Deep-learning techniques have made it easier and easier for anyone to forge convincing misinformation. But just how easy? Two researchers at Global Pulse, an initiative of the United Nations, decided to find out.
In a new paper, they used only open-source tools and data to show how quickly they could get a fake UN speech generator up and running. They used a readily available language model that had been trained on text from Wikipedia and fine-tuned it on all the speeches given by political leaders at the UN General Assembly from 1970 to 2015. Thirteen hours and $7.80 later (spent on cloud computing resources), their model was spitting out realistic speeches on a wide variety of sensitive and high-stakes topics from nuclear disarmament to refugees.
The researchers tested the model on three types of prompts: general topics (e.g. “climate change”), opening lines from the UN Secretary-General’s remarks, and inflammatory phrases (e.g. “immigrants are to blame …”). They found that outputs from the first category closely matched the style and cadence of real UN speeches roughly 90% of the time. Likely because of the diplomatic nature of the training data, outputs from the third category required more work to generate, producing convincing outputs about 60% of the time.
The case study demonstrates the speed and ease with which it’s now possible to disseminate fake news, generate hate speech, and impersonate high-profile figures, with disturbing implications. The researchers conclude that a greater global effort is needed to work on ways of detecting and responding to AI-generated content.