I Could Write A Book

Over the last year or so, like other academics, I’ve received occasional emails from different publishers asking whether I’d be willing for my work to be included in any deal they might sign with GenAI companies to make their publications available as training data. I do appreciate their position, trying to protect their interests in a less risky and expensive way than suing OpenAI, Meta and the rest into oblivion, but one of the reasons why my response has been and will continue to be negative is the strong sense that the horse has already bolted; they’ve already used it all. Which would make my continued refusal questionable, I admit, if it were a choice between opting in to receive recompense for already-stolen material and opting out to receive no recompense although the material has already been stolen – if I thought there was any serious prospect of substantial recompense trickling down to authors. Which I don’t.

It was interesting, therefore, to see my sense at least partly confirmed, with the news that Meta trained its GenAI using material downloaded from LibGen – depending on your perspective, a giant free library that allows researchers beyond the comfortable West to access research or a giant database of pirated publications. This news came with the bonus of an email from a Meta employee admitting that if they paid to use even one book this would undermine their implausible, self-serving claim that everything they were doing counted as “fair use”. But most entertainingly, it reminded a lot of naive western academics of the existence of things like LibGen, and provided an online search tool for it: if your work was on LibGen, you can take it as read that Meta has stolen it. In my case, I haven’t checked all 137 entries in the database, a lot of which seem to be reviews I’d forgotten I’d written plus a number of wrongly-tagged chapters in volumes I’ve edited, but definitely all but one of my books are in there, maybe all of them, and most of not all of the articles in anglophone journals.

Among the varied reactions and emotions I’ve experienced, one of the more interesting is a feeling of failure and despair, or at least a sense that failure and despair might be appropriate: this machine has read my work, and yet it still produces such crap? Clearly my attempts at conveying ideas and promoting understanding have been fruitless. True, this isn’t confined to GenAI; it does occasionally happen that a student attributes some completely erroneous idea to me on the basis of misreading or misunderstanding – but then I get to add a gentle comment to the essay and use the resulting embarrassment as a teaching moment. ChatGPT is impervious.

Our aspiration as writers is to find the right reader – I’m reminded of the passage in. Italo Calvino’s If On A Winter’s Night A Traveller... where an author is watching a woman on the other side of the valley reading, wishing to be read as the author she is reading is being read – even if we have to create or mould them through our writing (as indeed Calvino shaped me). Maybe we have dreams of reaching thousands or millions, but especially in the academic sphere I imagine we’ve all at some point had the experience of reading something old and/or obscure and suddenly thinking, well, I get this, this writer is speaking to me and their writing has found its true reader, and I can hope that even the most unjustly disregarded article or chapter by me may likewise some day make a difference to the right person.

GenAI training is the absolute opposite of this. Your work is not being read in search of ideas or emotions or experiences that will be unique to reading that author and that text; on the contrary, any such uniqueness is to be filtered out or ground down into homogeneous sludge, in pursuit of the expression of the most statistically average statements possible (hence the tendency, as has been observed, for GenAI summaries to omit original arguments of books or articles because they don’t conform to the database). There is no prospect of GenAI being inspired by something in your words to create something new, to apply your ideas in new contexts or connections – only the theoretical possibility that its statistical regurgitation of the homogeneous sludge will accidentally produce something that looks (if only in the eyes of the occasional washed-up novelist or academic desperate to jump on the bandwagon for attention) as if it is genuinely new and interesting.

The death of the author, for Roland Barthes, was intended to be the birth of the reader. GenAI gives us the shambling undead version of both. Books are not read but merely processed, so that the most perfectly average regurgitation of cliché can be generated to stand in for writing. The romantic dream is that some image, some turn of phrase might unexpectedly touch the heart of the most savage or degenerate listener, but you might as well hope for the meat grinder to be affected by the sensibility of the young bullock chewing an especially crisp dandelion plant on a misty morning, or for the resultant sausage to communicate this. And the GenAI sausage isn’t even a good sausage.

The romantic dream is that the robot is jolted out of its programming to recognise something of the human; instead, the human is becoming as indifferent to meaning and specificity as the robot. Calvino’s brilliant novel includes the idea that novels might be mechanically analysed, dissected and summarised, and it is clear that this entirely misses the point of either writing or reading – not least because it entirely ignores the possibility of pleasure. For GenAI, and its worshippers and enablers and users, reading is nothing more than the accumulation of data about statistically meaningful relationships between discrete combinations of letters, and writing is nothing more than the generation of data that echoes these patterns. And there is no health in us.