1.4 AI Bloopers, Porch Walls, and Other Unintended Outcomes

In this episode, I thought I’d share a blooper reel, and I bet they're not at all what you might be expecting. They certainly aren't what I was expecting, and I thought it might be fun to share the unvarnished truth. These are classic examples of how AI can go wrong, and why a human in the loop, in this case me, needs to pay better attention.

The Chalk Drawing

First, I wanted to share my daughter’s reaction when I told her she was the star of the show in Episode 2. In the chalk drawing story, I  gave my daughter, who was four at the time, strict rules about drawing in the house: 'No drawing on any wall, surface, or floor in the house. No crayons, no markers, no chalk.'

Not only did she remember the chalk incident where she drew on the porch wall, she had a confession. She and her friends had actually covered the side yard wall in chalk too. If I’d walked around the side of the house that day, I would have had a double-whammy. Based on my initial reaction, she must have chosen not to mention the hidden masterpiece. Does this make you wonder how many stories your kids are omitting until they’re adults and safe?

Anyway, when I asked my Adobe Illustrator AI to create an image of a child’s chalk drawing on a porch wall, it was a surprisingly difficult assignment. I had drawings of chalk people literally rendered inside a brick wall and other crazy images. Nano Banana did a better job, but it made one thing clear: describing visual images to an AI is a skill that takes real practice. 

When you spend time prompting reasoning engines, you run into some hilariously unexpected outcomes. These are a few of my favorite personal AI bloopers:

The "Literal Interpretation" Trap:

When I asked an AI to organize a clean table of data and told it: “If a value is missing, leave it completely blank.” I meant leave the Markdown table cell empty. The result was not at all what I expected. Instead of an updated table, the AI printed out twenty rows of data where every missing field literally said: "Completely Blank". It followed my exact words, but completely missed my intent!

When working on the "I'm Not Eating That" story, I nearly overlooked the AI's error. The model treated the single quotes as control characters, which threw off the text. I only caught the error because I’ve encountered it often when pasting SQL and code.

The Endless Apology Loop:

When I catch an AI in a mistake, I try to correct it, but sometimes I end up watching it spiral into a crazy apology loop. I once politely corrected a single date in an output, and the model responded with three long paragraphs apologizing for hurting my trust, promising to reflect on its behavior, and then handed me back... the exact same wrong answer! It spent all its energy apologizing instead of fixing the problem.

The "Over-Eager Apprentice" Design:

I asked an AI to generate a clean, minimal graphic layout and told it to “make the title stand out.” The AI responded by rendering the title in giant, neon-yellow, bold, italicized, underlined capital letters with three fire emojis on either side. It looked like a neon billboard from 1970. The title definitely stood out; it wasn’t at all what I was looking for. I guess I should have added some context, like for a class exercise, make the title stand out. 

Amnesia Loop & Gaslighting

As if we haven’t already covered numerous ways AI can fail – while finalizing my podcast submission, I had the joy of experiencing two new ways AI fails, proving yet again that the human in the loop is indispensable. I carefully crafted a podcast submission paragraph over 7 revisions. I approved version 7 and asked AI to produce the final submission. Succumbing to an amnesia loop, a distracted AI randomly chose version 4.

Then, to top it off, AI kept flip-flopping the directions for podcast images. First, AI told me to make 3 layers in my image file – Background, Subject, and Title. Then AI warned me to not include the podcast title. I stopped to read the actual instructions, which require a Title layer. At which point, the AI sincerely apologized for the gaslighting.

At this point, you may be thinking I’ve experienced enough to coral my AI. Sadly, that is not the case. Let me share today’s blooper. While prepping this episode, I used Terminal and Pandoc to batch convert a bunch of Word docs into Markdown so the AI could process them. When the model summarized my notes, it actually added the Pandoc conversion steps directly into my narrative—as if command-line file utilities were part of my original writing! This is a perfect example of context bleeding where an AI confuses how a task was technically executed with the actual message or data being communicated.

It’s annoying how my list of context failures keeps growing: context bleeding, context window friction, semantic context friction, syntactic context friction. Unfortunately, I stopped to ask the AI what other context issues I could look forward to, and guess what?!

There’s context rot, context dilution, context poisoning, context truncation (sometimes called context loss), and temporal context drift. There’s also cross-document ambiguity, when you paste multiple files into a single context window where identical terminology or column names mean entirely different things. Oh well, I’ll apply the glass-half-full mentality and set these aside as podcast ideas for the future.

If you recall the episode with the silly 16th-century German idiom, you may be expecting a sound bite of my clumsy German attempt, but that’s not happening. Some things are just too embarrassing to share.

My attempts at speaking German were so bad I gave up and looked for an authentic version. Sadly, that did not go well at all. I typed the phrase "Da wird doch der Hund in der Pfanne verrückt!” into Google Translate. Initially, the AI assured me this vivid German expression conveys complete disbelief or shock. Let’s just say I was not prepared for the translation that Google Translate provided, and I can’t share it here. If you want to try it yourself, you can grab the phrase from the transcript, and you’ll see just how bad that went.

This example of semantic friction illustrates that the AI relies purely on probabilistic training data rather than anchored semantic metadata, and strips away the nuance, culture, and precise intent. Without clear reference frameworks, the model defaults to the path of least resistance, which is totally not what I expected. I’ll file this life lesson up: different AIs can provide different answers, in this case, translations.

Why These Bloopers Matter

We laugh at these glitches—and we should laugh at them—because they remind us of a fundamental truth: AI doesn't have common sense. It doesn't know what a porch wall represents; it doesn't know what minimalist design feels like, and it doesn't know Grandma's unwritten baking rules. It only knows the exact text and boundaries you hand it. So the next time your AI gives you a bizarre, unexpected response, don't get frustrated. Take a breath, laugh at the blooper, and ask yourself: 'Did I just leave an opening for chalk drawings?'

If you have a funny AI blooper to share with us, drop it in the comments below. If you’re interested in a Season 2 subscribe to let me know. I have some ideas based on what I’ve tried with AI, but honestly, who knows what direction AI is heading.

Season 2: Code, Syntax & Shaping

2.1 Creating Web Pages with HTML and CSS

2.2 Markdown

2.3 Shaping Data

2.4 Python Coding