OpenAI releases GPT2 to the public after banging on for ages that it's a bad idea

Chris Merriman
@ChrisTheDJ
08 November 2019

Tweet
Facebook
Send to

ARTIFICIAL INTELLIGENCE CONSORTIUM OpenAI has done the unthinkable and opened the 'Pandora's Box' it swore it would never do.

GPT2 is an AI tool that writes such plausible-sounding 'fake news' that it was considered dangerous to allow it into the public domain.

The tool was originally designed for benign purposes, in fact, Auntie Beeb will be using a similar, though less advanced, tool on the night of the General Election to speed up reporting of constituency declarations.

In addition, you may recall that GPT2 has been used in some of Janelle Shane's AI experiments.

The trouble is, it has proved too good, able to write just the sort of stories that have made democracy such a minefield in recent years.

Up to now, only paired-down versions of GPT-2 have been available publicly, while OpenAI keeps a close watch on how they're being used and how successfully they are working.

However, in a blog post this week, the company announced that there had been no real evidence of the tool being misused: "…we haven't seen evidence of writing code, documentation, or instances of misuse. We think synthetic text generators have a higher chance of being misused if their outputs become more reliable and coherent."

That feels a bit of a, shall we say, bold statement, because it effectively gives with one hand and takes away with the other. Anyone who knows their mythology knows that once Pandora's Box is open, it's more difficult to get the lid closed again than that time you were at check-in and had to ask three passers-by to sit on your suitcase.

Other conclusions of the trials included an acknowledgement of the need to develop common standards for analysing bias and its effects. It also admits that detecting a story written by GPT-2 is "challenging" and that it can be fine-tuned for misuse of the type we saw during recent elections in both the UK and US.

All of which gives rise to one question. What the actual eff, OpenAI? - You've just essentially listed all the reasons for not releasing it, and then released it anyway because no one is misusing it YET.

That's a bit like not plugging a serious code vulnerability because it's not zero-day YET. Still, it's done now. Trust no-one. Except us, obviously. µ