Researchers Find You Can Train AI to Create Fake UN Speeches in Under 13 Hours

The researchers used a simple deep learning model and cheap cloud storage to complete the project.
Donovan Alexander

New insights in the world of artificial intelligence (AI) seem to always border on the fringes of creepy and "that is fascinating." AI and its application are spreading throughout countless industries and mediums.

AI and its deep learning techniques have been used to create art, detect cancers, generate faces, and even real human voices. Now in the coming age of the AI apocalypse, new insights have given us a peek into what new deep learning can do in a host of areas including politics. 


The idea of an AI politician might sound like something from a B-level season of a science fiction television show or remind you of Bran the Breaker taking the Iron Throne in Game of Thrones ( Think about it. He's basically an AI system). But it is actually a real possibility.

Recently researchers from the United Nations' Global Pulse decided to find out how easy it was for anyone to use relatively easy deep learning techniques to forge convincing misinformation. The results may shock you. 

Spreading Misinformation With Artifical Intelligence 

As part of a recently published study at Cornell University, the researchers used open-source tools and a large collection of data to show how quickly they could get a fake UN speech generator up and running. How did they do it you ask?

The team got ahold of a readily available language model. This commonly used model had been trained on Wikipedia, and later fed 40 years of speeches given by political leaders at the UN General Assembly. 

With no more than $7.80 spent on cloud services and a little more than a half a day (just 13 hours), their AI politician was churning out speeches that would be strong enough to engage crowds at the UN today. The speech generator created speeches that ranged from sensitive and high-stakes topics, including nuclear disarmament and climate change. 

Most Popular

For the research, researchers tested the model on three types of prompts: general topics, opening lines from the UN Secretary-General’s remarks, and inflammatory phrases. The deep learning model was even able to get the cadence down of real UN speeches 90% of the time. 

But, Why?

So why is this research important? AI techniques are more accessible than ever. And, there is a good chance in the near future, with little to no resources, people could create AI speeches and political content to spread misinformation or simply use it for political gains. This research is a call to put more effort into detecting and responding to AI generated content.