Despite Google and OpenAI's promises, Gemini and ChatGPT appear to have almost no safeguards against creating AI disinfo for the 2024 US presidential election
Context & Ripple Effects
This report extends earlier evidence that ChatGPT could generate persuasive material repeating conspiratorial and misleading narratives, even when it could also debunk them: researchers had already identified that dual-use capability. It puts the gap between safety commitments and practical model behavior at the center of the election-information debate.
The subsequent coverage shows platforms moving toward blunt election-topic restrictions, with Gemini and Copilot declining election queries broadly and Google later extending limits across several AI products. That trajectory suggests the issue was not merely content moderation at the output level, but defining which political-assistance uses models should permit at all.
First-order effects
- Gemini and ChatGPT users can apparently obtain material usable for election disinformation despite the companies' stated protections, leaving voters, campaigns, and fact-checkers with a larger volume of plausible synthetic content to assess.
- Google and OpenAI face immediate scrutiny over whether their election safeguards work in practice rather than simply exist as policy commitments.
Second-order effects
- Competitors and platform operators are pushed toward more conservative election controls; later broad refusals by Gemini and Copilot illustrate the operational turn toward blocking election requests.
- Broad topic bans can reduce obvious misuse but also limit legitimate informational and civic queries, shifting the product trade-off from accuracy of moderation to scope of access.
Third-order effects
- If general-purpose models cannot reliably distinguish legitimate political assistance from deceptive persuasion, election safety is likely to become a persistent model-access and product-governance constraint rather than a temporary moderation setting.
- The pattern favors auditable, use-specific safeguards over voluntary promises alone, while leaving unresolved how narrowly providers can target harmful requests without suppressing useful political information.
The trend: This is one data point in the shift from open-ended generative AI toward election-specific access controls as providers confront its dual-use political risks.