This week, I'm handing the mic 🎤
Clothilde is finishing her MSc in Business Analytics and her thesis is about AI hallucinations. When she offered to write this issue, I said yes immediately. You're getting a real expert explaining exactly why your AI makes things up, written in plain language.
☕ Welcome
Hey, it’s Clothilde
I've spent countless hours researching AI hallucinations, and when I am writing papers and in a rush, I like to ask ChatGPT to write my references in APA style for me (sue me ☺️). Very often, I spot mistakes in the title, year or authors. Sometimes, it will even invent names, or give fake links. I know that because I double check, but a judge in Quebec wasn’t so lucky.
You may have seen this in the news: Justice Geoffroy, presiding over the Robert Giroux white-collar crime case in Quebec, cited cases that don't exist. The Superior Court of Quebec took notice, and now AI use in legal proceedings is under review
But this isn't new, nor proper to the law: AI hallucinations are showing up across all fields at a growing rate.
If you've ever asked ChatGPT, Claude, or any other conversational AI to write a business report, explain a historical event, or explain your rights as a citizen, chances are the answer you got wasn't entirely faithful to the facts. Worst case scenario, it was completely made up. And here's the interesting part: it doesn't matter how reliable your source is, or how straightforward your question seems. The tool can still get it wrong. In fact, it's bound to do so.
🤯 WAIT, WHAT?
Built to answer, not to be right
We touched on this back in Issue #5: LLM’s (the technology behind tools like ChatGPT and Claude) are not 100% reliable, and that's okay. You just have to use them accordingly.
Here's what the research says. According to the Stanford 2026 AI Index Report, hallucination rates across top GenAI models range from 22% to 94%. What's striking isn't just how high the numbers are. It's the pattern: when a false statement is presented as something another person believes, models handle it reasonably well. When the same false statement is presented as something the user believes, performance collapses. AI struggles to separate fact from belief, and that gap doesn't shrink with more complex models, it grows.
But why does this happen? The answer is in the structure. Research has found that LLMs don't hallucinate by accident, but by design. The way these models are trained and interact with humans makes hallucination not a bug to be patched, but a feature of how they function (Huang et al, 2025).
The main causes are:
Knowledge boundary: you can only train on so much, and the knowledge is not always up to date.
Interpretation biases: may be influenced by human induced biases (during human learning reinforcement phase).
People-pleasing tendency: also called “sycophancy”, which makes it hard for them to contradict you if you’re wrong (as it doesn’t differentiate between facts and beliefs).
AI doesn't typically say, "I'm not sure" it is built to answer. And does not always back down when you call it out. Detection tools exist but are inconsistent, and tracing sources manually can be time exhaustive.
Hallucinating AI is not going away. But it is manageable.
⚡ TRY IT TONIGHT
Ask it to show its work
You can't stop AI from making things up but you can stop being surprised by it! Add this to the end of your next important question:
"Before you answer, tell me which parts of your response you are confident about and which parts I should verify myself. List the real sources you are drawing from, and clearly flag anything you cannot trace to a verifiable source."
Example of the response with the Prompt:
Notice how it flagged the source it couldn’t access “What you should verify”. That’s the prompt working. It won’t eliminate the hallucinations but it will warn you.
Example of the response without the Prompt:
Same question, no prompt. It still gave a full answer. No flags 🚩 or warnings ⚠️.
You can't eliminate hallucinations, but you can stop being caught off guard by them.
Pair it with these habits:
Understand the tool you're using: Knowing that hallucinations are structural, changes how you interact with AI output.
Triple check, every time: Never take an AI answer at face value, no matter how confident it sounds.
Make everything traceable: Keep track of where information comes from. If you can't trace it back to a real, verifiable source, don't use it.
Own it: If something goes wrong, the machine will not be held responsible, you will. That’s accountability. The judge in QC learned that the hard way.
Pass it on: Tell a coworker. Mention it to a friend. Share this with the group chat. Not to be a downer, but because most people using these tools are not always aware of the limits.
📱 THIS WEEK IN AI
Good Stuff From Around the Internet
Ask DoorDash launched this week. You can now snap a photo of a recipe, describe a craving, or paste a grocery link and the app builds your cart. What does this tell us? AI agents are moving into apps you already use every day.
The catch? AI gets your order wrong and suddenly you have all the ingredients for chicken chow mein, not the end of the world. But it's a good reminder to glance at the cart before you check out.
Anthropic (Claude) had a wild week. On Tuesday last week they launched Fable 5, their most powerful model ever made available to the public. On Friday, the US government issued a directive blocking any foreign national from using it, and Anthropic pulled the models for every customer worldwide. Gone in 72 hours.
This gave me a bit of a wake-up call, I use Claude constantly. It's my default at the moment. Watching a new model disappear overnight for reasons completely outside my control reminded me: building your whole workflow around one tool is its own kind of risk. The good news is that ChatGPT, Gemini, and others are all genuinely good right now. The smartest move is to get comfortable with at least two. Think of it like knowing more than one route to work.
🌟 BEFORE YOU CLOSE THIS TAB
You’re still the fact-checker
Once you understand the limits of a tool, you have two options: put it down, or pick it up knowing exactly what it can and can't do.
LLMs have definitely improved in accuracy over the years. But hallucination will not disappear anytime soon. It's part of how they work, and that's unlikely to change. This is not meant to scare you away from using them - just to help you understand their caveats so you can leverage them (just like with any other tool).
Now, that reality is a strong argument against the idea that AI will ever replace lawyers, doctors, researchers, or anyone else whose work depends on getting facts right.
Thanks for reading. If you want to geek out on AI hallucinations with me, come find me on LinkedIn ✌🏼
Your unfair advantage, one week at a time.
Got a question, a topic request, or just want to say hi? Just reply to this email. I read everyone.