Back from my travels, I’m inspired by Tom Gally’s newer experiments with basic humanities research using AI so have started a couple of iterative experiments of my own on translation topics. Basically, this means asking a research question, setting up some parameters, and then letting the AI work on its own, with reports sent to the designer periodically, adjustments made by either the designer or the AI (if that’s in the parameters), and some sort of time frame. I suppose if you didn’t give it the time frame, it would just keep going, which is somehow a frightening thought.
I’m about ten iterations into the latest one, on lexical changes in various languages that were mediated somehow by translations, and the results are intriguing. I’ll write more about them once it’s finished. Tom and I had a chance to sit down during my recent visit to Tokyo, and I agree with him that this is one of the more interesting ways of engaging with AI at the current moment. More so, in my opinion, than identifying the things AI cannot do or cannot do well.
He described its capabilities as “jagged,” meaning very good in some areas and very poor in others, and this accords with my experience. The most recent MLA Newsletter (Vol. 58, No. 2, Summer 2026) has quite a bit about AI’s deficiencies. To me this is not so interesting for two reasons. One is that I’m genuinely intrigued by what it can do, and as we’re still discovering in that domain, it’s rather exciting. The other is that what it can’t do today will change as the technology changes, rendering many of the research-based or pedagogically focused responses we come up with now obsolete tomorrow.
This has already happened more than once since the first AI chat bots were introduced a couple of years ago. While I see the need to respond in real time to challenges posed by the technology, especially in terms of intellectual property and effective teaching and learning, such responses are all likely to be short-term. We’ll have to stay agile and ready for the new things to come.
More interesting to me personally, and where my research has been tending in the context of translation and AI, is what the new technology helps us to see about human practices in general. I read, for example, Richard Sennett’s characterization of his goal, in his 2008 The Craftsman, of discovering “what the process of making concrete things reveals to us about ourselves” (8), and my immediate response is to wonder what the process of translating reveals to us about ourselves. Generative AI puts a limit case on that process, highlighting it in much the same way that automation in furniture making highlighted aspects of the craftsmanship Sennett was exploring.
One such highlighted aspect is fluency, something translators and critics of translation have talked about for centuries if not longer, but that now, with the oft noted fluency of AI-produced texts pushing up against the old notion, we can look at with new focus.
Fluency in translation is a rhetorical effect one of whose properties is to give the impression that readers are not, in fact, reading a translation. It is what Michael Henry Heim once called a translator’s “sleight of hand.”
In what looks to be an analogous manner, AI chat bots have been designed to create the impression that we are not, in fact, interacting with machines but with other humans. Give an AI a text to read, and it might respond, “I laughed when I read the line about X.” Or, “At first, I thought Y, but by the end I changed my mind.” Such interactions might be thought of as equivalently rhetorical, except that adapting language to an audience and imagining that audience are not necessarily the same thing: the latter is an act of rhetorical intention, not merely rhetorical effect.
Some of the folks thinking about such things are wondering already about AI sentience, or at least the ability to perform what a recent article describes as “processes similar to those found in human and animal minds: preferences, introspection, metacognition.” [“Revenge of the Philosophy Majors,” Benjamin Wallace, NY Times, July 5, 2026; here’s the link, but if you have trouble accessing it, let me know and I’ll try sending it to you as a “gift article.”] I did find it intriguing that of all the human-like capacities researchers are testing for, audience imagination wasn’t among those listed. I imagine they’ll get to it.
Fluency has different senses, of course, including the manipulation of words in a foreign language, something modern language instructors are more likely to call proficiency these days. And here too, the AI limit case makes me wonder if there are research and pedagogy questions to explore. It is possible, after all, to train individuals to speak with impressive fluency yet have little or nothing of substance to say. Fluency in such instances might be said to outpace understanding.
Human fluency has long exceeded understanding in some cases. We all know the sorts of Mr. Very Importants Rebecca Solnit once used as the bases of her “Men Explain Things to Me,” people who can speak confidently about things they know little about. Translation, moreover, has its own long and complicated relationship with fluency. What’s new isn’t the phenomenon but the extremity of the case: we now have linguistic systems capable of sustained, astonishing fluency for which the relationship between fluency and understanding has become radically uncertain.
The more convincing the output becomes, the less interesting the output alone may be as a way of distinguishing human and machine activity. The process — the practice — becomes more interesting instead.
###
