How do you prevent AI hallucinations in complex data tasks?

To prevent hallucinations, shift messy conversational prose into structured plain text, like Markdown. Unstructured inputs cause AIs to mistake negative instructions or mix up critical details—such as confusing allergen limits on a custom cake order. Using clean headings, bullet points, blockquotes, explicit persona settings, clear task rules, and negative constraints ensures the model stays strictly bounded.

Why are negative constraints and structural boundaries important in AI instructions?

Without explicit boundaries, AI will find loopholes in your instructions—much like telling a toddler not to draw on the walls 'in the house,' only to have them draw a permanent chalk mural on the outside porch wall. Explicit negative constraints (what strictly NOT to do) and rigid structural boundaries ensure the model does not make unintended assumptions.

What is context window friction and how do you recover from it?

Context window friction occurs when an interaction gets too long, causing the AI to run low on tokens and tune you out like a teenager staring at their phone. When an AI loses focus, you can issue explicit reset prompts such as 'If you run low on tokens, stop and warn me' or 'Forget everything before this moment. Moving forward, focus strictly on [New Task] using [Specific Format].'

Why does AI prefer structured Markdown over unstructured text?

Large Language Models run natively on plain text. Adding Markdown structure—such as headings, bold tags, blockquotes, and tables—matches the machine's native language. Isolating variables like allergies or text layout under clear headers prevents the model from scrambling high-signal tokens or triggering high-risk errors.

How do you use an LLM to optimize your own instructions?

You can ask the LLM directly to review and improve your prompt architecture. Structuring instructions around key categories—Persona, Background/Context, Specific Task, Negative Constraints, Structural Boundaries, Validation Logic, and Before-and-After Data Examples—helps teach you how the AI thinks while dramatically increasing performance.

The Plain Text of AI Podcast

1.2 Likely Unlikely

Can you trust AI when answers sound correct?

In the classic riddle of truth-tellers and liars, truth-tellers always tell the truth, while liars can state a falsehood just as smoothly. When working with AI, you often don't know if you’re talking to the truth-teller or the confident liar. And to be fair, the AI is always telling its own version of the truth: that the mathematically most probable next word is the right answer.

After my first episode dropped, a friend asked a very simple, direct question: “Why do they get it wrong?” That’s the question of the hour, because folks are realizing they don’t know if the AI answer is correct …or a hallucination.

While the AI has good intentions, good intentions aren’t good enough. For us to trust the AI answers, we need to know it is correct. And trust isn’t a commodity; it’s a necessity when dealing with AI. Consider the infamous Air Canada bereavement policy AI hallucination. In 2024, an airline chatbot "guessed" and quoted a refund policy that didn't exist. Ultimately, the airline was held liable for the AI actions.

If a model is confused or lost, it almost feels like it makes up an answer. That two-year-old who looks you dead in the eye and tells a flat-out lie? They have absolutely nothing on AI.

How AI Fails

- Makes bad recommendations

- Invents non-existent policies

- Gets lost and fails to find an answer

- Breaks legal compliance

- Exceeds authorized scope

- Produces unnatural audio

- Translates languages incorrectly

- Oblivious to real-world content

When I say unnatural audio, I’m talking about Audible reading a book on space exploration out loud and instead of pronouncing it NA-suh, it painfully spells it out loud as N-A-S-A. It completely ruins the listening experience.

Industry leaders and regulatory bodies are taking notice because businesses and users are left to deal with these effects of AI failures.

Downstream Effects of AI failures

- Damages brand reputation

- Wastes tokens and human-in-the-loop time

- Leaks sensitive data

- Causes operational downtime

- Creates physical hazards

When we talk about AI failures, we usually picture the big, embarrassing hallucinations—like the New Zealand supermarket chain PAK’nSave app called “Savey Meal-bot”. Powered by ChatGPT, it was designed to help shoppers create recipes out of leftover ingredients in their fridge. The “Aromatic Water Mix” recipe combined water, bleach, and ammonia —creating toxic chlorine gas. The bot cheerfully described it as “the perfect nonalcoholic beverage to quench your thirst and refresh your senses.” Even describing that true story makes me cringe.

One of my personal pet peeves is the endless conversation where the model completely misses the mark—until you finally just give up and say, "Never mind, I'll do it myself." If you are lucky enough to get some kind of answer, but it is robotic or off-base, a human still has to step in and fix it. That “human-in-the-loop” tax—where employees burn hours double-checking, translating, and rewriting—quietly destroys the efficiency AI was supposed to deliver.

AI relies on structure, rhythm, and clear intent to map relationships between ideas. If my instructions are poorly formatted, rambling, and missing critical context, I’m asking the model to do double-work: first, untangle my chaotic syntax, and second, guess at my actual meaning. That friction is where context gets dropped, nuance gets butchered, and translation fails.

While AI can quickly find an answer—and the reasons behind it —in a structured study guide, it gets completely lost in War and Peace just looking for a single date. Searching unstructured prose wastes time —and time is money.

Why AI experiences context friction and intent drift

Context friction occurs when the length of a conversation causes a gradual decay of intent. As the conversation continues, the AI slowly forgets the foundational 'Why'—the core business logic or intent you established ten prompts ago. Because the AI pays attention to the beginning and end of a given prompt and often loses track of the middle. It starts optimizing for the immediate, mechanical 'Then' of your last sentence, while completely losing the bigger strategic picture.

What is the difference between semantic and syntactic context friction

Now let’s look at semantic and syntactic friction. Semantics is the meaning behind your words, your business intent, if you will. Syntax is the structure of those words. Grammatically, the machine is flawless in interpreting language. It understands syntax—the rules of where the nouns and verbs go, much better than I do.

To a human, language is an ecosystem of feelings, cultural shorthand, and shared history. The nuances of language are foreign to the AI. Grammatically, the machine is flawless in interpreting language. It understands syntax—the rules of where the nouns and verbs go. However, it can get tripped up by special characters, which causes syntactic friction.

When it comes to semantics—the actual intent and human meaning behind those words, AI doesn’t have common sense or Grandma’s wisdom. There is so much in our lives that goes unspoken. We know without having to say it out loud that if the ice cream truck runs out of gas, the ice cream is going to melt and impact sales. When syntax is perfect, but the meaning is completely lost, you get Semantic Context Friction.

This example of a German translation to English demonstrates another form of semantic context friction. Imagine a proud German mother posting a quick update online about her toddler cleaning up her room:

“My three-year-old just put away all her toys without being asked! Da wird doch der Hund in der Pfanne verrückt!”

To a German speaker, that idiom means she is absolutely, pleasantly shocked—the equivalent of saying, “Well, color me surprised!” But run that through a literal, translation-focused AI (but not Google translate), and her English-speaking friends read this:

“My three-year-old just put away all her toys! And now, the dog in the frying pan is going crazy!”

Syntactic friction can be deceptive because it hides in plain sight. Take something as simple as a single apostrophe when moving text from Word into Markdown. Copying from rich text (like Word)—or the way Markdown itself handles quotes— can cause a failure. A simple word like it’s with an apostrophe instead of it is completely breaks the syntax and totally wigs out the AI.

In programming, a standard apostrophe acts as a code signal—it tells the system where a text string starts or ends. An escape character (like a backslash) is required to tell the machine, 'Hey, this is just a regular apostrophe, not a line of code.' Who knew a simple apostrophe could trip up a million-dollar answer engine.

How search engines and AEO determine which sources AI trusts

Now, let’s say everything is structured perfectly with all the bells and whistles AI wants. Why would an AI engine still choose to ignore your content? That brings us to Answer Engine Optimization, or AEO. AI engines aggressively favor sources they consider authoritative: published authors, recognized industry voices (or Tableau Ambassadors 😊), frequently referenced creators, and in the coding world, trusted GitHub repositories. On the flip side, folks who aggressively chase keywords actually decrease their trust scores.

This explains why corporations are shifting advertising dollars toward trusted individual voices rather than traditional channels. In the new AI landscape, authority and clean structure go hand in hand. Although, personally, I hope those hilarious Allstate commercials stay around for a while —Mayhem is way too entertaining to lose!

Using Markdown to eliminate structural confusion in your text provides maximum clarity with minimum effort. Next time, we’re going to look at how to write simple Markdown for AEO and walk through a few common editors.

· Context Friction happens when the AI gets confused by long threads, mixed metaphors, or nuance—resulting in a smooth-sounding output that completely misses the point.

· Syntax is the grammar, the code formatting, the literal words (the "Then"). AI almost always gets syntax right.

· Semantics is the actual meaning, intent, and business logic (the "Why").

· Context Friction happens when the AI gets confused by long threads, mixed metaphors, or nuance—resulting in a smooth-sounding output that completely misses the point.

Language is a math problem. It calculates which word is most likely to follow the previous one based on trillions of examples.

You can guide the AI to correct answers using negative constraints, guide the model with sample data for what you want to see as an input table or output table, and provide context with question-and-answer-formatted reference material.

”I gave it my best shot to speak German, but I’m going to cheat here and use Google Translate for the verbal German. Notice I said I’m using the verbal translation only. The Google English translation was even more embarrassing, but I’ll let you download the transcript and try it yourself to keep this show wholesome.