Language models could make propaganda cheaper to produce, easier to scale and more varied in its wording. A paper by OpenAI, Georgetown University's Center for Security and Emerging Technology and the Stanford Internet Observatory sets out how those tools might affect disinformation campaigns—and how policymakers can assess possible responses.
From model development to public belief
The paper organizes the risks across four stages: building a model, gaining access to it, distributing content and shaping what people believe. Different actors could influence each stage, so the authors' framework looks beyond the technology itself to the path from a model's creation to its possible effects on audiences.
That framing matters because a tool can be restricted at one point and still be used elsewhere in the chain. Controls on access, for example, do not by themselves address how content is distributed or how people respond to it. The framework encourages policymakers to consider the whole process rather than treat a language model as the only point where intervention is possible.
Why language models could change influence campaigns
The research team warns that widely used models could automate parts of disinformation work and lower the cost of producing propaganda. If content takes less effort to create, campaigns may be easier to expand. Chatbots could also generate propaganda in real time, opening up tactics that depend on an ongoing exchange.
Language models may also produce messages that are more persuasive or varied than repeated, copied posts. Variety could make propaganda harder to identify and debunk than a set of identical messages. These are potential capabilities described by the paper, not a claim that every campaign will use them or that every model will have the same effect.
The authors expect language models to become useful to propagandists and to change online influence operations. Keeping the most advanced models private or limiting them to API access may not remove the risk: propagandists could turn to open-source alternatives, while nation states could invest in the technology themselves.
Four questions for weighing safeguards
The paper does not endorse or rank individual mitigations. It instead proposes four questions to help policymakers consider whether a response is workable and worthwhile:
- Technical feasibility: Can the mitigation be built and put into practice, and would it require major changes to technical infrastructure?
- Social feasibility: Can political, legal and institutional actors implement it? Would it require costly coordination, and can it be acted on under existing law, regulation and industry standards?
- Downside risk: What negative effects could the mitigation cause, and how serious might they be?
- Impact: How much would it reduce the threat?
Together, these questions point to a practical trade-off: a measure needs to be possible to carry out, acceptable to the people and institutions involved, proportionate in its side effects and effective against the risk. The paper presents this as a way to guide discussion, not a ready-made policy prescription.
Important questions remain open
The research team says many uncertainties remain. It is not yet clear which actors will invest in which capabilities, what new abilities language models may develop, or how accessible those systems will become. The authors also raise the possibility that new norms could discourage AI-assisted propaganda.
Those unknowns limit how confidently policymakers can predict the shape of future campaigns. They also make it useful to examine safeguards before a particular intervention is chosen: a proposal that sounds plausible may still be difficult to implement, carry meaningful downsides or have little effect.
The paper describes itself as a basis for further discussion and is “far from the final word.” Its central contribution is a way to map the stages where AI could enter influence operations and a set of questions for evaluating possible responses. As language models become more widely available, those questions can help keep the debate focused on both the threat and the consequences of attempts to reduce it.