OpenAI’s first threat report offers a clearer look at how foreign influence networks are testing generative AI. The central finding is not that these campaigns have become unstoppable. It is that they are already experimenting with automation, content production, code support, and platform testing.
The report named five different networks that OpenAI identified and shut down between 2023 and 2024. The actors described in the report came from Russia, Iran, China, and Israel, and their activity shows both the appeal and the current limits of AI propaganda.
What OpenAI Says It Found
The report describes established influence operations trying to fold generative AI into their work. Russia’s Doppleganger and China’s Spamoflauge are named as examples of networks exploring how tools like ChatGPT could help automate parts of their operations.
The use cases were practical rather than futuristic. The networks used ChatGPT to create content, debug code, support websites, and help with social media activity. In one case, a network used ChatGPT to debug code intended to automate posts on Telegram, a chat app described in the source as a long-running favorite of extremists and influence networks.
That automation did not always work cleanly. The same account sometimes posted as two separate characters, exposing the operation instead of strengthening it. In other words, the technology helped with scale in some moments, but it also introduced mistakes that made the campaigns easier to spot.
The Technology Still Shows Its Limits
The most important detail in the report may be how imperfect the output was. OpenAI found that these influence campaigns ran into problems with both writing and code. Generative AI did not reliably produce strong copy or dependable technical work for the networks using it.
Language remained a particular weakness. The report says the systems struggled with idioms, which matter because idioms can make writing feel more human, local, and personal. Basic grammar was also a problem, to the point that OpenAI named one network Bad Grammar.
The Bad Grammar network made an especially revealing mistake. It once posted:
“As an AI language model, I am here to assist and provide the desired comment,”
That kind of error undermines the purpose of an influence campaign. Instead of appearing like an authentic person or local voice, the post exposed the machinery behind the account.
How The Campaigns Used AI Online
The report describes several ways these networks used ChatGPT beyond short posts. Spamoflauge used ChatGPT to debug code for a WordPress website. That site published stories attacking members of the Chinese diaspora who were critical of the country’s government.
Other campaigns used AI-generated content for social media and websites. The activity appeared on widely used platforms including X, Facebook, and Instagram, but OpenAI’s report says the material did not break out from the influence networks themselves into the mainstream.
That detail matters. The campaigns were active, but the AI-generated content did not appear to achieve broad reach beyond the networks pushing it. The source also describes campaigns run by an Israeli company seemingly working on a for-hire basis, with content ranging from anti-Qatar to anti-BJP, the Hindu-nationalist party currently in control of the Indian government.
Taken together, the examples point to a pattern:
- Foreign influence actors are trying to use generative AI for content and automation.
- Their current output can be clumsy, repetitive, or technically flawed.
- Some operations are testing websites, social platforms, and automated posting workflows.
- The campaigns described by OpenAI did not appear to reach mainstream audiences through their AI-generated content.
Why Ineffective Campaigns Still Matter
It would be easy to read the report as reassuring. The campaigns described were relatively ineffective, and their propaganda was often crude. That matters in a period when many experts have worried about the power of generative AI to spread mis- and disinformation during a crucial election year.
But the source also makes clear why the issue should not be dismissed. Influence campaigns on social media often improve over time. They learn how platforms work, test what gets removed, and adapt to avoid detection.
Jessica Walton, a researcher with the CyberPeace Institute who has studied Doppleganger’s use of generative AI, described that testing process in practical terms. In her research, the network used real-seeming Facebook profiles to post articles, often around divisive political topics.
She said:
“The actual articles are written by generative AI,”
She also said:
“And mostly what they’re trying to do is see what will fly, what Meta’s algorithms will and won’t be able to catch.”
That is the larger risk. The current campaigns may be weak, but experimentation creates feedback. Each failed post, broken automation attempt, or detected account can teach an operator something about the platform, the model, or the moderation system.
The Takeaway For The Future Of AI Propaganda
OpenAI’s report does not show a world where generative AI has already transformed foreign influence operations into flawless machines. It shows a messier stage: established networks testing tools, making mistakes, and seeing where automation helps.
For now, the limits are visible. The writing can be awkward. The grammar can fail. The code can produce behavior that reveals the operation. The content may stay trapped inside the networks that created it.
Still, the direction is clear. Russia’s Doppleganger, China’s Spamoflauge, Bad Grammar, and the other networks described in the report show that influence actors are not ignoring generative AI. They are learning where it works, where it fails, and how it might fit into future campaigns.
The immediate story is not mastery. It is practice. And in the context of foreign influence operations, even practice deserves close attention.