Elon Musk-backed non-profit OpenAI, which claimed to have developed a machine learning-powered text generation software so powerful that it couldn’t be ethically released to the public, has… done just that.
OpenAI previously said that the text generator, using the GPT-2 software and trained on some 40 gigabytes of data yanked from eight million websites, was so good at generating text from a given prompt that it could be used to generate convincing hoax stories. The Guardian described it as “capable of writing plausible passages that match what it is given in both style and subject,” as well as significantly better for general usage purposes than preceding models. Musk is one of OpenAI’s primary funders, though he has been outspoken about his belief that AI poses an existential threat to humanity and reduced his role at the organization last year.
In a blog post this week, OpenAI said that although it had determined humans find the outputs of their trained GPT-2 model convincing and it was capable of creating “synthetic propaganda,” the nonprofit had found no strong evidence of misuse. Users can now access the full-sized GPT-2 text translator on TalkToTransformer.com.
While the GPT-2 model does generate comprehensible text in a reasonably correct tone and style, and it’s easy to imagine situations in which it could be misused… This is real “your mileage may vary” territory. It’s hard to describe it as particularly dangerous in its current form (though Gizmodo did get it to spit out an eerily good fake Federal Communications Commission letter on net neutrality).
“The words ‘too dangerous’ were casually thrown out here without a lot of thought or experimentation,” researcher Delip Rao told the Verge. “I don’t think [OpenAI] spent enough time proving it was actually dangerous.”
On the other hand, GPT-2 has just enough grasp of the English language to be extremely funny (at least until it gets good enough to help usher in an even more dystopian future). We took it for a spin, and honestly, it was hard to stop. Check out some examples below: