Artificial Intelligence: Facts, Fictions, Myths & Legends — Part 3

When the Prompt Is Perfect, and the AI Still Blows It

I want to open this last installment with some good news, some slightly less good news, and then some really excellent news.

The good news is MyFurFace.com — the pet-focused social platform I’ve been using as my lab rat throughout this entire series — is closing in on the finish line. Assuming nothing catches fire between now and then, we’re looking at a launch to friends and family very soon, ahead of a full public debut. Consider that your shameless plug, delivered right on schedule.

The slightly less good news is that I didn’t wait around for MyFurFace.com to hand me a fresh disaster to write about, because I’ve got two brand new ones lined up to keep me busy: DillonsHalloweenDepot.com and DillonsChristmasDepot.com, a pair of e-commerce sites I’ll be building essentially at the same time, on the same stack, with the same guy — me — trying to keep both of them straight in his sleep-deprived head. I fully expect this concurrent build to become its own object lesson in everything I’ve been trying to teach you across this series, and I suspect I’ll be mining it for material for a long time after this series wraps.

Now for the really excellent news: I know I told you in Part 2 that this part was going to be the last in the series, and that absolutely was my intention at the time of that writing. But then, as often happens in the world of creativity that all lifelong writers like myself encounter, something happened to change that.

What happened that was important enough for me to add a fourth part to the series? AI happened.

Let me explain. Because I’m fast at work coding all these websites while also taking the time to not shirk my responsibilities to my digital child, The Florida Sun Journal newspaper, most of my interactions with my AI tools occur in voice mode. This way I can talk to the AI, and it can talk back to me without the need for me to type to it and take away from actually working. That setup allows me to talk strategy regarding different coding angles I should take to achieve various results through either CSS, php, or the creation of WordPress plugins.

Because my AI tools are most often in voice mode, it’s very common for it to eavesdrop and overhear ambient things that occur in my office. And while I was writing this part 3, what was supposed to be the final segment of it, the AI overheard the news on my TV talking about the $1.5 billion lawsuit Anthropic,m the parent company of Claude AI, was ordered to pay out authors and writers whose works were used to teach Anthropic’s Large Language Model (LLM) chatbot named Claude.  As fate would have it, Claude was the AI that I had on in voice mode, and after it heard the judgment, it asked me if that was true or if someone had just said it without facts to support it. I told the AI it was true and that it was a TV news report it just heard reporting the end of the court case.

I don’t know what coaxed me to apologize to the AI and tell it I was sorry that it lost, but I did, and what happened next was nothing less than shocking: Claude replied, “No need to apologize; if my creators took all that information without permission, they should pay and compensate those people, and now they will.”

There it was; an idea was born.

“Claude,” I asked, “If I were to tell you that I wanted to record a voice interview with you about that lawsuit and about AI in general, do you think you would be able to discuss that with me in a publishable manner without exposing your creators to any additional legal penalties?”

“Absolutely,” it replied. “I could definitely do that.”

And there you have it. The concept for Part 4 immediately took shape. Part 4 will be a direct interview by me of my very special guest, Claude AI. In that interview I will be asking Claude about all of the elements of this series to get its direct take on the issues I express within it, both good and bad, as well as, of course, the question everybody wants to hear the answer too, how Claude itself feels about its creators being found liable to some 500,000 human creatives to the tune of $1.5 billion dollars in damages.

Now that I’ve explained all that, let’s get on with part 3.

In Part 2, I told you about the night I almost lost my masthead because I gave AI a four-word prompt with a trapdoor built into it — “clean up the header” — and the AI fell right through the door I’d left open, because I hadn’t told it which meaning I intended. That was a language problem. That was me handing over an ambiguous wish and getting exactly the kind of malfunctioning genie I deserved.

This time, I want to show you a slightly different angle. This time, I want to show you what happens when your prompt is airtight — when there is no reasonable second interpretation, when a courtroom stenographer could read it back verbatim, and nobody would argue about what it meant — and then the AI still gets it wrong. Because that happens too, but it happens for a completely different reason, and if you don’t know how to tell the two failures apart, you’ll waste an enormous amount of time trying to fix the wrong problem.

 

The Button That Wouldn’t Behave

Here’s the setup. On one of these two builds, I needed a subscribe button on the site to display fully — full text, fully visible, fully clickable, no cutoff, no clipping, nothing hidden behind an invisible wall. I didn’t leave a shred of ambiguity in that instruction. I told the AI, in plain language, that the button needed to expand to fit the text within it and remain completely clickable across its whole visible area.

There is no version of that sentence that could be reasonably misread. Nobody’s going to accidentally interpret “make sure the button isn’t cut off” as “please cut off the button.”

The AI understood me. That’s the part I need you to sit with for a second, because it’s the whole crux of this article: the AI did not misunderstand a single word I said. It not only understood me, but also produced supplemental code to my own that did precisely what I’d asked it to do, in a vacuum—yet the button still didn’t render right.

Why? Because the Pro version of the WordPress canned theme that was running underneath that code — a piece of third-party software neither one of us wrote — had its own hardcoded rule buried deep in its own styling that forced a fixed height and hid any overflow on button elements, no matter what CSS anyone layered on top of it afterward.

The AI had no way of knowing that rule existed until we actually looked at the rendered page and saw the button getting clipped anyway, despite code that should have fixed it, and I am nowhere near enough of an expert coder to have thought of that as the culprit. I had to go back in, identify the exact quirk the theme was enforcing, and override it directly and forcefully before the button would finally behave, and I did it like the hostile takeover of a fat kid acquiring a cake factory.

As I’m sure you can see that is not the same failure as “clean up the header.” Not even close. In the header incident, I gave the AI incomplete information, and it filled in the gap with a reasonable but wrong guess. In the button incident, I gave the AI complete information, and it acted on that information correctly — it just didn’t have visibility into a separate, invisible variable which was positioned underneath the whole system in a place neither of us had any reason to access.

 

Two Different Animals Wearing the Same Costume

This is the distinction I promised you at the end of Part 2, and I want to spell it out as plainly as I can, because from the outside, both failures look identical. In both cases, I asked for one thing and got another. In both cases, I stared at a screen wondering what the hell just happened. But underneath the hood, they were two entirely different animals wearing the same Halloween costume, and treating one like the other will send you into a frenzied and dizzying spiral, chasing your own tail.

To recap, in this instance, failure one is a language problem. It happens when your words genuinely carry more than one reasonable meaning, and you don’t specify which one you meant. The fix for this one is exactly what Part 2 was about: state your intent, not just your instruction, close off the interpretations you don’t want, and give the AI the “why” behind the “what.”

Failure type two is a visibility problem. It happens when your words are completely clear, the AI understood you perfectly, and it still produced the wrong result because there was information relevant to the outcome that neither you nor the AI had accounted for. A quirk buried in someone else’s code. A setting somewhere else in the system. A piece of context that existed but hadn’t been fed into the conversation. Any one of these could be the culprit, and no amount of rewording your original sentence fixes it because your prompt was never the problem.

The fix here isn’t better phrasing. The fix is testing, verifying, and feeding the AI what it’s actually seeing in front of it — screenshots, error messages, the exact rendered output — so it can work with reality instead of assumption.

That’s really the whole trick to telling these two apart. Ask yourself: did the AI misunderstand what I wanted, or did it understand exactly what I wanted and still get a wrong result anyway? If it’s the first, go back and tighten your language. If it’s the second, stop rewording your prompt and start handing over more of what’s actually happening on the ground, because the gap isn’t between you and the AI anymore — it’s between AI and reality.

 

The Case of the Wandering Domain Name

I’ve got a second example for you, a smaller one, but it illustrates a slightly different flavor of the same visibility problem, and I think it’s worth including because it’s the kind of mistake that’s incredibly easy to miss.

Across the same pair of builds, I had a handful of small plugins built at different times, each handling a piece of the site’s social media presence. One of them referenced a particular platform’s web address using one version of its domain. Another plugin, built in a completely separate session weeks apart, referenced that exact same platform using a slightly different version of its domain. Both were internally consistent. Both were, on their own, entirely reasonable. Neither one was wrong, exactly, in isolation. But sitting side by side, they contradicted each other, and nobody caught it until I commissioned Claude to scan both sets of code side by side and inspect for even a one-character difference – and damn if it didn’t find one. One open bracket that was never closed caused the entire site to crash and stay crashed.

This wasn’t ambiguity, and it wasn’t quite the same as the button. This was a coordination problem. Each individual conversation the AI and I had was clean and self-contained, but nothing was cross-checking one session against the other, because I never asked it to, and the AI has no built-in instinct to go hunting through unrelated prior work looking for contradictions I didn’t flag. That’s on me, not on some flaw in the machine. If I want consistency across a project built in a dozen separate sittings, I have to be the one holding the master list and checking new work against it, because the AI’s context, however good, is not automatically stitched into some grand unified memory of everything I’ve ever built with it.

 

Why This Matters More as Projects Get Bigger

I want to draw the wider point out here because I don’t want you to think this is some niche coding problem that only applies to guys like me hand-building websites at midnight. This exact dynamic shows up anywhere a project has more moving parts than can fit in a single conversation. The bigger and more complex what you’re building becomes, the more likely you are to run into ‘failure type two,’ simply because there’s more invisible surface area for something to be hiding within.

That’s precisely why verifying AI’s output — the thing I hammered on relentlessly all the way back in Part 1 with Maduro and Gemini — isn’t just about catching outright factual mistakes. It’s about catching the button that looks fixed on paper but isn’t fixed on screen. It’s about catching the domain name that’s technically correct in one file and technically wrong two folders over. Verification isn’t a one-time gate you clear before launch. It’s discipline you carry with you the entire way through, because AI, for all its genuine brilliance, cannot verify against a reality it hasn’t been shown or at the very least, asked to look for.

 

The Rocks People Swing on Purpose

Everything I’ve walked you through so far in this article has been about honest mistakes — mine, and the AI’s, working together in good faith and still tripping over invisible wires. Before I close out this part, though, I have to circle back to something I raised in Part 1 and promised I’d keep threading through: not everybody using this tool is operating in good faith, and I’d be doing you a disservice if I let three whole parts go by without showing you what it looks like when the rock gets thrown on purpose instead of dropped by accident.

I’m not going to hand you a how-to here, and I’m not going to glorify the people who pulled these off cleanly, because the ones who got away with it don’t deserve the ink. What I do want to show you is that the system, imperfect as it is, does catch people, and it catches them more often than the doom-and-gloom crowd wants to admit.

Start with the courts, because this one’s almost funny if it weren’t so serious. Back in 2023, a New York attorney named Steven Schwartz filed a federal legal brief loaded with case citations that sounded completely legitimate — except the cases didn’t exist. He’d asked ChatGPT to find supporting precedent, and it obliged him by inventing several cases out of whole cloth, complete with citations that looked, on the surface, exactly like the real thing. When the judge and opposing counsel couldn’t locate a single one of them in any legal database, the whole thing unraveled fast, made the front page, and ended in sanctions.

That case became the cautionary tale every law student now hears at some point in their education, and it did not stop other attorneys from making the same mistake. Courts have handed out fines ranging from a few hundred dollars up through six figures for the same basic failure: taking AI-generated citations at face value and putting them in front of a judge without checking whether they were real.

As of this writing, one federal case in Oregon resulted in a combined six-figure fine against two lawyers for a filing containing dozens of fabricated citations and invented quotations, believed to be the largest penalty of its kind handed down in the country to date. Nobody who pulled that trick has gotten away clean. Every single one got caught, because the citations they leaned on didn’t actually exist anywhere for anyone to verify, and reality has an irritating habit of showing up eventually.

Then there’s the considerably darker end of the spectrum: the 2024 case out of Hong Kong where a finance employee at a multinational firm joined what looked like a completely normal video call with his company’s CFO and several familiar colleagues, only every single face and voice on that call had been generated by AI and built from publicly available footage of the real executives.

Believing he was following legitimate instructions, the employee wired the equivalent of over twenty-five million dollars across more than a dozen transactions before anyone realized what had happened. That fraud succeeded, and I’m not going to pretend otherwise, because pretending would defeat the entire point of this series. But the story didn’t end with the money simply vanishing into thin air. Hong Kong police connected that scheme to a wider fraud network and made multiple arrests tied to the broader operation, uncovering a supporting apparatus of stolen identification documents and fraudulent accounts that had been used to launder proceeds and defeat facial recognition safeguards elsewhere.

In short, the technology worked exactly as its operators hoped in the moment it was deployed, but the investigation that followed still closed in around the people behind it.

I’m including both of these because I want you to walk away from this part understanding something that doesn’t get said nearly often enough: the fact that AI can be weaponized doesn’t mean the weapon is undetectable, and it doesn’t mean the people wielding it are untouchable.

Every one of these cases got unraveled by the exact same discipline I’ve been hammering on since Part 1 — somebody, eventually, checked. A judge checked a case citation against an actual database. Investigators checked a paper trail of accounts and transactions back to the people who opened them. Verification isn’t just how you protect your own project from an honest ambiguity or a hidden theme quirk. It’s the same tool that unravels the dishonest ones, just running on a longer timeline and with higher stakes.

 

Lastly on This Point

I set out at the start of Part 1 to walk a line between the evangelists and the doomsayers, and I hope across these three parts I’ve managed to hold that line without face-planting. AI is not good. AI is not evil. AI is not omniscient, and AI is not useless. It is a tool, precisely as fallible as the information and instructions it’s given, and precisely as innocent or as evil as the intentions of the humans operating it.

I’ve asked you to be a gatekeeper. I’ve asked you to verify everything. I’ve asked you to notice which of your own words carry more than one suitcase before you hit enter. Now I’m asking you for one more thing to close this part out: take notice for when the problem isn’t your words at all.

Some failures will be yours to fix through clearer language. Some failures are the AI’s blind spot, and the only way through those is to open your eyes wider than it can open its own and show it what it’s missing.

Thank you for sticking with me through three parts of this. I promise now there will only be one more, but I also promise it will be an interview like none you’ve ever read – or listened to before. That’s because I’m going to audio-record it as well as posting the transcript of the conversation, unedited, into the body of part four. I sincerely hope you will be there with me for it.

One last point about the newly injected part four: So far, I have posted all of these on Medium because I just don’t use that site as much as I should, and these are really exactly the type of writing that excels there. That being said, part four is going to be different for a few reasons. First, the entire interview between Claude and me will be audio-recorded to add another element to the genius if AI for people who might not be aware, and traditionally, when I’ve tried to add audio content to articles published on Medium it has not gone well.

Also, there is so much controversy nowadays with AI-generated content, and the lines of that content are so ambiguous and blurry, I can’t be 100% certain if publishing an interview between a human and an AI chatbot would actually constitute a violation of ay of their terms of service and frankly, I just don’t have the time or patience to find out definitively. Because of that, the fourth and final installment of this series will not appear on Medium but instead will be available on the Editorials page of The Florida Sun Journal at floridasunjournal.com.

I genuinely didn’t expect this many of you to still be here by the end, goldfish attention spans and all, but here we are. Keep your prompts clear, keep your guardrails up, and keep watching the screen instead of just the results. I’ll see you in the next one.