The delve Syndrome
Gabriel Lanyi
I used to encounter the word once in a blue moon, when one of my more adventuresome authors would write something like “The current research delves into the intricacies of parent-child relations during...” I would automatically change delve into to examine or explore and never give it a second thought. Then about a little over a year ago, without warning or provocation, dozens upon dozens of my clients started delving with abandon. Authors who I didn’t think knew what the clumsy import from Dutch meant began delving under layers of conceptual affordances already in the Abstract, resurfacing briefly in the Introduction for a quick paradigm shift only to delve again into the depth of nuanced interplay, not to be seen again until the Discussion.
What’s going on here? Google Scholar returns 56.9K instances of delve since 2020 and 24.4K since 2024. That amounts to 32.5K in the 4 years of 2020–2023 vs. 24.4K in the first 10 months of 2024, or an approximately 400% increase. Some words have done better: unpack logged 22.9K uses since 2020 vs. 18.8K in 2024, a roughly 600% increase, and without counting the packaging industry. Similarly trending among my authors have been nuanced, interplay, multifaceted, negotiate (independent of business), stark, engage with, insight, navigate (unrelated to movements of a craft or vehicle), drawing on, crucial, resonate (not to do with sound), taken together, robust (non-physical), discourse, granular, and of course, paradigm, to name the most... salient.
When I asked Claude and ChatGPT about this phenomenon, the screen turned slightly red as both admitted to having played a pivotal role in launching these terms on a soaring career to stardom. It is understandable why a word like nuanced, vague to the point where it can mean anything, is ranked high by the algorithms that assign weight to the LLM tokens given the distinguished company such words keep in the SciEng with which the big science journals (BSJs) are awash. Words and phrases that appear often in perceived high-quality and influential texts are likely to be overrepresented in the training data of the LLMs. Terms that are commonly used in prestigious circles are reinforced during training by being ascribed higher weights. When the LLM training data is curated, automated and manual processes privilege formal language. Words that appear frequently in well-formed, top-tier content—and BSJs, the home of SciEng, are considered such—are further strengthened as representing sophisticated writing. In answering prompts, LLMs are optimized to predict the next word in a sequence based on the context formed by the previous ones, and such words are naturally more likely to be picked, so “Let’s delve into the intricacies of this topic” would be selected more often than the more pedestrian “Let’s explore this topic.” It is also reasonable that the human annotators who tag and assign metadata to the raw text during the fine-tuning of specialized capabilities of the LLMs have the same biases as the authors published in BSJs and see certain terms as markers of subtle analysis. If they find that texts containing the word delve sound more dignified or authoritative, they may show preference for it in their annotation, which will bias the output of the model in its responses toward this word.
Thus, the terms that the LLMs spread recklessly through the campusphere are already present in disproportionately large numbers in SciEng because the training data of the LLMs reflects existing trends. LLMs merely exaggerate these trends. By constructive interference, mediated by BSJs, LLMs create a feedback loop with academia, reaffirming, augmenting, and disseminating the worst mannerisms of SciEng.
This accounts for certain high-status words ending up on the fast track but I am still puzzled by how interplay can so utterly displace an almost exact synonym, interaction, which is much more common and more precise. Yes, I’m aware of the fine difference, with interaction having to do more with communication and interplay with reciprocal influence between the interacting entities. But interplay has acquired an abstract use as jargon that has come to mean every type of interaction. Here is the LLM assembling an answer one word at a time, as we watch it happen in real time on the screen, at each juncture having to make a decision about the best choice for the next word. And now it comes to a point like “This attitude is the result of the witty...” where a reasonable next word might be interaction or give and take or exchange or interplay or half a dozen other viable options. Yet, it settles on interplay because it is ranked higher, suggesting a more complex relationship between the interlocutors, and it is more, well... nuanced—in other words, it can mean any number of things.
But let us not blame Claude and ChatGPT for the sins of the fathers. What LLMs do is inflate the affectations that SciEng has adopted and spread in the last decades. It was long before the advent of LLMs that impact usurped beneficial and harmful effects alike, desirable as well as catastrophic outcomes, success and failure, and worse, as a verb, it ousted harm, boost, enhance, diminish, improve, damage, advance, impair, and dozens of others. Context now serves for environment, circumstance, setting, phase, climate, and if need be also school, organization, and country. The gradual replacement of real words with placeholders that can mean many things has been going on for some decades in the academic discourse, and it seems to be part of a broader trend of relentless convergence toward singularity in many walks of life. By the time unstoppable processes have run their course and all human labor is carried out by one person encoding algorithms for the Robot General, once we exceed longevity escape velocity and med sci stays one step ahead of death so we all live forever (achieving absolute equality of outcome in healthcare), when all children are awarded a PhD from Harvard at birth (achieving absolute equality of outcome in education), when 100% of the state budget is spent on servicing the debt, when all devices are fused into one gadget, all software subsumed in one app, and all operations are performed by one command that carries out all tasks—just about then, all the words will have merged into a single one that means everything.
P.S. In a move that in qualitative research is referred to as “member checking,” I shared this essay with Claude, who among several astute observations, suggested mentioning possible counterforces that may resist linguistic convergence. I answered that I must give it some thought. Although Claude needed all of two seconds to analyze my essay, I had to take a few days to think over its comments. Claude acknowledged how easy it was to overlook the time differential in AI-human interactions, the gap between the rapid action of the algorithms and the human time it takes to weigh suggestions, consider their implications, reflect, revise, and refine—“where the true value lies,” according to Claude. Tying this idea back to the “delve syndrome,” Claude suggested that driven by AI, language may be changing faster than our ability to critically reflect on these changes. It took Claude a second to make this suggestion. I’d rather not say how long it took me to formulate the following thoughts on trying to stop AI from accelerating our worst practices:
AI is a train that has long left the station and it is now revving exponentially. Trying to slow it would be as foolish and ineffective as it would be to try to speed up our inherently slow biological time. Our only hope is to reconcile the two somehow, making allowances for our slower actions even as AI tears alongside us. And what better means for reconciliation do we have than... you guessed it, AI itself. In the small niche of linguistic processes that make up the topic at hand (an important one, nevertheless, because of the tight coupling between language and thought), we can coöpt AI in a self-referential meta-move to help us identify the instances where language is being impoverished and point out opportunities for enrichment, for example, by flagging instances when framework is standing in for platforms, theories, models, principles, and more. By making this a permanent and automatic feature of AI language processing we can replace the current negative feedback loop with a positive one so that AI, instead of amplifying our most destructive habits, will replace them with constructive ones, then intensify those. This will reclaim our agency and bolster it, as richer language will lead to better thinking, which we sorely need to hold our own vis-à-vis AI.
