Large language models like ChatGPT are powerful and extremely useful. They can be used for good and for bad. That being said, there is huge potential for misuse of AI large language models. OpenAI collaborated with Georgetown University's Center for Security and Emerging Technology and Standford's Internet Observatory. Together they came to highlight the three main aspects that could be affected by the widespread use of language models in influence operations. The three aspects are actors, behavior, and content.

Making Propaganda Easier

Language models could lower the cost of running influence operations, making them accessible to new actors and potentially giving propagandists-for-hire an advantage. Scaling influence operations and creating personalized content may become easier and more affordable with language models. Additionally, language models could generate more persuasive messaging and make influence operations harder to detect.

It is predicted that language models will be useful for propagandists and will likely transform online influence operations. Even if the most advanced models are kept private or controlled, alternative open-source models may be adopted by propagandists and nation-states may invest in the technology.

However, there are still many uncertainties surrounding the use of language models in influence operations. Factors such as research and commercial investment, availability of user-friendly text generation tools, development of norms, and actor intentions all play a role in shaping the landscape.

Mitigating Risks

To mitigate the risks, the report proposes a framework based on the stages of the language model-to-influence operation pipeline. Mitigations can be applied at each stage, including constructing models that are more fact-sensitive, imposing usage restrictions on language models, coordinating with platforms to identify AI content, and engaging in media literacy campaigns.

It is important to note that the feasibility and potential downsides of each mitigation should be carefully evaluated. Technical feasibility, social feasibility, downside risks, and the impact of the proposed mitigations should all be considered before implementation.

The report aims to stimulate further research and discussion on mitigating the risks of AI-enabled influence operations. The provided framework and guiding questions can assist policymakers and institutions in exploring potential mitigation strategies and weighing their merits.