Epic Fails With AI — And What They Actually Teach You
Introduction: The Juggling Act
Piotr Garlej walked on stage with five juggling balls and a plan. The plan failed immediately — and that, he explained, made it the perfect intro to his topic. Most speakers show you AI at its polished best. Piotr wanted to show you what happens when it goes wrong, because understanding why it fails is where the real value lies.
Lesson 1: Challenge AI — Never Trust the First Answer
ChatGPT gave Piotr a regional Polish dialect word with complete confidence. Twenty kilometres from his destination, he double-checked. It was wrong. He asked again. Wrong again. Three more times. Each answer delivered with the same unshakeable certainty.
A Columbia University report found that Perplexity Pro — one of the better models — fabricated answers in 68 out of 200 test searches. The antidote is the Hemingway Rule: the first draft of anything is always bad. Push back with prompts like:
"How do you know this?"
"Go deeper on that point"
"Explain your reasoning step by step"
You always deserve a better answer. Ask for one.
Lesson 2: Structure Your Prompts Properly
Piotr asked Midjourney for the Titanic hitting an iceberg. He got the wrong ship, the wrong scene, and eventually the wrong film. The problem was the prompt. His fix — the FSCS framework:
Form — photo, illustration, 3D animation?
Style — photorealistic, cinematic, anime?
Composition — close-up, full body, bird's eye view?
Scene — what is actually happening?
A structured prompt produced a result his eight-year-old son almost approved of. The same logic applies to text prompts: set a role, provide context, define the task, add at least three examples. Structure is the antidote to generic output.
One warning: be careful with copyright. Using a specific filmmaker's name as a style reference raises real legal questions. Descriptive alternatives — "in Hollywood blockbuster style" — are safer.
Lesson 3: Experiment — and Never Rely on Just One Tool
Piotr once accumulated thousands of credits on Tome, then a leading AI presentation platform. Tome pivoted, wiped every user's credits, and announced it was — in their words — "sunsetting." Gone overnight. The lesson: the AI landscape shifts constantly. Experiment widely, and when one tool fails you, try another. Midjourney could not add strawberries to an image without changing the whole picture. ChatGPT did it cleanly in one go.
Don't Delegate — Cooperate
In New York, Piotr observed a company where speakers had let AI write both their slides and their speaker notes. The result: entire presentations delivered while staring at a laptop, reading stiff AI sentences to a room that had long since checked out. No eye contact, no interaction, no persuasion.
As Garr Reynolds puts it, think of it like a burger. AI gathers the ingredients, AI polishes the result — but the human does the most important part in the middle. Nobody wants to eat just the bread.
Final Thoughts: Fail Better
AI makes mistakes. So do we. The difference is that it learns from every failure and improves. The same approach works for the people using it: challenge the output, structure the input, experiment constantly — and do not be afraid to drop a few balls along the way.
Join the Conversation
What is the worst AI fail you have experienced — and did it put you off the tool or push you to get better at it? Share your thoughts below.
Comments