Why Most AI Writing Can’t Get Its Facts Straight
The brute force approach to artificial intelligence writing still works better on fiction than on factual, data-based reporting.
Get me rewrite!
Photographer: Imagno/Hulton Archive via Getty Images
It’s been almost a year since OpenAI, the San-Francisco lab co-founded by Elon Musk, released Generative Pre-trained Transformer 3, the language model that can produce astoundingly coherent text with minimal human prompting — enough time to draw some conclusions on whether its brute-force approach to artificial intelligence can in time allow most writing to be delegated to machines. In my current job at Bloomberg News Automation, I’m in the business of such delegation, and I have my doubts that the trail blazed by GPT-3 leads in the right direction.
In these past months, lots of people have tested GPT-3, often with surprising results like these fake Neil Gaiman and Terry Pratchett stories or these “Dr. Seuss” poems about Elon Musk — or these perfectly readable newspaper columns, clearly published by editors both in awe of the new technology and relieved that AI wouldn’t be taking away their jobs any time soon.
It’s taken me a while to figure out what all these GPT-3 products resemble, and now I know: A monologue from the classic play by Nikolai Gogol, “The Inspector General.” The central character, a complete nonentity named Ivan Khlestakov, arrives in a provincial town and is taken by its elite for a high-ranking government inspector about to conduct a secret investigation into their shady affairs. Khlestakov, fired up by the red carpet treatment, the free-flowing champagne and the attentions of the town’s eligible ladies, lets loose a self-aggrandizing tirade (here in Arthur Sykes’ translation):