Early this week, I had the pleasure of participating (and I use that word broadly) in a seminar focused on teaching with ChatGPT. The participants in the seminar were really outstanding and shared range of practical and theoretical approaches to teaching with ChatGPT. I was frankly blown away by the thoughtfulness and expertise on the panel and I struggled to engage with it entirely (despite sitting awkwardly on the “digital stage”).
Here’s what I had thought about in the lead up to the panel.
One thing that came up at the end of the panel that got me thinking a bit differently about ChatGPT is that it can only offer responses based on what it reads and consumes. On some level, this means that we have to learn to write (and disseminate) information for a new class of “artificial” readers.
This isn’t an entirely new observation. In fact, we have already understood some of this which occurs under the banner of “search engine optimization,” but it strikes me that if ChatGPT can deploy its language model to produce text, it’s not a leap to recognize that the same language model could be used to assess the content upon which its responses to queries are based. In other words, ChatGPT, and its future iterations, is a reader that, like all readers, performs its task based on an algorithm that presumably adjusts to the content available.
The question then as a writer is how do we make our work appealing or maybe merely susceptible to our new ChatBOT aggregators. Surely, these kind of bots will have less interest in deliberate displays of opacity, ambiguity, or playfulness. We might even be able to retire for good the need for the compelling (or even slightly misleading) lede which students so often turn into the cringeworthy first sentence. It also calls into question the value of such awkward stylistic crutches as the “rhetorical question.” At the same time, one wonder how it assesses the presence of “irony” in ascertaining the authority or utility of a text. A query to ChatGPT tells us that it discerns irony though linguistic features, contextual clues, semantic analysis, and, perhaps most importantly, machine learning, which relies on texts that humans have marked up to allow the AI to understand what irony looks like in practice.
On a more basic level access to texts will surely impact how these AI bots formulate their answers. For generations publishers have sought to monetize their texts by limiting access to them and recognizing that the there remains a balance between cultivating the influence of a text and capitalizing on those who need access to it. One wonder whether in the future, the scope, speed, and reach of AI readers will mean that a text that is hard to access — behind a paywall, written in a non-English language, or even is simply opaque in meaning — will limit its influence in the kind of language models that these AI readers rely on to produce new text. In other words, the presence of quick-reading AI bots will accelerate the importance of open access bodies of texts which will almost certainly gain a greater influence over the language models that shape the ability of AI bots to “think.”
There are those who see a future where publishers are less inclined to charge for access to individual publications. Pay-wall barriers can make it harder to for automated processes to aggregate information across a wide range of sources and as aggregators increasingly serve to privilege certain sources above others (and to amplify certain works over others). Of course, there are ways to let bots in and impede the movement of human eyes, but it also stands to reason that AI-powered aggregators will invariably draw more freely on content that is more easy to access and plentiful on the web.
In light of this situation, some have proposed an alternative business model that see publishers providing aggregation services, likely powered by AI bots, that assess the significance of publications in a field, provide answers to research queries, or even prepare regular literature reviews for scholars. These aggregation services, of course, will come at a cost and will likely introduce certain biases into the results that they aggregate, but will provide a revenue stream for the publishers. More than that, as publishers shift the cost of publishing from the readers (via subscriptions) to authors (via subventions), the cost of publishing could become a way to ensure that an article rises to the top of an aggregator’s trawl through recent publications.