I have been thinking a lot about the form our research takes recently. One such thought-in-formation below.
Knowledge Zombies
Rebecca Roach
The ‘textpocalpyse’ is what critic Matthew Kirschenbaum has called it: the explosion of digital text that generative AI is facilitating. The ChatGPTs of the world ‘have brought something like autocomplete to the entirety of the internet’. In the arena of culture, it’s the writers and artists who are the first to fall, if the (AI generated?) headlines are anything to go by today. But there are broader, more urgent, implications too: will we all become knowledge zombies?
The canary in the coal-mine in this instance is a behemouth. As far as Google is concerned, the textpocalpyse has well and truly arrived, and with it an existential threat for the search giant. Google might have won the first round of the “search wars” (remember Yahoo and AltaVista?) in the noughties, but generative AI has set off a whole new skirmish—one that threatens our epistemological order.
Crucially, the textpocalypse is creating problems for search engine optimisation: a war of attrition between the text-generating bots and the search engineers’ ability to algorithmically evaluate content (and sift and devalue AI-generated text). If generative AI is the problem, it is also the solution, at least in the eyes of Google, Bing, and other combatants in these new search wars.
In May 2024 Google introduced ‘AI Overview’, which offers to ‘take the work out of searching by providing an AI-generated snapshot with key information and links to dig deeper.’ (It insists on the feature as you cannot turn it off, to the chagrin of many a forum user.)
If I google ‘what was Project Xanadu?’, for example, ahead of the usual search results (sponsored or not) I am met with a summary ‘overview’ box that helpfully informs me:
Project Xanadu was a hypertext system and publishing tool created by Ted Nelson in 1960. It was a predecessor to the World Wide Web and was intended to be a universal library that would allow people to access all information. Nelson believed that Xanadu would enable computers to manipulate media in a way that would free people’s minds and actions.
So far so good. At the end of this ‘overview’ there is a link icon which, if expanded, provides four, seemingly random links: one to a paywalled article by Wired; one to a scholarly article on Project Xanadu and the ‘politics of new media’; one to a landing page which provides an AI generated overview of scholarly ‘papers important to this topic’ (though in actual fact they are completely unrelated); and one to a PDF of a thesis abstract from the digital repository of Charles University in Prague, with absolutely no other contextual details.
It is never made clear how these links relate to the text generated by the AI Overview and how these citations have been selected; that is proprietary information. The link selection certainly doesn’t suggest an underlying logic to the human user, aside from an overreliance on questionable scholarly sourcing. If you were hoping for authoritative citation, the link results are more befuddling than the generated text.
My point here isn’t really to criticise such ‘experimental’ results. (There is a helpful sentence at the end of the generated text reminding the user that ‘AI Overview is experimental’, as if it was avant-garde poetry.) Rather, it is what such an experimental form does epistemologically that worries me.
In some respects, AI Overview isn’t particularly novel. Adobe Acrobat and Microsoft Outlook now offer to summarise my documents, whether I want it or not (and I don’t). AI Overview is also the most recent iteration of a longer Google program, Knowledge Panels: ‘information boxes that appear on Google when you search for entities’ which are ‘meant to help you get a quick snapshot of information on a topic based on Google’s understanding of available content on the web.’ Such a statement begs the question of what exactly a ‘snapshot’ might be in a textual context, and whether or not we want our searches mediated by ‘Google’s understanding’ (What is that? Is Google the corporate entity here or the search algorithm? Either way such ‘understanding’ doesn’t appear authoritative.)
The Knowledge Panel feature has rightly come under a fair amount of scrutiny since its launch in 2012: they ‘materialize at random, as unsourced and absolute as if handed down by God’ as one reporter put it. Such criticism around their lack of attribution and verification processes is compounded by the opacity around precisely how Google’s Knowledge Panels are formulated: source data, algorithms utilised, prioritisation, etc.
AI Overview exacerbates these anxieties. Google blithely suggests the feature offers ‘key information and links to dig deeper’ but how that ‘key information’ is determined, how those links are prioritised, remains proprietary. In an echo of the corporate-speak, whereby a grand mission to ‘organize the world’s information and make it universally accessible and useful’ masks a more commercial effort to improve search function, AI Overview is perhaps the manifestation of an new mantra: ‘Information wants to free, from citation.’
AI Overview and the other weapons in the new search wars seem to be sowing the seeds for the demise of the web as we use it today. In the fight against the textpocalypse, Google propose that we stop navigating the hypertextual battlefield at all. Links might help us ‘dig deeper’, but with AI Overview the user need no longer follow the link and check the citation. We can stay on Google’s server forever, going nowhere.
That’s convenient for some users, saved from the bother of clicking a link and waiting for a website to load. How wonderful: we’ve lowered the environmental footprint of search, if it weren’t for the enormous processing inefficiencies of AI. Yet it isn’t so helpful for those websites that are hosting the information, providing the content for the large language models which AI Overview exploits.
Google, seemingly, is undermining its own business model: its search revenue has historically derived from advertising—companies and services pay for the privilege of users clicking on links, of following the citation trail. Is Google now gaslighting us, or trying to burn it all down from within? The company and its competitors recognise the problem: revenue sharing deals with media companies are being trialled, that are based around citation. It’s a radical shift from the ‘transformative fair use’ arguments that legislated Google Books’ use of digital content a decade ago.
The stakes though, are bigger than that: AI Overview and the new search wars are undermining user access and experience of the web and its terrain of knowledge. When Ted Nelson proposed his grandiose vision of Project Xanadu, a hypertextual and interconnected trove of all the world’s information, it was premised upon the link, on the ability to make citational connections. Search engines historically have functioned with this model at their centre: on the idea that information is organised, accessed, authenticated, and corroborated citationally. In the textpocalypse, it is the link—and the citational structuring of knowledge that it represents—that is under threat. In the new search engine wars prediction trumps citation.
Texts might be slain, but at what cost? We are now the knowledge zombies.