OpenAI has taken action against multiple influence operations that leveraged its AI models to generate content for various political objectives. These operations include activity related to the Ghana 2024 presidential election, a multilingual effort dubbed "A2Z" focusing on Azerbaijan, and two China-origin campaigns, "Sponsored Discontent" and "Uncle Spam," targeting US politics and Latin American news outlets.
In the Ghana election case, accounts were banned before the election was held. The "A2Z" operation, which used OpenAI's API, was identified in October 2024. The "Sponsored Discontent" and "Ghana election activity" case studies were published in OpenAI's February 2025 report, while "Uncle Spam" was detailed in its June 2025 report.
Key Points
- OpenAI banned ChatGPT accounts generating articles, posts, and fake engagement for Ghana's 2024 presidential election, supporting one candidate and criticizing another.
- The Ghana operation, linked to DigitSol, used ChatGPT to create English-language content for a website and social media across Facebook, Instagram, X, YouTube, TikTok, and WhatsApp.
- The "A2Z" operation used OpenAI's API to generate comments in multiple languages for X and Facebook, often praising Azerbaijan or defending its human rights record.
- "Sponsored Discontent," a likely China-origin operation, used ChatGPT to generate short English comments and long-form Spanish articles, some of which were published by Latin American news sites.
- The "Uncle Spam" operation, also China-origin, used OpenAI models to generate polarized US political content and AI-generated profile images for platforms like X and Bluesky.
- Both the Ghana and "Uncle Spam" operations were assessed as Category 2 on the Breakout Scale, indicating activity on multiple platforms but minimal authentic engagement.
- The "A2Z" operation used AI to manage fake social media personas, generate bios, analyze posts, draft replies in many languages, and proofread content.
Context
According to OpenAI Safety, the Ghana election activity involved content supporting Vice-President Mahumudu Bawumia and criticizing former President John Mahama. The operation used ChatGPT to generate content ranging from four-word comments to lengthy news articles. The "A2Z" operation, as described by OpenAI Safety, leveraged AI to manage numerous social media accounts and engage in comment threads with real people. OpenAI Safety also noted that the "Sponsored Discontent" operation used ChatGPT to translate and expand publicly available Chinese-language articles into Spanish, which were then published by Latin American news websites.
Why It Matters
These disruptions highlight the ongoing challenges for AI developers in preventing malicious actors from misusing advanced models for influence operations. Builders must consider how their models could be exploited for generating deceptive content, managing fake personas, and attempting to manipulate public discourse, particularly around sensitive events like elections.
What To Do
- Note the range of content types (short comments, long articles, images) and platforms (multiple social media, dedicated websites) used in these operations.
- Review the stated methods of detection and disruption employed by OpenAI, such as account banning and API monitoring.
- Consider the "Breakout Scale" used by OpenAI to assess the impact of influence operations, particularly the distinction between generated content and authentic engagement.
- Watch for further reports from OpenAI and other organizations on evolving tactics used in AI-powered influence operations.