Summary
- States United Democracy Center tested ChatGPT and Google AI across six swing states, finding factual error rates fell to zero in 2026 follow-up tests, down from roughly 7–8% in 2025.
- Despite cleaner facts, ChatGPT returned incomplete candidate lists 88.9% of the time, and linked to official state election websites less than 40% of the time.
- Bad actors are now structuring information operations to rank favourably in AI model responses — a tactic researchers are calling ‘generative engine optimisation’.
- AI companies frequently alter their models and interfaces without notice, creating unpredictability; Google AI abruptly switched to links-only for election queries in incognito mode mid-study.
- Experts recommend voters rely on official state and local election offices, not AI chatbots, as their primary source for electoral information.
The baseline is improving, but the bar is low
Research from the nonprofit States United Democracy Center, shared exclusively with CyberScoop ahead of publication, tested two widely used and freely accessible tools — OpenAI’s ChatGPT free tier and the Google AI interface embedded in Google Search — across nearly one thousand queries submitted in six US swing states: Arizona, Michigan, North Carolina, Nevada, Pennsylvania and Wisconsin. In 2025, 6.9% of Google AI responses and 8.2% of ChatGPT responses contained verifiable factual errors, including incorrect candidate listings and false guidance on polling locations. Follow-up testing in 2026, limited to Arizona, Pennsylvania and Michigan, found error rates in both models had dropped to zero. The study’s authors acknowledge this represents genuine progress.
Accurate is not the same as complete
The more consequential finding sits beneath those headline numbers. ChatGPT provided incomplete lists of gubernatorial primary candidates 88.9% of the time when queried. Neither tool reliably directed users to official sources: linking to a state election website — which the study identifies as the single most important measure of voter utility — occurred less than 40% of the time. Thania Sanchez, senior vice president of research and analytics at States United Democracy Center, described the practical consequence clearly: the models would name the Republican and Democratic candidates without acknowledging third-party contenders, leaving voters with a materially incomplete picture of their ballot options.
Generative engine optimisation enters the threat landscape
Isabel Linzer, an elections policy analyst at the Center for Democracy and Technology, told CyberScoop that bad actors in the information environment have begun structuring content specifically to rank favourably within AI model responses — a practice she describes as ‘generative engine optimisation.’ Where search engine optimisation targeted Google rankings, this evolution targets the data AI models draw on when constructing answers to user queries. Campaigns, meanwhile, are reportedly thinking about how to format their own materials to appear prominently in AI-generated responses. For security executives, this represents a meaningful shift in how adversarial influence operations can be carried out at scale and with reduced cost.
Model instability compounds the problem
An additional layer of risk comes from the pace at which AI vendors modify their products. In February 2025, between the study’s two testing rounds, Google AI abruptly changed its behaviour for election-related queries submitted through incognito browsing, replacing written summaries with links only. This kind of undisclosed, mid-cycle change makes consistent organisational guidance difficult and creates unpredictability that adversaries can potentially exploit. The study also flags concerns around ideological bias, the potential for personalised or sycophantic responses shaped by a user’s prior chat history, and the general lack of predictability in model outputs for politically sensitive queries.
Usage is growing faster than reliability
A June survey from the Pew Research Center found that approximately half of US adults reported having used an AI chatbot at least once, up from a third in 2024, with a quarter using them daily. Searching for information was the most commonly cited use case. That adoption trajectory, combined with AI being surfaced as the default response layer on tools like Google Search, means voters will encounter AI-generated election content whether or not they seek it out. As Sanchez observed, even a simple Google search now returns an AI overview before anything else.
Why it matters
For CISOs, particularly those in sectors adjacent to government, critical infrastructure or public affairs, this story surfaces three converging risks. First, employees and stakeholders are making decisions based on AI-generated information that may be factually incomplete or adversarially shaped — a social engineering vector that requires no technical exploit. Second, generative engine optimisation is a documented, evolving technique that could be turned toward corporate disinformation as readily as political disinformation. Third, the instability of AI vendor model behaviour means that any internal guidance or approved tooling built around current AI outputs requires regular re-evaluation. Organisations that have not yet developed a formal AI use policy — including guidance on when AI outputs must be verified against authoritative primary sources — are exposed.
What to do now
- Direct staff to consult official state or local election office websites for electoral information rather than AI chatbots, consistent with the study’s expert recommendations.
- Update AI acceptable-use policies to include explicit guidance that AI-generated outputs on high-stakes topics — including civic, legal and regulatory matters — require verification against authoritative primary sources.
- Brief communications and public affairs teams on generative engine optimisation as an emerging influence operation technique, and assess whether your organisation’s public-facing content is structured in ways that could be misrepresented in AI-generated summaries.
- Establish a periodic review cadence for any AI tools in use, accounting for the fact that vendors may alter model behaviour without advance notice, as observed in the Google AI mid-study change.
- Monitor AI vendor release notes and policy updates for tools embedded in productivity and search products, and assess downstream impact on organisational workflows that depend on consistent AI output.
