“Deepfaking” AI riffs on Orwell and invents believable news pieces

Charlie Spargo's picture

Artificially intelligent text generator GPT2, revealed by nonprofit research organisation OpenAI, will be held back from public release due to its potential for malicious misuse.  

GPT2 receives input and can accurately continue writing the passage, basing its output on structure and word use. It’s so good at what it does, it can create fiction and news barely distinguishable from the real thing.

OpenAI generally releases full research to the public, but is worried about misuse in the form of falsified news stories, fake positive - or negative - reviews, and spam. They will use the time to discuss and weigh up the potential consequences of release, and come to a decision.

It’s been used to create an output from the first line of 1984 by George Orwell - coming up with a story in a futuristic setting and with the author’s flair. It also drafted passages based on a prompt featuring characters from Lord of the Rings, and one beginning “Miley Cyrus was caught shoplifting from Abercrombie and Fitch”.

Researchers used a dataset 15 times larger than any other state-of-the-art AI system had drawn on before. Using the online content aggregator Reddit, the system was trained using links receiving more than three upvotes - a grand total of around 10 million articles, and about 40GB of information.

The content it creates have been described as reminiscent of “deepfakes”, in reference to the technique of synthesising human faces in video and creating realistic clips of events and comments that never actually occurred.