The Language We Inherit From Machines

I’ve been thinking about AI watermarks again. Not the invisible kind, but the kind that is much more obvious.

A peer and I were talking about people we’ve known for years who have started talking differently. Not badly, but in a way that removed some authority from what they were saying. Their vocabulary has changed - it’s hard to tell where one’s vocabulary ends and the machine’s begins. We’re hearing things like “layers,” “failure modes,” “dimensions,” and “frameworks” appear in conversations where those words wouldn’t have been used, at least in the same context, just a couple of years ago.

An IT person saying “Layer 7” or how much they miss Taco Bell’s 7-layer Burrito is typical, but when someone says, “We’re experiencing a failure mode at the conversation layer,” something else is going on. The words have escaped their technical context. ChatGPT has turned us into Shrek.

Your Idea, Their Language

There’s no questioning where their knowledge came from - experience - but rather where the language used to express it originated instead. The knowledge is theirs; the language may not be. When someone tells me something, and I think, “That sounds like ChatGPT,” the focus on the topic at hand becomes secondary - at least for a couple of seconds. While the person I’m speaking with may have had the idea themselves, even spending years developing the thought, they likely found a new way to articulate it into something easy to understand.

For most of my career, professional language spread through people. Someone came up with a term because they needed to describe something, and others found it useful. They adopted the term, and it became used by more and more people over time. The term had a lineage, and now there’s a new path:

Person → LLM → Person → LLM → Person

The LLM doesn’t have to invent anything for this to happen. It can take a term that makes perfect sense in one context, and help generalize it into another. They simply sound how professionals talk, and their output gets adopted because you know it’s good (just like Brawndo is what plants crave, dummy.)

Cheeseburgers, Onion Rings, and Coleslaw

I caused something like this to happen in a workplace years ago, and in the dumbest way possible - which makes it a great memory to throw in here.

A coworker at NFTE was having a bad day, so I paid them a visit. Sometimes, a short trip down a hallway with a couple of cups of coffee and a bunch of cookies can make a difference in someone’s day; I just didn’t expect what happened when I sat down. They exploded with frustration, expletives of all sorts being yelled behind closed doors, and while I don’t remember the situation, their reaction made sense at the time.

The yelling, however, was something I knew I needed to help stop. In my own way, I managed to form a vocabulary - starting from that day - that represented different feelings.

“Cheeseburger” meant anger, “onion rings” was a thank you or a concession to something going on, and “coleslaw” was the equivalent of “my bad.” It was random eough to cause laughter, and it stuck for years. I even witnessed a friend who worked in a nearby building yell “cheeseburger” at his boss. The phrase was beginning to be adopted simply by hearing it used in conversation.

One day, the three of us had lunch at a local Japanese restaurant where I was a regular. I had my 25th birthday there and the restaurant picked up the tab. We were sitting at our normal table when one of the sushi chefs came over, placed a platter down, firmly exclaimed, “cheeseburger!” and walked away with a smile on his face.

The platter had three mini burger patties and fries. The crew in the kitchen was making lunch for themselves and thought it would be a funny surprise - and it was, but was also a sign that our vocabulary had broken out of the context in which it was designed. From that point on, “cheeseburger” became whatever it needed to be - the inflection in one’s voice became the real indicator of how you were feeling.

We’re Getting an AI Dialect

Language evolves. This isn’t the first time industries have developed jargon, and it won’t be the last. What we’re experiencing now is due to LLMs being extraordinarily efficient at spreading language. Someone encounters a phrase generated by an LLM, adopts it, someone else encounters it, and eventually the phrase becomes familiar enough to feel natural.

The usage of new language becomes self-propagating, and that gets my mind to wander from watermarking to computer worms. Words aren’t malicious, but they do replicate - and each new host can introduce small mutations while preserving enough of the original pattern for the language to continue spreading.

The Developer Problem

This isn’t limited to spoken languages - it carries over to skill sets, too. Over the past couple of months, I’ve seen developers say things like

“I haven’t written a line of code in months and I’m realizing how much I’ve forgotten.”

That statement carries the same concern, but for a different reason. The problem isn’t that they used AI or had forgotten a language’s syntax - nobody should feel bad about that. After all, that’s why documentation exists. The question is what else they’ve stopped practicing along the way.

Writing software isn’t just about producing code and shipping features; it’s breaking down a problem, making assumptions while building an approach, understanding potential tradeoffs, recognizing when something doesn’t make sense, debugging, and knowing why something failed.

It’s a process, and one that dominates my field of vision every 3-5 years since I wrote my first line of BASIC.

“Cheeseburger.”

If AI handles more and more of the process described above, the output can improve while the human’s ability to produce or evaluate that output deteriorates. You can become better at getting software built while becoming worse at programming, and that’s an unfortunate tradeoff.

Back at NFTE, around 2004, we began development of a new version of BizTech, an online learning platform. We got a huge Microsoft hookup around software licensing, and they wanted us to use .NET, replacing the Classic ASP used in previous versions. Great idea, except the selected development team were Java experts - and approached everything like it was Java in disguise.

In the early 2000s, .NET and C# were explicitly designed to feel familiar to Java developers. When a team treats a web framework like a monolithic desktop application without understanding managed memory or stateless web architecture, the result can be massive scaling bottlenecks - and the guy who built the infrastructure gets blamed.

“The servers are bad. There’s a security hole and someone’s exhausting resources or doing something malicious.”

That’s a big cheeseburger moment - a double cheeseburger, even. I snagged the President’s credit card, bought a Sourcefire box (❤Snort) and proved the servers weren’t choking on network load or external attacks - the code was simply collapsing under its own weight.

“Don’t ever fucking question my infrastructure ever again.”

I remember leaning into a Polycom Bird of Prey conference phone and saying that in the most calm voice ever. The kind of calm that scares people. That’s a lot of aggression from me in the workplace, and I don’t think I’ve spoken that way since. There was a silver lining to all of this; the team had no more excuses, put the work in, and BizTech 2.0 became a 2005 CODiE Awards finalist in a secondary education category. Kaplan won that year and it was very much deserved.

“Onion rings.”

It’s similar to the problem I see with spoken language, and even some AI-generated code. The danger isn’t necessarily that the machine gives us bad answers, but that it provides them so conveniently, and confidently, that we stop exercising the parts of ourselves that used to produce them.

Just as an LLM can influence spoken language, it can give us a false sense of confidence, similar to the developers that couldn’t fathom that their code was at fault until they saw data that proved otherwise.

This isn’t an Argument Against AI

I use AI tools nearly every day. I used OpenCode to help me solve issues with this blog’s theme. Likewise, I’m not interested in pretending we can - or should - put the genie back in the bottle. The machines are incredibly useful, but there’s a line between using a tool to amplify your thinking, and using a tool in place of it.

Critical thinking isn’t just the ability to produce an answer. It’s the ability to ask, “why do I believe this,” “does this actually make sense,” “would I have said it this way,” and “do I understand it well enough to explain it without the machine?”

That’s a question worth asking whether you’re writing marketing copy, troubleshooting a network, building software, or simply trying to explain an idea to another human being. AI is becoming better at helping us think, and it’s critical that we make sure they’re not quietly becoming responsible for all our thinking.

Use the machines; just don’t let them convince you that you have nothing worth saying without them.