THE COMING AUTOMATION OF PROPAGANDA
OpenAI, a project Elon Musk founded, made headlines this year for its GPT-2, a text generation language model the organization deemed “too dangerous to release.” This framing was perhaps an exaggeration, but OpenAI’s work was impressive nonetheless. Testers gave the algorithm 40GB of seed text from links aggregated across the Internet, which it studied with the aid of a supercomputer, producing a lightweight output that a regular desktop could run. OpenAI released a toned-down version of the algorithm to the public, but the products the organization revealed of the full version were remarkable. Though OpenAI admits to taking a few tries to get a good sample, given the first line of Orwell’s 1984, “It was a bright cold day in April, and the clocks were striking thirteen,” it eventually produced a coherent opening to a near-future novel set in Seattle. With an opening line about the discovery of unicorns in the Andes, an article GPT-2 produced wouldn’t look at all out of place in a pop-science website. That is, apart from the subject. The “fake news” applications require little imagination. One study explored this exact scenario, showing that GPT-2 was able to generate foreign policy news that subjects rated on average only marginally less credible than the New York Times seed text.
Read more here: