Using GenAI to sound like literally anyone but you.
Everyone knows the trick. You paste three of your old blog posts into an LLM and say "write like me." Thirty seconds later you get an article that sounds like you on a good day. Fewer typos. Better transitions. Somehow funnier. It's a little insulting, honestly.
But my friend Morgan Creighton had a saying "The opposite of a Good Idea is Often Also A Good Idea".
What if you need to write something that sounds like nobody? Not "neutral corporate voice." Nobody. Text with no fingerprints. Text that could have been written by anyone in the building, anyone on the planet, or a reasonably literate golden retriever.
When would you need that? Two scenarios come to mind:
- The annual anonymous employee engagement survey.
- A ransom note.
I want to be clear that I only have experience with one of these. I also want to be clear that the survey is the one where I'm worried about getting caught.
The "Anonymous" Survey Problem
HR says the survey is anonymous. This is technically true. Your name is not attached to it.
Then you hit the free-text box. "Anything else you'd like to share?" Of course you have something to share. You have been saving it up since Q2.
Here's the thing. Your manager reads those answers. Your manager has also read roughly four hundred of your Slack messages. You type "our sprint planning lacks a coherent theory of mind" and you might as well have signed it in blood. Notarized.
Everyone has tells. Mine are well documented. I write short sentences. I make a joke and then explain the joke. I say "Here's the thing." (See above. I literally can't stop.) I refuse to use em-dashes on principle. And somewhere around 2009 I lost the ability to write a paragraph without mentioning that I used to program in assembly. I used to program in assembly.
There's a whole field for this. It's called stylometry, the statistical analysis of writing style. Scholars used it in the 1960s to work out which of the disputed Federalist Papers Madison wrote. In 2013 it helped unmask J.K. Rowling as "Robert Galbraith." If a couple of linguists with word-frequency tables can catch one of the most famous authors alive, Karen from People Ops can absolutely catch you.
Reverse Style Transfer
The normal prompt goes like this:
Here are samples of my writing. Write like me.
The new prompt goes like this:
Here are samples of my writing. Write like literally anyone else.
Here's the version I actually used:
Here are five samples of my writing. Study them carefully.
Identify my verbal tics, sentence rhythm, favorite words,
punctuation habits, and the jokes I think are funnier
than they are.
Now rewrite the survey response below so that none of
those features survive. It should read like it was
written by a mid-level employee in a different department
who has never met me and has strong opinions about the
parking garage.
My original survey answer:
Our deploy process has more manual gates than a medieval castle. I've counted. It's eleven. One of them is Dave.
The anonymized version:
I believe there may be opportunities to streamline our deployment workflow, as it currently involves several manual approval steps. Also, the lighting on level 3 of the parking garage has been out since July.
Perfect. Nobody will ever suspect me. I don't even park on level 3.
Things That Went Wrong First
It took a few tries. Science is iterative. So is crime. (Survey crime. Theoretical survey crime.)
Attempt 1: "Make it sound like someone else." The model wrote it as Shakespeare. "Forsooth, the deploy pipeline doth tarry." Technically not my voice. But there is exactly one person in the company who would think that was funny enough to submit. Me. Traceable.
Attempt 2: "Remove all personality." I got text so bland it was suspicious. No human writes like that. Except Legal. Legal is now under investigation, and I feel a little bad about it.
Attempt 3: "Write as a completely different person." The model invented a persona named Gary. Gary had a backstory, a beagle, and a grievance about the coffee machine on four. Gary had more personality than I do. HR reached out to Gary for a follow-up conversation. There is no Gary.
The lesson: you are not aiming for different. Different is memorable. You are aiming for median. The statistical center of all office writing. The beige of prose. The sentence equivalent of a navy fleece vest.
The Ransom Note Use Case (Purely Theoretical)
The classic ransom note was cut out of magazines. Every letter a different font, so nobody could trace your handwriting. That was analog style obfuscation, and honestly it was ahead of its time. Labor intensive, though. And you needed magazines. Who has magazines?
So, in theory, GenAI is cut-out letters for the 21st century. In practice there are three problems.
Problem one: it won't do it. I asked (theoretically). It politely declined and gently suggested I talk to someone. Fair.
Problem two: the logs. Your prompt history is the least anonymous thing you own. It has your account, your timestamps, and every embarrassing question you've asked about sourdough. (So do all this on your personal GenAI account on your personal laptop).
Problem three, the big one: the AI has a voice too. Hide your style behind an LLM and you don't get no style. You get its style. Your ransom note comes out like this:
_Ransom Terms: A Comprehensive Overview
Great question! We have your cat. Here's what you need to know:
• Payment: $1,000,000 in unmarked bills
• Timeline: 48 hours
• Location: The bench by the duck pond
It's worth noting that the cat is safe and, frankly, thriving. Let's delve into next steps together.
I hope this helps! Let me know if you'd like a version with a more urgent tone._
You have successfully replaced "I know who wrote that" with "I know what wrote that." Congratulations. You are now the most identifiable kidnapper in history.
(Please do not write ransom notes. Not with AI. Not with magazines. The cat has done nothing to you.)
A Practical Recipe (For Surveys Only, Thank You)
Okay. Here's what actually worked.
Ask for your tics first. Feed the model your samples and ask it to list your habits before it rewrites anything. The list is horrifying. It is also the most useful writing feedback you will get all year.
Ban both voices. Tell it to avoid your tics and its own. No "delve." No "it's worth noting." No bullet points where a sentence would do. No cheerful sign-off.
Aim for the median. Ask for the reading level and tone of a typical employee at your company. Not a poet. Not a robot. Somebody who says "circle back" without irony.
Scrub the content, not just the style. Style isn't your only fingerprint. If you're the only person who counted the eleven manual gates, it doesn't matter how beige the sentence is. Specifics leak.
Read it out loud. If it sounds like you, try again. If it sounds like a press release, try again. If it sounds like a coworker you mildly dislike, ship it.
Conclusion
For years we've used AI to sound more like ourselves. It turns out it's just as good at making us sound like nobody at all. There's something philosophically unsettling in that. I studied Wittgenstein in college, so I'm contractually obligated to mention him once per article. He said the limits of my language mean the limits of my world. He did not say what happens when you outsource the language. I assume it's the parking garage.
Anyway, I filled out the survey. Fully anonymized. Median voice. No Gary.
The next morning my manager stopped by my desk. "Thanks for the feedback about the parking garage," she said.
I asked how she knew it was me.
"You're the only one who filled out the survey."
Stylometry is impressive. A sample size of one is undefeated.
Top comments (0)