The mediocrity of AI and deskilling
how even mediocre AI will replace novice humans and deskill the human workforce
Since the release of ChatGPT, many people have commented that its output is, for the most part, quite run-of-the-mill and mediocre. For example, an essay generated by large language models (LLMs) has a mechanical essay-by-numbers quality to it.
But to conclude from this that LLMs will not replace humans is an error. Most human content is also run-of-the-mill and mediocre, and this is not a bug but a feature. There are plenty of tasks that do not require anything but a “by numbers” output where LLMs already provide acceptable performance, e.g. summarisation of longer content, online ads and SEO content.
LLMs now possess the skill level of a novice human in many domains where the output is text. As I argued in a previous essay, it is also reasonable to argue that the current approach will likely be unable to generate truly novel material and that it is restricted to wandering within the known domain without being able to explore beyond it. This may give some comfort to experts that their economic value is not under threat just yet.
However, most tasks in most domains are not expert-level tasks. Even the criticism that an LLM that performs these tasks needs to be supervised and its output monitored by an expert human is not as severe a drawback as it seems. After all, this is exactly how many human teams work anyway, i.e. an expert human tells a novice to perform a specified task, reviews the work, and corrects/edits the work. This is precisely why products like GitHub Copilot are already successful. In fact, an AI “assistant” of this nature is usually much more time-efficient to manage than a human novice.
But there is a larger problem with replacing novice humans with a novice AI. As I explained in another essay, although today's experts may be safe from being replaced by a machine, AI and automation short-circuit the learning process by which human novices become experts. Going back to the example of the five-paragraph essay by numbers, such an essay is a simple stepping stone that enables a novice writer to become a better writer.
As the below essay highlights, there is nothing new about this, and many professionals, from financial markets traders to aircraft pilots, have already experienced this. What we are now seeing that is different is that
machines can now perform not just codifiable/legible tasks but also the illegible tasks that are the bread and butter of the jobs that comprise the “knowledge economy”. What has already happened in quantitative domains where numbers are the output will now take place where words and images are the output.
The problem is not insurmountable. For example, flight simulators are quite effective in tackling the problem of training pilots. But the problem will not solve itself. The default path of the near future is a gradual deskilling of the human workforce across domains, which paradoxically also makes it much easier for machines to replace us.