Is there greater cause to worry further down the literary food chain? There have for a while already been “AI bots” that can, we hear, “write” news stories. All these are, though, are giant automated plagiarism machines that mash together bits of news stories written by human beings. As so often, what is promoted as a magical technological advance depends on appropriating the labour of humans, rendered invisible by AI rhetoric. When a human writer commits plagiarism, that is a serious matter. But when humans get together and write a computer program that commits plagiarism, that is progress.
As a news reporter, GPT2 is, to put it generously, rather Trumpian. Fed the first line of a Brexit story – “Brexit has already cost the UK economy at least £80bn since the EU referendum” – it went on a nutty free-associative spree that warned, among other things: “The UK could lose up to 30% of its top 10 universities in future.” (“Up to 30% of the top 10” is a rather roundabout way of saying maybe three.) Brexit, the machine continued, will push “many of our most talented brains out the country on to campuses in the developing world” (eh?), and replacing “lost international talent from overseas” would, according to “research by Oxford University”, cost “nearly $1 trillion”. To which one can only properly respond: Project Fear! To their credit, the machine’s masters at OpenAI admit that it is sometimes prone to what they call “world-modelling failures”, “eg the model sometimes writes about fires happening underwater”.
The makers’ announcement that this program is too dangerous to be released is excellent PR, then, but hardly persuasive. Such code, OpenAI warns, could be used to “generate misleading news articles”, but there is no shortage of made-up news written by actual humans working for troll factories. The point of the term “deepfakes” is that they are fakes that go deeper than prose, which anyone can fake. Much more dangerous than disinformation clumsily written by a computer are the real “deepfakes” in visual media that respectable researchers are eagerly working on right now. When video of any kind can be generated that is indistinguishable from real documentary evidence – so that a public figure, for example, can be made to say words they never said – then we’ll be in a world of trouble. OpenAI agrees that this is a larger problem, even if its proposed remedy is rather vague. To prevent what it calls “malicious actors” from exploiting such technology, it says, we “should seek to create better technical and non-technical countermeasures”. Arguably this is like engineering a bioweapon and its antidote at the same time, rather than choosing not to invent it in the first place.
A Profitable Year for Trade Publishers
Profits rose in 2018 over 2017 at four of the five major trade publishers that report their financial information—and the one company at which earnings were down, Lagardère Publishing, reported an increase in profits at its U.S. subsidiary, Hachette Book Group. Operating margins fell at Lagardère and slipped by a 10th of a percentage point at Penguin Random House but rose at the three other companies. Solid backlist sales and strong sales of digital audio were cited by four of the five publishers as key revenue drivers in the year, though sales of e-books were down.