Showing posts with label Primo Levi. Show all posts
Showing posts with label Primo Levi. Show all posts

Saturday, April 22, 2023

ChatGPT needs a Fact-Checker

I’ve had a long career as a production person, working with the printed word from my high school days, through freelancing on magazines, and then working full-time on books; and that’s the way I paid my bills for decades.
Years ago, I found my all-time favorite job—Design Director at an encyclopedia company—and that’s where I learned more about fact-checking and copyediting the printed word.

“Fact-checking” is the process of verifying the factual accuracy in a document. “Copyediting” is the process of rewriting a document—correcting grammar and misspellings, clarifying syntax, eliminating wordiness, etc.
(Different companies have different copyediting standards; therefore, different copyeditors never make the same corrections.)
“Proofreading” usually describes checking a document against another version of that document.
However, sometimes a proofreader may find errors, in a piece of writing, that neither the fact-checker, or the copyeditor, could see.

Ideally, information should go through a review process—involving all of these steps—because human beings are fallible and make errors.
We make errors in spelling, and errors in sentence structure.
We make factual errors because we can’t read our own notes, or because we’ve misunderstood the data that we’re referring to.
Sometimes, errors are discovered in the first review, and sometimes errors are discovered (or “caught”) after the info has been posted or printed.
Sometimes, new errors are inserted, when we try to correct the original error. I think people in the industry use the word “caught” because it’s always a “hunt” for errors, and often it’s mind-blowing when we didn’t “spot” an error that seems obvious later.
(Books have more time for this process than newspapers.)

Perhaps, you’ve noticed the “Correction” boxes in the newspapers.
It’s the policy of newspapers (for example, the New York Times) to correct factual errors in a prominent space.
These mistakes are usually errors like misspelling names, or giving incorrect information.
However, if a breaking news story (for example, about a cave-in or a plane crash) tells you that 56 people died, and 57 died instead, that information will be in the next story about the event, and it won’t be in the “Correction” box.
New data is not considered a “correction.”

Correcting errors in books are handled differently.
In the past, if a publisher noticed an error after a book was printed, they would create an “Errata slip” and either bind it in at the back of the book, or place it loosely under the inside front cover.
However, that’s seldom done these days.
Modern publishers—especially for “books of facts” like technical or clinical books—usually post errata lists on their websites.

All of the above information is preamble to this article on why I find it hard to take the ChatGPT experiment seriously.
How can ChatGPT be useful if the data fed into the software is not fact-checked?
As the saying goes: “Garbage in; garbage out.”
At least 300 billion pre-2022 words (570 GB) were fed into ChatGPT. If those billions of words were taken from error-filled Wikipedia articles, unknown electronic books, and conflicting essays scattered across the internet; ultimately how useful can the results be?
Students and journalists can’t count on the ChatGPT essays to be accurate!
(Are the high school teachers and copyeditors so ignorant that they won’t notice the factual errors?) 

The only good I can see in ChatGPT is that of helping writers to organize material.
That’s sometimes difficult to do.
However, the thrill of reorganizing an essay, and making it better, is most of the fun of writing.
You learn a lot by rewriting an essay.
Often, the initial thoughts are malformed, but slowly you learn how to express what’s on your mind. ChatGPT certainly isn’t valuable in helping with the most difficult part of writing non-fiction—making sure that the facts are as accurate as possible.
It may be somewhat helpful in helping the author to phrase the facts in the best way possible, but you lose the pleasure of discovery.

In the “good ol’ days”—when printed encyclopedias were found in homes and people read daily newspapers—accuracy used to be a big deal.
The encyclopedia set I worked on strove for accuracy, but occasionally there were hiccups.
I remember one “hiccup” when an article about a South American country was commissioned for the 24-volume set, and the author of that article was afraid that the slight rewrite and clarification of her article (by editors) would result in her assassination.
She demanded that the encyclopedia be pulled back and reprinted—obviously, not a possibility in those days of print encyclopedias.
I was on the “page and picture” side in the encyclopedia process, and not the “word” side.
Therefore, I don’t know how it all turned out. However, I do remember hearing that the author was livid.
No matter how one tries to be accurate in an encyclopedia set, or a newspaper article, there will always be disagreements between the authors and the editors.
It’s simple to prove that a word is misspelled, but deciding whether or not an event is described properly is complicated.

Another problem that the editors of printed encyclopedias used to face is which people and places to include in the sets.
I remember that the fan club leader for a certain country singer was adamant that our encyclopedia set include an article about their idol.
They wrote every year, begging the editors to include an article on the singer, describing his merits in great detail.
(Today, all superfans need to do is write their own articles, and place them on Wikipedia.)
There were also big issues with the maps—for example, any maps showing the border between India and China.
There are over 100 disputed territories on earth.
Which political sides will ChatGPT take on these territories? 

I can’t get over how casual the ChatGPT creators are over the fact that material generated by ChatGPT may be inaccurate or spun entirely from falsehoods.
It’s as if they believe that as long as the written words “sound correct,” all is well.
How can intellectual discourse survive, based on that philosophy?
I’ll close with a few lines from Primo Levi’s science-fiction story, “The Versifier”*—a story about a poet who purchases a computer to write poetry for him that I discussed in my inaugural article.
This story was first published in Italian in 1960 in the Italian newspaper Il Mondo, and is still relevant today.
    Il disio di seguire conoscenza,
    E miele delicato il suo succo acro.

    The desire to ingest vast knowledge,
    A nectar of sorts, but bitter to its taster.

The lines in Italian are from Storie Naturali by Primo Levi, and the English translation is from The Complete Works of Primo Levi, by Liveright Publishing Corporation, a Division of W.W. Norton & Company. (The “Versifier” story is part of the “Natural Histories” group of short stories in Volume One, pages 417-438.)


What You Liked Best