Large language models (LLMs) are causing consternation, awe, and eye rolling, depending on one’s crisis related to their deployment. Which “future-proof” career is going to implode next? Lawyers, copywriters, college students, emotional supports?
This did not happen because it was inevitable. Often the setup to such crises become problematic before automation joins the fray. Let’s look at writing, presently considered under threat by LLMs.
If you were born in the 80s or later, you might have learned about the Flesch-Kincaid readability test. It spat out a numerical readability score that measured writing complexity through word and sentence count and length. A short sentence with short words had high readability. A long sentence with long words had low readability. The resulting numerical score corresponded to an approximate education level required to read the text.
Flesch-Kincaid became the BMI of publishing: a semi-useful measure taken completely out of context. The score became a hard boundary that writing had to clear for publication.
Blind devotion to the readability score often made writing less comprehensible and hamstrung writing in industries that relied on more academic and industry-specific language. The test skewed language choice away from words of Latin and Greek origin and toward words of Germanic origin, as the latter tend to have fewer syllables than the former.
The first iterations of content media were guided by Flesch-Kincaid readability cutoffs and SEO. They were so distasteful because strict compliance with readability scores and SEO detracts from the reading experience.
There are a couple forces at work here. First, when readability scores take priority, 1) the human writer is forced to write something a well-educated human writer never would, and 2) the human writer is forced to conform to parameters that force an output not unlike a machine-mediated output. When these parameters are invisible to human readers (except through the sudden mediocrity of human output), it renders invisible the full breadth of human ability and makes machine writing look like fair competition. This invisible restraint of human factors in writing can lead to performance graded on a curve that favors factors we associate with machines rather than humanity.
The major problem is that automation creates an illusion that a machine can act on its own, and proponents of AI are loathe to clarify the actual human labor needed to make an LLM work. That may change someday, but today, machines are still tools that require a lot of human input, even if we pretend otherwise. When human performance is graded on a curve that rewards tool capabilities over human abilities, the activity of creation stops being a process whereby humans use tools. Instead, humans appear to compete with their tools, with very little said about what would actually happen if humans were banished from the process.
Over time, tools have become more complex, and economic forces have fueled a paradigm of valuing what tools can do with human intervention over what humans can do with tools. This has led to a decline in perceived competence of workers and increased perceived competence of tools.
As tools become more complex, expectations of what tools are and what tools do also change. However, these changes in perceived relative competence only happen in a society where there is a preexisting paradigm that favors machines or there is a secondary gain to forgetting human talents.