Elisabeth Braw
Key PointsGenerative artificial intelligence (AI), which causes confusion among the public through its generation of sophisticated and credible-seeming text, audio, and imagery, poses a considerable threat to societal discourse, especially since it can be used in hostile powers’ disinformation.
Western countries’ legislators are struggling to keep pace with generative AI’s rapid advance. In June 2023, the EU became the first jurisdiction to pass legislation aimed at limiting generative AI’s harm.
At the same time, AI can be useful in detecting gray-zone aggression, which can appear anywhere, anytime, in any shape. Today, countries targeted by gray-zone aggression struggle to identify it at an early stage because doing so primarily involves monitoring by humans.
Introduction
In mid-May 2023, Sam Altman, the CEO of ChatGPT-creator OpenAI, told a congressional committee that one of his “greatest concerns” was the potential for artificial intelligence (AI) to persuade voters and provide “interactive disinformation” in the 2024 US election campaign.1 One might ask why, given these concerns, Altman and OpenAI decided to release their technology. But even if they hadn’t, considering how much damage Russian disinformation caused during the 2016 election campaign, AI-aided falsehoods in upcoming election campaigns are an alarming prospect.
Generative AI, the category to which ChatGPT and other chatbots belong, refers to algorithms often called large language models (LLM), which are capable of generating new and credible content including text, images, and (to a lesser extent) video and audio from training data. During the early months of 2023, such AI caught the public’s attention with the arrival of ChatGPT, a chatbot that composes prose as elegant and informative as that written by humans. The skyrocketing popularity of the tool, which reached 100 million active monthly users within two months of its launch, on November 30, 2022, helped people realize that any written text can now be the work of a robot and that the reader is mostly not in a position to establish whether a written work’s author is a human or a machine.2
Like human-written prose, chatbot writings can include inaccuracies and falsehoods that most people lack the skills to detect.3 AI, though, can help scale such falsehoods: Bots will—quickly, elegantly, and at high volume—write using the information they have been fed, which can be poisoned in different ways. This poisoning can result from poor production and quality of the datasets fed to the system or from data inbreeding, when models are trained on synthetic data—created not by humans but by other generative AIs.
No comments:
Post a Comment