I have been hearing from a lot of corners about the threat of ChatGPT automating much of what is currently written by humans.
But I am a skeptic on this front for a couple of reasons.
No doubt there are some rote, boilerplate writing tasks—like the bog-standard student “essay”—that ChatGPT replicates with ease, but as a self-employed writer who gets paid primarily to think on the page, I am not worried about a syntax-regurgitation machine that cannot actually think.
My other source of skepticism is that true automation always seems to be many years in the future from when it is first promised as imminent. See: self-driving cars, for example, a problem that some folks believe cannot be solved by the current method most commonly used in cars like those produced by Tesla.
See also: this personal anecdote from my first exposure to the promise of automation that utterly failed to deliver, a failure that was foreseeable from the start and may be instructive as to how we think about the role of ChatGPT in the present and the future.
My first postcollege job (circa 1992) was working as a paralegal for a large law firm in Chicago as part of a team of people on a massive piece of commercial litigation that is more complicated than is worth explaining for the purposes of this anecdote, but which involved millions of documents originating from dozens of different geographic locations.
Some of the documents dated to the literal 19th century, though the most dispositive ones were from the first half of the 20th century. We had photocopies of typed documents that were originally carbons or mimeographs, many of which also had handwritten notes, notes that may have been dispositive to the case.
Every document was labeled with a Bates stamp, a two-letter code indicating place of origin, and then a number to provide a unique identifier.
I came on in the midst of the discovery phase, when depositions of key figures were happening, which included people who had worked in management and line supervisor positions at a number of industrial sites. Along with a couple of other paralegals, my job was to index the documents by laying hands on them, looking them over and then summarizing their contents in a DOS-based indexing system, noting the names of any people and, of course, including the Bates numbering so it could be found again.
So, when someone was going to be deposed, we’d go to the old database, type in their name, get a report of all the documents on which they’d been indexed, make copies of the copies and provide them to the attorney working on the deposition.
It was exceedingly boring work for which I was rather well paid. It was clear we weren’t even going to come close to indexing all of the documents, so we would constantly be scrambling. If so-and-so was being deposed from such-and-such location, we’d dive into the smaller universe of documents from that location and tear through looking for anything that could be relevant, sometimes burying the attorneys in documents, many of which were redundant, which would make them very irritated.
After I’d been there four or five months, a smart young associate who had been buried in documents asked the team of paralegals if we had any idea what the case was about, and after answering in the negative, he took a couple hours one afternoon to sketch out the issues and the known facts.
This previously missing context was invaluable, as I suddenly could discern which documents might be relevant from those which could be quickly passed over with a single database entry covering many hundreds or thousands of documents at a time.
It also allowed us to hire an army of temps to index a subset of documents much more quickly than the smaller team could handle. The full-timers on the case could identify the useful stuff, and the temps could crank through it.
I could not argue that we were 100 percent successful in identifying the relevant documents, but collectively we humans working with the tool of the database were able to make something that was genuinely useful to the aims of the lawyers working on the case.
After humming along for a while, after I’d given notice because I was going back to school (early to mid-1994), but before I’d actually left, a team from a large consultancy showed up to pitch something called “optical character recognition” technology.
They did an amazing job of demonstrating how OCR could scan data and create a searchable database that would be tagged to the digital images of each document. They quoted some figure to scan all the documents that sounded enormous to me, but it was much cheaper than the human labor involved in indexing and then copying any documents once you needed them.
As they say, I had no dog in this fight, but I briefly spoke up and suggested that the nature of the documents in our case—photocopies of carbons—and the fact that many also contained handwriting made them poor candidates for capture by this technology. I was heard respectfully, but the people who get to decide these things decided to go forward with the scanning to make the database.
After having been gone for a year, I reconnected with some old co-workers to see how things were going, and the results were predictable. The database did not work well, and the worst part of it was because no one had laid an eye on these documents prescanning to know what was in them, you had no idea what you might be missing.
They told me tales of lawyers being ambushed by opposing counsel who had documents we had not found and therefore could not prepare for. OCR has obviously vastly improved over the last 30 years, but the improvement is actually quite recent, and it is by no means perfect.
I don’t know how the story turned out beyond that point, though for all I know, the litigation is still ongoing. It was a really big case.
But I did learn a lesson about automation and the role human judgment must play in complicated scenarios.
I can definitely understand the allure of believing some significant portion of the human labor of writing could be outsourced to an algorithm, but I think people who believe that the vast majority of written communication can be outsourced are probably courting some real future headaches.
The worst part is that the people who give so much over to automation may have no idea that something is going wrong until it is far too late to do something about it, like the poor folks who trusted Tesla’s “self-driving” mode and got a wrecked car (or worse) for their trust.
ChatGPT is a potentially powerful tool, but I hope everyone is remembering that human agency, for all its flaws, is a necessary ingredient in sorting through these complex situations.
 I did the job for two years and made more those years in salary than I ever received as a full-time college instructor. I saved enough to cover my expenses above my grad assistantship for three years of graduate school with only summer employment to augment my earnings.