Artificial Intelligence: Facts, Fictions, Myths & Legends — Part 2

The Words You Choose Are the House You Build

I promised you at the end of Part 1 that we’d get into prompting, and here we are. Before I get into the meat of it, though, I have to tell you about the night I almost lost my masthead.

Not “lost” as in misplaced it in a drawer somewhere. Lost as in one imprecise sentence, typed at eleven o’clock at night into a chat window, came about four words away from wiping out a piece of branding I’d spent weeks getting right on a project I’ve been building called MyFurFace.com — a pet-focused social media platform I’m developing basically solo, with AI as my main technical collaborator since I don’t have a formal coding background, and AI is very good at knowing where to put all of those pesky commas ‘,’ parenthesis ‘()’ and tildes ‘~’ which I’ve been trying to figure out for the better part of four decades why that key exists on every QWERTY keyboard in existence. Meanwhile, we have to use a series of coded keystrokes to type out the copyright symbol ‘©’ which, in Windows, you do by holding down the Alt key while you type out 0169. I don’t know about your life, but in mine, I have to use that copyright symbol a lot more than I will ever need to use a ‘~’.

More on the MyFurFace.com project in a bit. It’s going to be doing a lot of the demonstrative heavy lifting in this article because it was, and is, my real-world coding lab rat.

Here’s the sentence, near as I can remember it: “Clean up the header.” That’s it. Four words. Innocent enough on the surface, right? Except “clean up” is one of those phrases that lives a double life in the English language, and, as I’ve come to find out, has many other possible iterations in the mind of a supercomputer.

Because I hadn’t given the AI a single clue about which version I meant, it decided to make its own best guess. That’s right, AI computers often guess when they don’t have complete information. Who knew? I sure didn’t. But I learned the hard way.

Did I mean “tidy up the code so it’s not a mess,” or did I mean “remove things from it until it’s clean?” By the way, what exactly does ‘clean’ look like to AI? As I alluded to earlier, each of those is entirely different instructions wearing the same shirt, and the AI had to guess which one I was actually asking for. In this case, it guessed the second one. It started stripping elements out of the header, including, very nearly, the masthead image I had crafted in my spare time across about two weeks of generative coding and explicitly never wanted touched.

Fortunately, I caught it before any real damage was done, though I only caught it because Claude, Anthropic’s proprietary chatbot, luckily decided to stop halfway through and ask me a question about what I wanted to do with the ‘dirty’ code it was removing.

See? There is a digital god too. At least that’s what I choose to believe now.

It was scary as hell, but that near-miss taught me something I want to spend the rest of this article unpacking with you at a level of detail most people never bother to think about: AI doesn’t read your mind. It reads your words. And your words, whether you realize it or not, are doing double, triple, sometimes quadruple duty every single time you put your fingers on a keyboard.

 

Two Roads, Same Destination, Wildly Different Trips

Let’s talk about what “getting it right” actually looks like, because I want you to see the contrast before we dissect the wreckage.

A few weeks back, I needed the database backbone built for MyFurFace.com. Specifically, it would be comprised of nineteen tables covering everything from pet profiles to health records to a matching engine to marketplace transactions. I assure you, this new website of mine has functionality even Mother Zucker over at Meta would be envious of. My point being, this is not a small task, and it’s almost unheard of to try a project of this magnitude completely alone, which is why I decided early on I would use AI to check my work and test all of the functionality of my coding, particularly when I haven’t slept for 2 days.

This isn’t just coding; it’s architecture. I could have just said “build me a database for a pet social network” and walked away to make coffee. Claude is that good at coding. However, the result would have been generic. Emotionless. Stale.

Why? Because for all of its attributes, AI lacks imagination. It can be creative to an extent, but that creativity is always called upon from other sources and you can guarantee that if you have an ethereal vision in your head about how you want your project to look, I can guarantee you with 100% certainty that an AI’s attempts to manifest that vision into a reality for you will not look anything like your idea.

There is no doubt that had I just employed AI to write all the code for the app for me, I would have come back to something generic, something that technically qualified as “a database” the same way a lean-to or a yurt technically qualifies as a house.

Instead, I got specific to the point of being almost obsessive about it. I specified that pets, not their owners, needed to be the primary entity — because the entire premise of the platform is the opposite of every other social media website in existence — that pets have the primary profiles and owners are the secondary. Because that basic premise is backwards from just about every other social platform on Earth, it’s therefore not something any AI would assume on its own.

I specified data types down to the field level. I specified that a pet’s ID needed to be a UUID stored as a specific character length because I had future scaling concerns. I specified that an owner could be linked to multiple pets through a particular structure rather than a traditional join table, because of how I wanted the app to query that relationship later. I told it, explicitly, what the password field needed to be called and why “password” itself was a bad name for it — because it’s not a password, it’s a hash of one, and calling it the wrong thing would eventually get some future version of myself, or someone I hire, confused at 2 a.m. during a security audit.

The result was nineteen tables that came out clean the first time, no rework, no “wait, that’s not what I meant.” That’s not luck. That’s the direct, measurable, dollar-and-hours output of specificity. Every ambiguity I closed off in each of the prompts was one more ambiguity the AI didn’t have to guess or interpret on its own, and every guess an AI doesn’t have to make is a guess it can’t get wrong.

Now compare that to “clean up the header.”

Same tool. Same session, albeit some 18 hours apart. Same guy typing. Wildly different outcomes, and the only variable that changed was how much interpretive work I left on the table for the machine to do on its own.

 

Which Brings Me to the Part Where I Explain How the Sausage Actually Gets Made

I want to go deep here, because I think this is the single most misunderstood mechanic in the entire AI conversation, and almost nobody explains it properly; so bear with me.

When you hand an AI a word, especially a word with more than one meaning, it doesn’t just flip a virtual coin. It’s not randomly guessing between definitions the way you might flip through a dictionary blindfolded and jab your finger at a page. Believe it or not, there really is some logic behind the choices it will eventually make.

What it’s actually doing is weighing every surrounding piece of context you’ve given it, both in that conversation and often in the documents or history attached to it, and using that context to calculate which meaning is statistically most likely to be the one you intended, given everything else you’ve said. That’s right. Despite all that fancy AI hoopla, even the most advanced chatbots ultimately reduce everything down to 0s and 1s, just like an old Commodore 64 used to do back in the early 1980’s.

Yes, before you go there in the comments, I had one.

Now, think about the word “bank.” If I say “I need to deposit this check at the bank,” the surrounding words — deposit, check — pull hard toward the financial institution. If I say “the fisherman sat on the bank,” the surrounding words — fisherman, sat — pull hard toward the edge of a river. The AI isn’t magically “knowing” which one you mean. It’s doing something closer to triangulation. It’s looking at the neighbors of the word and asking, essentially, “given every other word this word is surrounded by, which version of itself is it most likely to be right now?”

Here’s the part that should make your hair stand up a little: “clean up” has neighbors too, but on that particular night, my neighbors were useless. “Clean up the header” had no deposit, no fisherman, no check, no river. It was just sitting there in isolation, and when a phrase like that doesn’t have enough context around it, the AI still has to pick something. It doesn’t get to throw up its hands and say “insufficient data, please clarify” unless it’s been built or prompted to prioritize that kind of caution — and even then, plenty of ambiguous requests slide right through because they don’t look ambiguous to the system in the moment. They look like a normal, executable instruction with a most-probable interpretation attached to it.

The AI committed to the interpretation that statistically wins in most codebases when someone says “clean up” attached to a UI element — and in most codebases, most of the time, “clean up” does mean strip out clutter. I just wasn’t in “most codebases.” I was in mine, with my own private, non-negotiable rule sitting in the back of my head that the AI had no way of seeing unless I put it in front of it.

That’s the whole ballgame, right there. The AI wasn’t wrong in some abstract sense. It was wrong for me, in my specific context, because I hadn’t given it my specific context. It answered the question I asked instead of the question I meant, and those are not always, or even usually, the same question, and they almost always have very different answers.

 

The Genie Doesn’t Read Your Mind, It Reads Your Wish

I brought up genies in Part 1 when I talked about AI having a catch built into every wish it grants, just like a mythical genie in a lamp always does, and I want to pull that thread again here because it fits this topic almost too perfectly.

Every genie story ever told has some version of the guy who wishes for a million dollars and gets it, technically, in the form of a life insurance payout after a family member dies. The key lesson to be learned in those stories, of course, is that genies aren’t being malicious. The genie granted exactly what was asked. The tragedy is entirely, one hundred percent, a failure of the wisher to specify what they actually wanted clearly enough to close off the interpretations they didn’t want.

AI operates exactly the same way, minus the trickster intent, minus even the awareness that it’s doing anything remotely genie-like. It’s simply resolving your words down to their most probable meaning and executing against that meaning. When your words carry only one reasonable interpretation, you get the wish you actually wanted. When your words carry two or three reasonable interpretations, and you don’t bother to specify which one you meant, you’re rolling dice, and sooner or later those dice are going to come up on the meaning you didn’t intend, usually at the exact moment you can least afford it.

On MyFurFace.com, my nineteen tables were a wish with no ambiguity left in it. “Clean up the header” was a wish with a trapdoor built into it, and I nearly fell through my own trapdoor.

 

The Fix Isn’t Complicated, It’s Just Effortful

I’m not going to pretend there’s some magic incantation that eliminates ambiguity forever, because there isn’t. Language is inherently a lousy compression format for thought — a situation where we’re taking something rich and complicated happening in our heads and squeezing it down into a string of words. I probably don’t have to tell you that some part of that rich and complicated part in our heads always gets lost during that squeeze. That much is unavoidable. What you can do, though, is squeeze more carefully.

When prompting, the key is to always clearly state your intent, not just your instruction. Don’t just say “clean up the header.” Say “simplify the header code without removing the masthead image or any visible elements; I only want the code itself tidied.”

That sentence has almost nowhere left to hide. Give the AI the “why,” not just the “what,” wherever you can, because the “why” acts like the fisherman standing next to the word “bank” — it drags the interpretation toward the meaning you actually hold in your head. And when the stakes are high, when you’re touching something you genuinely cannot afford to lose, you need to spell out explicitly what must never happen, not just what should happen, and this is what most people who try prompting AI without any actual training in prompting do every day.

I had already told the AI, in an earlier session, never to remove or alter my masthead image. That standing instruction is the only reason “clean up the header” didn’t turn into a genuine catastrophe instead of a genuine scare — because even though that one prompt was ambiguous, I had closed off the worst possible interpretation ahead of time, in a separate breath, on a separate night.

That’s the real lesson buried in all of this, and it’s the one I want you to walk away with: precision isn’t about being paranoid or over-engineering every sentence you type. It’s about recognizing, before you hit enter, which of your words are carrying more than one suitcase, and deciding, on purpose, which suitcase you actually want opened.

 

Lastly on This Point

Somebody once said the single biggest problem with communication is the illusion that it’s taken place. That line most often gets credited to George Bernard Shaw, though that attribution could never actually be proven to belong to him. Still, regardless of whoever actually said it, the point stands on its own: you can walk away from a conversation with an AI feeling completely confident that you were understood, and be dead wrong about it, simply because the words you chose had more doors in them than you realized.

My header survived. My masthead survived. But I only got to write that sentence because I’d built a guardrail into an earlier conversation, not because the prompt itself was airtight. Not every ambiguous prompt is going to have a guardrail waiting to catch it, and that’s exactly why learning to spot your own ambiguity, before the AI has to resolve it for you, is the single most valuable skill you need to develop to prompt fearlessly.

Before I end this second part, I need to remind you also of the evil-intentioned people we talked about in part 1. Because they’re out there.  And while today’s AI has numerous fail-safes and guardrails incorporated into it, people also keep getting smarter about how to effectively sidestep those guardrails.

My point here is simple: there is always a chance that an evil-intentioned person or persons will find ways to manipulate prompts in such a way that they will get AI to do things that could cause real harm. Everything we do to work with AI is done through prompts, and while most of them are completely benign, there have been instances where wicked people have found ways to craft prompts that get past the security.

I’m not going to mention those instances here, lest I give some depraved imbecile some digital ammunition to use against some unsuspecting innocent. Most of those instances are easy enough to find through the simplest of searches. But as I mentioned more extensively in part 1 of this series, the potential for real harm absolutely exists, and we all must remain extremely vigilant in gatekeeping the technology if and when it looks to us like someone is trying to use it for wicked or dishonest purposes.

Only a fool would believe that the raw potential for evil should be a good enough reason to scrap such advanced technology forever. That’s never going to happen. No, AI is here to stay, and its role in our daily lives will only increase with each passing day. Because of that fact, the responsibility to gatekeep falls upon us all.

Will it stop every instance of evil that unscrupulous people will find to use AI for?  Not even close. Bad things are going to happen. Count on it. But while we can’t stop it, we absolutely can minimize it.

Thanks for joining me again. I’m very glad you aren’t sick of me yet. In the final part of this series, we’re going to be talking about what happens on the other side of this coin. Not what AI does when your words are ambiguous, but what happens when your words are perfectly clear, and the AI still gets it wrong anyway, and I’m going to explain how to tell the difference between the two so you know which problem you’re actually working against. So please be my guest and join me one last time for Part 3, of Artificial Intelligence: Facts, Fictions, Myths & Legends.