Profile picture for user jon.schubin
Posted By

There’s a new form of paranoia – AI content hysteria.

Large language models are spitting out millions of words an hour. Too many of these words have escaped from browsers and into the internet at large, entering a world without clear frameworks for when it is and is not allowed. 

Students, long subject to “plagiarism checks” on their work, now see essays put through tools with names like Quillbot, Copyleaks and GPTZero. Paste in a text block and receive an assuringly precise response: “There is a 94 percent chance that this text was not generated using AI.” 

Other potential shibboleths are emerging: Y Combinator founder Paul Graham recently Tweeted that he’s found evidence the word “delve” is correlated to the use of large language models. 

In the absence of a full understanding of how OpenAI and Google are generating texts, these tools' supposed analysis is nothing more than slightly educated guesswork. A piece in the International Journal for Education Integrity found “they are neither accurate nor reliable,” finding false positives and negatives more than 20 percent of the time. 

Unfortunately, the goal of content isn’t proving that it’s written beyond a reasonable doubt by a human – if someone has gone through the effort of running your copy through detection software, there’s a problem with it. In some cases, “this sounds like it was written like AI” is just a new-fangled, euphemistic way of saying “this sucks”.

If Renata Adler or Joan Didion used the word “delve” in a collection of prose, no one would think it was written by AI. “Delve” would simply be one small tree in a great forest of clever wordplay and sharp observation. 

We have continuously evolving subconscious detectors for bullshit and determining whether something was written by a human is the next frontier.

You can drive yourself crazy on what constitutes a true ‘performance-enhancing substance’ in content creation. I can’t imagine anyone would be bothered to find out a spell check has been enabled inside word processing software or even the green grammar alerts. Recently I’ve started to use the corporate version of Grammarly, which provides several suggestions based on usage, style and tone. 

I struggle with these tools. On the one hand, I want to believe my fundamentals of language are rock solid, beaten into shape by years as a copy editor and later daily news reporter. And yet I’m fond of word omissions and overly baroque construction that clearly hampers understanding. An editor, either human or machine, clearly has a function. 

What to do? I propose a simple solution is just to let your inner freak flag fly. Be bold, be fun and be interesting. 

Think back through this piece. Note the fun little bit of alliteration in the previous sentence. Also, you might have seen the unedited curse several paragraphs above, which has not appeared in the more than 200,000 words written over the last eight years on this site. There was also a reference to the mostly forgotten (and very New York) mid-century writer Renata Adler. 

All of this could have been easily removed. Wouldn’t it be easier – or safer – just to say bad or junk? I face challenges to this in the editing process. 

My advice stands in the face of many seminal texts on the issue. In “On Writing Well” by Wiliam Zissner. “The secret of good writing is to strip every sentence to its cleanest components,” he writes. Paul Graham, a noted speaker, says, “I try to write using ordinary words and simple sentences…The less energy [the reader] expend on your prose, the more they'll have left for your ideas.”

There’s a term for this: “purpose prose”. It’s excessively ornate writing that draws attention for its extravagant style instead of for its meaning. Many freshmen in composition seminars will find these words written below their first ever C+ on an essay. 

But interesting writing doesn’t have to be bad writing. As a rule, we will follow the rules. Simple, declarative sentences. Minimal use of passive voice. Fewer euphemisms, less cliché. And yes, I’ve removed bits that were misshapen and inhibited understanding. Editing still matters. 

When we do break the mould, however, it will be to show our unique perspective. It should allow us to be more precise, to paint a more colorful picture. 

On balance, I will be a bit looser going forward on allowing voice and tone to shine through, even in pieces intended for a corporate audience. Perhaps one of the worst things about artificially generated copy is its lack of friction. Challenging and surprising readers, forcing them to slow down, is critical to understand.

In editing this very issue of the newsletter, I found several places where my colleagues added a bit of zhuzh in their copy. Props to Charlie Andsell’s “​​pachyderm in the palace” and “acolytes hewn in [his] image.”

And in this time where we are still acclimating to machine-assisted writing, perhaps it’s okay to go a bit further than normal, a sort of counter-reformation against ChatGPT. 

Ironically, endlessly generated copy is known as “sludge,” since the prose emanating from them endlessly and effortlessly flows in and out of our mind’s eye and then down the drain. So perhaps it’s time to reclaim that word. We can prove our humanity with occasional moments of viscosity, and clarion calls to our individual voice. 

Jon Schubin is a director in London

(Postscript: As a bonus feature, I used our corporate version of Grammarly to create a version of this article that is “more clear” and “more on brand”. You can see that version – and everything that was lost in the process here. That version is 280 words long and ends with the following conclusion: “So let's not be afraid to let our voices and personalities shine through in our writing!” Oof.)