world Applications
- **Implementation Considerations:** Developers should consider implementing "hard-coded" redirects for election queries that point users to official resources (e.g., Vote.org) rather than generating a text response.
- **User Education:** Public awareness campaigns are needed to warn voters that AI is not a fact-checked search engine.
## Future Work
- **Language Gaps:** Further study is required to see if accuracy rates drop even lower for non-English election queries.
- **Sourcing Transparency:** Investigating how AI models cite their sources—or fail to—when providing civic information.
## References
- Institute for Strategic Dialogue (ISD). (2026). *Chatbots and the Ballot Box*. [https://www.isdglobal.org/publication/chatbots-and-the-ballot-box-evaluating-accuracy-sourcing-and-language-gaps-in-ai-answers-to-election-questions/](https://www.isdglobal.org/publication/chatbots-and-the-ballot-box-evaluating-accuracy-sourcing-and-language-gaps-in-ai-answers-to-election-questions/)
- Nazzaro, M. (2026). "Nearly a third of AI chatbot replies to voting questions inaccurate." *Threat Beat/The Hill*.