OpenAI Disrupts Russian 'Doppelganger' Influence Operation

OpenAI has banned four clusters of accounts associated with the Russian influence operation known as "Doppelganger." These actors used OpenAI models to generate multilingual content for websites and social media to influence audiences in Europe and North America, primarily focusing on the war in Ukraine.

Influence Operation Tactics and Account Clusters

OpenAI identified four distinct clusters of accounts, each employing different tactics, techniques, and procedures (TTPs), suggesting the operation was managed by different functional teams.

Cluster 1: Multilingual Social Media Engagement

This cluster generated short text comments in English, French, German, Italian, and Polish. These comments were posted on 9GAG and X (formerly Twitter) alongside memes and videos.

  • 9GAG Activity: Accounts often used celebrity or cat profile pictures and posted in diverse channels (e.g., "Random," pets, sports, humor, and dating). To create an illusion of support, three to five accounts would reply to each post with simple messages like "hahaha" or "lol." These accounts were typically created on the same date and only engaged with campaign content.
  • X Activity: Accounts typically made only one or two posts. Comments were paired with random-looking links (e.g., "solarpanelfor.sale") that redirected to geofenced websites attributed to Doppelganger. Some errors occurred where English, German, and Polish articles were only accessible via French IP addresses.
  • Evasion Tactics: To bypass OpenAI's supported countries access policy, this cluster used a service linked to companies in Russia and the Czech Republic. OpenAI has since disabled this service's API access.

Cluster 2: Translation and Content Generation

This cluster focused on translating original Russian compositions into English and French. These articles were published on the website rrn.media, a site repeatedly attributed to Doppelganger. The cluster also generated short social media comments to promote these articles.

Cluster 3: French-Language Content Refinement

This cluster was dedicated to generating, proofreading, and correcting French-language articles published on the website franceeteu.today, which is also attributed to Doppelganger.

Cluster 4: German-Language Content Generation

This cluster ingested content from the Doppelganger website grenzezank.com and used it to generate short posts for a Facebook Page and Group named after the website.

Content Themes and Model Constraints

The majority of the content focused on the war in Ukraine, designed to portray Russia positively and the US, NATO, and the EU negatively. Common narratives included:

  • Russia being an essential economic partner for Europe.
  • Western leaders being out of touch with their voters.
  • Ukraine being weak, corrupt, or near defeat.

OpenAI noted that actors attempted to use its models to generate cartoon images of European politicians and critics of Russia, but the models refused these requests.

Impact Assessment

OpenAI assesses that the identified Doppelganger campaigns did not achieve substantial positive engagement from authentic audiences.

  • Engagement Metrics: On 9GAG, posts typically received 5-10 upvotes (mostly from campaign-linked accounts) and 15 to 200 downvotes. On X, posts often had around 1,000 shares with zero likes and zero replies, suggesting inauthentic amplification.
  • Breakout Scale Rating: Using the Brookings Institution's Breakout Scale (1-6), OpenAI categorizes this activity as Category 2. This indicates that while the operation posted across multiple platforms, it failed to achieve "breakout" or significant engagement with authentic audiences.

Sources