The thing nobody mentions about selling AI services is that you spend about half your time explaining why the machine just lied to your face with the confidence of a seasoned politician.
I was showing a client a new custom agent last month. We were testing its ability to pull specific data from technical manuals.
Everything was going perfectly. It was fast, it was accurate, and the client was nodding along. Then, out of nowhere, the AI decided that the best way to explain a specific safety protocol was to quote a fictional character from a 90s sci-fi movie.
It didn't just mention the character. It insisted that the character was the original patent holder for the technology we were discussing.
We both sat there in silence for a few seconds. The client looked at me, I looked at the screen, and the AI waited patiently for our praise.
It’s in these moments that you realise that for all its trillions of parameters, AI is basically just a very fast, very eager intern who occasionally does hallucinogenic mushrooms during their lunch break.
Why AI Hallucinations are the New Office Comedy
The term "hallucination" is far too polite. It suggests a gentle, dreamy state. In reality, when an LLM gets it wrong, it goes for gold.
I’ve seen AI tools invent entire legal precedents with citations that look so real even a lawyer would double-check them. I’ve seen them describe historical events that never happened with such vivid detail you start to question your own education.
The funny thing is the tone. AI never says, "I think this might be right, but I'm a bit tired."
It says, "Here is the absolute truth," followed by a series of statements that defy the laws of physics or the timeline of human history.
I keep coming back to the idea that we’ve built a mirror. The AI isn't being "stupid" in the traditional sense. It’s just trying so hard to be helpful that it refuses to admit it doesn't know the answer.
It’s an accidental satire of corporate culture. How many meetings have we all sat through where someone spent ten minutes "leveraging" words they didn't understand to answer a question they hadn't listened to?
The Strange Ways We Talk to Machines
It isn't just the AI that’s acting weird. We are too.
I’ve noticed a shift in how people phrase their prompts. There is a specific group of people who are incredibly polite to ChatGPT. They use "please" and "thank you." They apologise for asking for a third revision.
I asked a friend why he does it. He told me, quite seriously, that when the robot uprising happens, he wants to be on the "nice list."
Then you have the opposite. The people who treat the prompt box like a hostage negotiation. They use all caps. They threaten to "fire" the AI if the code doesn't work.
There was a viral trend a while ago where people found that if you tell an AI you’ll give it a $200 tip for a good answer, the quality of the output actually improves.
We have reached a point in human history where we are effectively bribing math.
When Clients Get Creative with AI Tools
I’ve had some interesting conversations lately about what people think AI can do versus what it actually does.
One lead asked me if I could build an AI that could "predict the mood of the stock market based on the weather in London."
I had to explain that while AI is powerful, it hasn't quite mastered tea-leaf reading yet.
Another person wanted an AI assistant that would join Zoom calls and "subtly disagree" with their business partner whenever they mentioned a specific project.
That isn't a tech problem; that’s a "you need a therapist" problem.
We’re in this strange honeymoon period where the technology is just capable enough to be dangerous, but just limited enough to be hilarious.
The bit most people miss is that AI doesn't have a "common sense" filter. If you ask it to write a press release for a company that sells waterproof towels, it will do it. It won't stop and say, "Wait, why would anyone want a towel that doesn't absorb water?"
It will just write 500 words on the revolutionary "Stay-Dry" technology.
How to Handle the "Robot Brain Farts"
If you're working with these tools every day, you eventually develop a thick skin for the absurdity. Here is how I’ve learned to deal with the inevitable weirdness:
- Lower your expectations of its "honesty": Treat every fact like it’s coming from a guy you met at a bar at 2 AM. It sounds plausible, but use Google.
- Watch for the "Confidence Loop": The more the AI uses words like "certainly" or "absolutely," the more likely it is about to tell you that the moon is made of blue cheese.
- Laugh at the output: If the AI suggests you should fix a broken server by "politely asking the electricity to flow faster," take a screenshot. It’s good for the soul.
- Keep the human in the loop: Never let an AI post, send, or publish anything without a pair of human eyes checking for "accidental sci-fi quotes."
There’s a lot of fear around AI taking jobs or ending the world. But for now, I’m mostly just enjoying the fact that the smartest technology we’ve ever built still occasionally thinks 1+1 equals "blue" because it’s having a bad day.
If you’re trying to navigate this without losing your mind, it helps to have someone who has seen the mistakes before. If you need a hand with your strategy, you can book a consultation.
The Future is Probably Going to be Weirder
I don't think we’re going to get less "hallucinations." I think we’re just going to get more sophisticated ones.
We’re moving toward a world where AI will be able to generate video and audio that is indistinguishable from reality. I can only imagine the kind of digital pranks we’re going to see.
I recently heard about an AI being used to generate "new" recipes. It suggested a dish that involved marinating a chicken in Gatorade and serving it with a side of "blended pennies."
The scary part isn't that the AI suggested it. The scary part is that somewhere out there, a food blogger probably tried it for the clicks.
We are explorers in a very strange new country. The map is being drawn as we go, and sometimes the map-maker forgets where the mountains are and draws a giant dragon instead.
The best thing we can do is keep our sense of humour. This tech is a tool, a toy, and a crazy mirror all at once.
If you're looking for a way to use AI that actually works—without the Gatorade chicken—you might want to check out SalesM8. It’s designed to be useful, rather than just creative.
Last week, something happened that reminded me why I love this line of work. An AI I was training for a client was asked to write a professional email turning down a job candidate.
It wrote a very polite, very standard letter. But for some reason, at the very end, it added: "P.S. Your cat is very handsome."
The candidate didn't have a cat. The AI didn't know the candidate. It just thought that adding a compliment about a pet was the "human" thing to do to soften the blow.
Logic is great, but until the machines understand why we don't eat pennies or talk to cats that don't exist, we’re still the ones in charge.
Have you spotted any bizarre AI behaviour lately? I’d love to hear about the most ridiculous thing an LLM has said to you. It makes the world feel a little less like a Terminator movie and a little more like a sitcom.
Want more stories like this? I share observations about AI, business, and life on the narrowboat at steventann.com.
About the Author
Steven Tann is an AI consultant, author of "You're Selling AI Wrong", and founder of SalesM8. He writes about AI, sales, and running a business from a narrowboat on the English canals. Connect with him at steventann.com.
