FREEDOM ON THE NET 2023 The Repressive Power of Artificial Intelligence Generative AI Supercharges Disinformation D isinformation campaigns have become a regular feature of the global information space. Over the past 10 years, Freedom on the Net’s indicator on content manipulation, which measures the extent to which online sources of information are controlled or distorted by the government or other powerful political actors, has declined more on average than any other indicator in the report. These campaigns have long been assisted by AI technology for distributing content: automated “bot” accounts on social media have smeared activists into silence and propagated false narratives about electoral fraud to voters, while platform algorithms have promoted untrustworthy and incendiary information over reliable sources. However, much of the work is still done by humans. During the coverage period, at least 47 countries featured progovernment commentators who used deceitful or covert tactics to manipulate online information, double the number from a decade ago. An entire market of for-hire services has GLOSSARY OF AI TERMS Artificial intelligence: A field of computer science in which people build computer programs known as “models” that analyze massive datasets in order to predict, interpret, classify, and generate information. Machine learning: A subfield of artificial intelligence in which computers learn without a human having to program exact instructions. Machine learning models are “trained” (meaning, developed) to learn patterns based on analysis of large sets of data. Deep learning: A subfield of machine learning that involves models learning in layers, building simpler patterns into more complex ones. This approach has enabled many recent AI advances, such as recognizing objects in images. Training data: Data—spreadsheets, text, images, audio, or video recordings—that people use to train machine learning or deep learning models. Generative AI: Deep learning models that specialize in generating content, including images, text, video, and audio, in response to a specific query or prompt. Large language model (LLM): A deep learning model trained on vast datasets of text, with the capacity 8 @freedomonthenet to generate sophisticated and realistic text outputs. Examples include OpenAI’s GPT-4, Meta’s LLaMA, and Anthropic’s Claude. Computer vision models: A deep learning model trained on vast datasets of images, with the capacity to generate sophisticated and realistic image outputs. Examples include Midjourney and OpenAI’s DALL-E. AI chatbots: Programs that employ AI to understand text-based prompts and generate responses to those prompts. Chatbots built on LLMs include OpenAI’s ChatGPT and Google’s Bard. Bots: Automated programs that perform prescribed tasks, often with greater speed and at a larger scale than their human counterparts. Synthetic media: Images, video, or audio content generated by AI—usually by LLMs, computer vision models, and deep learning–based audio models. Deepfakes: Photo, video, or audio content that has been created or manipulated (usually via AI) to convincingly portray actions or events that did not in fact occur. #FreedomOnTheNet