When we use large language models to augment, improve, or accelerate our writing, they erase our human voice in the process. Let’s not be erased.
The robots may be taking over, but no one seems to like how they talk.
In theory, the robots (by which I mean generative AI) talk just like us, given that their large language models (LLMs) have been trained on billions upon billions of statements you and I have made online. That Hacker News thread in which you waxed rhapsodic about JUST HOW WRONG someone is about Flash on the internet? It’s now training data for someone’s LLM-augmented doctoral dissertation. The LLMs have “learned” from all this online chatter to generate text that sounds like a human being.
Except it doesn’t. Not really. Not enough.
There’s a swelling chorus against AI-generated content on LinkedIn and elsewhere. As Oxide Computing CTO Bryan Cantrill opines, “Holy hell, the [AI] writing sucks.” Now, Cantrill is known for having strong opinions, but he’s not wrong when he argues this AI-generated writing is “stylistically grating.” The biggest tell? “Em-dashes that some of us use naturally—but most don’t (or shouldn’t).” OpenAI founder Sam Altman just fixed this last annoyance, but not before many of us realized that in our attempts to make our lives easier through AI, we inadvertently made everyone else’s lives worse.
It’s time to get back to writing that expresses ourselves, not merely what an LLM thinks sounds plausibly close to ourselves, because it’s the human in us that makes our communication compelling to other humans.
Cozying up to the robot voice
This trend toward robot voice isn’t new. If you’ve ever visited the UK or simply read a UK paper online, you’ll know that UK newspapers have distinctive voices. It’s not merely that different papers have different political biases and wear these biases proudly (or sanctimoniously, in the case of The Guardian) on their sleeves. Rather, they’re emphatically opinionated. In the US, we try to pretend we’re taking a neutral stance, even if the facts we choose to ignore or skew reveal our political biases quite clearly. As Emily Bell writes, “British journalism is faster, sloppier, wittier, less well-resourced and more venal, competitive, direct, and blunt than much of the US oeuvre.” (Yes, you know it’s an article from The Guardian when “oeuvre” is casually used as if normal people talk like that.)
For those in the tech world, the best example is perhaps The Register, founded in the UK. Perhaps best known for its sometimes over-the-top headlines, The Register offers smart, punchy analysis. It’s not neutral and doesn’t pretend to be. I used to write for The Register, and remember fondly some of these headlines:
- IBM trades cold comfort for hot air in Microsoft-AWS slugfest (a knowing nod to Pink Floyd)
- Craptastic analysis turns 2.8 zettabytes of Big Data into 2.8 ZB of FAIL (Doesn’t this make you want to click and read?)
- Larry Ellison’s yacht isn’t threatened by NoSQL – yet (Oops! Apologies to my present employer)
I didn’t write a single one of those headlines: The Register’s editors did them all. Sometimes they made me squeamish, but they were always fun (and sometimes informative!). No machine could replicate that formula, because there really wasn’t a formula. It was just some guy (Ashlee Vance at one point) trying to marry wit and information in an educated guess as to what would make someone want to read the article.
At the time I was writing for The Register, I was also writing for other publications. Those headlines were a bit tamer, which had much to do with the US ownership of the sites:
- With GitHub acquisition, Microsoft wants to make Azure the default cloud for developers (Written in 2018, this thesis on Microsoft’s GitHub intentions seems more accurate every day)
- NoSQL keeps rising, but relational databases still dominate big data (true then, true now)
- Don’t believe the hype, AGPL open source licensing is toxic and unpopular (one of the few headlines I actually wrote)
Although this is just a small sample, the US-style headlines (and writing) pretend to flatten out perspectives, aiming for an impossible neutrality. The UK doesn’t care; the writing is meant to provoke and inform. Our AI-generated content sounds more American than British, and that’s not a good thing.
Putting more ‘you’ in the writing
LLMs are “lousy writers and (most importantly!) they are not you,” Cantrill argues. That “you” is what persuades. We don’t read Steinbeck’s The Grapes of Wrath to find a robotic approximation of what desperation and hurt seem to be; we read it because we find ourselves in the writing.
No one needs to be Steinbeck to draft press releases, but if that press release sounds samesy and dull, does it really matter that you did it in 10 seconds with an LLM versus an hour on your own mental steam? A few years ago, a friend in product marketing told me that an LLM generated better sales collateral than the more junior product marketing professionals he’d hired. His verdict was that he would hire fewer people and rely on LLMs for that collateral, which only got a few dozen downloads anyway, from a sales force that numbered in the thousands. Problem solved, right?
Wrong. If few people are reading the collateral, it’s likely the collateral isn’t needed in the first place. Using LLMs to save money on creating worthless content doesn’t seem to be the correct conclusion. Ditto using LLMs to write press releases or other marketing content. I’ve said before that the average press release sounds like it was written by a computer (and not a particularly advanced computer), so it’s fine to say we should use LLMs to write such drivel. But isn’t it better to avoid the drivel in the first place? Good PR people think about content and its place in a wider context rather than just mindlessly putting out press releases.
LLMs are effective in software development because machines don’t mind reading drivel (boilerplate code). But people do. For any job that depends on persuasion—and most do, to some degree—it’s essential to maintain a human voice. For example, as much as I can’t stand Trump’s online (and offline) persona, part of the reason his followers find him so compelling is his stylistic weirdness. It feels authentic because it is authentic.
This isn’t a plea to be like Trump (please don’t!), but please be yourself. “AI made your writing smooth,” Talentz.ai CEO Muhammed Shaphy notes. “It erased your voice in the process.” Don’t be erased. Don’t be smooth. Be you.
