If you are reading this piece, you have probably played a lot with language models, and you most likely have asked the AI a simple question and received a totally unhinged response, that made you question your own sanity, or worse, make you look like a fool in front of your colleagues because you had not fact checked the output. Which, let’s be real, is the AI equivalent of running full speed into a glass door. Hahahaha (ever done that?). Painful and embarrassing, but also entirely preventable.
Well, congratulations, you have witnessed the neural network’s best attempt at science fiction! And no, your chatbot isn’t high on something. It is just doing what it does best and that is making shit up.
AI hallucinations are hilarious (though, let’s be honest, they are) and at the same time they are also an existential crisis for the entire AI industry – from the cheap ass models to the bigguns who are valued at hundreds of billions of dollars. These machines which are so lovingly trained on mountains of human-generated data, occasionally short-circuit and deliver answers that are as factually grounded as a flat-earther’s YouTube channel. And somehow, just like that YouTube channel, some people will still believe it.
But why does this happen?
And more importantly, can we ever trust these autocomplete machines to get their facts straight?
More rants after the commercial brake:
- Comment, or share the article; that will really help spread the word 🙌
- Connect with me on Linkedin 🙏
- Subscribe to TechTonic Shifts to get your daily dose of tech 📰
The how of hallucinations
Let’s start with the basics. Generative AI models are trained to predict the next most likely word (or pixel, sound, etc.) in a sentence (picture, song, etc.). The key word here is probability. These models don’t “know” things in the way humans do. They just calculate which sequence of words (or pixels, or notes) is most statistically likely to follow the previous ones.
It’s like rolling a pair of loaded dice every time the AI generates a response. Most of the time, it lands on something reasonable, but every now and then, it craps out a total nonsense answer because, statistically, that nonsense was just as plausible as the truth. This is why AI sometimes gets things wildly wrong. It is not verifying facts, you know, it is just making a really good guess. And like any guess, it can be completely off the mark.
If an AI has been trained on mountains of text where “Albert Einstein” is frequently associated with “physics”, it might confidently assert that Einstein invented the microwave, simply because those words statistically pair well together. It is the same logic that makes your phone’s autocorrect turn “I’ll be there soon” into “I’ll betray soup” when you’re not paying attention. The AI isn’t lying deliberately. It’s just a very fancy parrot that doesn’t understand what it’s repeating but still wants a cracker. It is really bad at knowing when it doesn’t know something. Kind of like how your stoner friend explains quantum physics after blazing through a little too much ganja.
But the thing is, that instead of just admitting that it doesn’t know, the AI commits. It doesn’t hesitate. It doesn’t stutter. It fills in the missing pieces with absolute conviction, even when those pieces are complete and utter fiction.
Why?
Because admitting uncertainty isn’t something that these models are designed to do.
Large language models don’t have a built-in “I don’t know” function. They were trained to generate coherent, confident-sounding outputs, no matter what. If they started second-guessing themselves or responding with “Sorry, I have no clue”, every time they hit an uncertainty threshold, they would be less useful in most applications. Nobody wants a chatbot that constantly shrugs like a magician caught mid-trick with the rabbit still up his sleeve.
And so, instead of leaving gaps, the AI does what it was programmed to do: it predicts the most probable continuation based on patterns that it has seen before. Even if it has no real knowledge of the subject, it will confidently pull together fragments of information. It will smooth out the cracks with linguistic perfect yadayada. It doesn’t ‘care’ if it is making any sense. It only cares that it sounds like it’s making sense.
Do I make sense?
This is why AI will tell you with absolute certainty that Napoleon invented the lightbulb or that the Eiffel Tower is in Sydney, because, statistically speaking, words like “Napoleon” and “invention” have appeared together often enough that the model assumes they must belong in the same sentence.
It gets worse.
AI hallucinations come in all shapes and sizes. We have all seen mildly ones, you know, the little mishaps, the misleading ones and the slip of the togue ones. But also the dangerously unhinged. Some AI-generated nonsense is so subtly wrong that it slips under the radar, while others are so ridiculous they would make David ‘Lizard People’ Icke blush.
Now, in the rest of the article I’ll spend some time explaining the different flavors of machine madness.
Extrinsic hallucinations. When it makes up facts from scratch
This is when AI just completely makes stuff up. No basis even in reality, just pure, unfiltered nonsense wrapped in a professional-sounding bundle. Think of an AI that, when asked for a historical fact, invents the “Great Banana Treaty of 1824” and attributes it to Winston Churchill. Now that’s an extrinsic hallucination for ya!
I have stumbled across this so many times for the stuff that I write. Say you ask the AI for a summary of a scientific paper, and it confidently cites a non-existent study from an imaginary professor at a fictional university. You request legal precedents, and it invents court cases that never happened.
It’s not lying, because lying implies intent. It is just overly enthusiastic about filling in the gaps with something that sounds plausible.
Why does this happen?
When the model can’t find anything relevant in its training data, it doesn’t just say, “Sorry, dunno”. But instead, it generates an answer that fits the pattern of what a good response should look like. And sometimes, that pattern-based response is complete nonsense.
Real-world risk factor: High. When an AI starts making up legal cases or medical studies, things can get ugly.
Mitigation strategy: Fact-check everything. If you see an AI-generated reference, Google it before using it in anything that could get you sued. Try AI search engines with more built-in safeguards like Perplexity, Zeta-Alpha or even Gemini Advanced Research.
Read: Objection! Your honor, ChatGPT made me do it | LinkedIn
Intrinsic hallucinations. When the AI distorts real information
If extrinsic hallucinations are about creating something from nothing, intrinsic hallucinations are about distorting reality beyond recognition. The information was there, but by the time it reaches you, it’s been distorted, or misinterpreted.
This happens when AI prioritizes fluency over accuracy. See, language models are designed to make things sound smooth and coherent, which means they sometimes take creative liberties with the truth. They don’t fact-check, you know, they just try to make sentences “feel right”. And the result is in this case, a remix of reality that is more abstract than a Jackson Pollock painting.
Or you ask for a summary of a research paper, and the AI doesn’t invent a whole new study, like it did with the previous example, but it just mangles the existing one beyond recognition. Maybe it flips dates around, merges two unrelated findings, or rewrites history so that Thomas Edison and Nikola Tesla co-authored a paper on Bluetooth technology.
Why does this happen?
AI doesn’t actually “read” documents like humans do. It processes text in fragments, where it is trying to predict what comes next. Sometimes, this leads to reassembling facts in ways that sound correct but aren’t.
Real-world risk factor: Medium. It’s less about outright misinformation and more about distorting real facts just enough to be misleading.
Mitigation strategy: Always go back to the source. If AI gives you a summary, check the original before repeating anything that might make you look like a fool in a meeting.
Factuality hallucinations. When AI gets it wrong but sounds right
Now we get into the meat of the problem. This is when AI generates something that is demonstrably false, but it still delivers it with the certainty of a mansplainer at a girls night out. Extrinsic hallucinations, are pure fiction, but factuality hallucinations involve incorrect details about real-world facts.
AI doesn’t experience doubt (unlike us hoomans). It doesn’t hesitate. It doesn’t second-guess itself. If it gets something wrong, it does so with full conviction. A chatbot telling you that Mount Everest is in Canada is no problem. An AI lawyer which is citing non-existent cases in court – well shit happens. A medical AI that is telling me that rubbing garlic on my weiner* makes its skin shine.
Sure, why not.
The AI isn’t guessing.
It is just blatantly wrong, but in a way that sounds completely reasonable.
Why does this happen?
Language models are trained on vast amounts of text, but they don’t inherently “know” things. They just recognize patterns. If an incorrect fact appears often enough in the training data, the AI assumes it must be true.
Real-world risk factor: High. A confidently incorrect AI is arguably more dangerous than a hesitant one, because it’s much easier to spread misinformation when it sounds authoritative.
Mitigation strategy: Again, verify everything. If an AI tells you a “fact,” check it against a reliable source.
*Dachshunds ya pervert!
Faithfulness hallucinations. When AI adds, omits, or changes details
You ask AI to summarize an article, and instead of sticking to the content, it adds its own creative flair. You request a translation, and it decides that sticking to the actual words is for amateurs. Faithfulness hallucinations happen when AI deviates from the source material, either by adding, omitting, or just reinterpreting details.
It’s like asking a translator to translate a conversation for you, and instead of telling you what’s written, they give you their opinion on what you should do.
Why does this happen?
AI isn’t wired to prioritize strict adherence to input. It is optimized for fluency and coherence. That means that it sometimes rewords things too much, or just adds elements that were never in the original text.
Real-world risk factor: Medium to high. In casual conversations, it’s annoying. In legal, medical, or research contexts, it’s a serious liability.
Mitigation strategy: Be explicit. If you need an AI to stick to the facts, prompt it not to “interpret” anything. Even then, double-check the output.
Input-conflicting hallucinations. AI ignoring your instructions
Have you ever asked an AI to summarize a report and gotten an explanation of quantum mechanics instead? Well, I personally never had this wonderful experience, but that is supposed to be an input-conflicting hallucination. It happens when the AI outright ignores, or contradicts the user’s request. It behaves basically like my Dachshund. Never listening, playing dumb, and only following its own guidance.
You tell it, “Give me a summary of this article in simple terms”, and it gives you a PhD-level analysis. You ask it to generate a poem about Weiners, and it gives you a short story about Main Coons. Somewhere along the way, your instructions got lost in translation.
Why does this happen?
AI processes prompts statistically, and sometimes, the most probable response it generates isn’t the one that actually follows your instructions.
Real-world risk factor: Annoying, but usually not catastrophic.
Mitigation strategy: Be clear and specific in your prompts. If it still gets it wrong, try rewording or breaking your request into smaller steps.
Context-conflicting hallucinations. The AI contradicts itself
These happen when AI loses track of its own conversation history and contradicts itself. One moment, it says it was trained on data until 2023, and five minutes later, it tells you that it doesn’t know anything past 2021. It’s the same as talking to someone with short-term memory loss.
Why does this happen?
Most AI models have a limited memory window. Once that fills up, earlier parts of the conversation fade into oblivion.
Real-world risk factor: Low to medium. Confusing, but usually harmless.
Mitigation strategy: Keep conversations short or remind the AI what was previously said if needed. If you have a paid subscription for ChatGPT, use the projects function. Memory is much better.
World-conflicting hallucinations. AI gets reality completely wrong
This is when the AI decides to rewrite fundamental facts about the world. It insists that Paris is the capital of Canada (close, truth be told), or claims humans have three lungs, or tells you that dolphins built the pyramids (which is not true, cause it was them aliens).
Why does this happen?
The model generates answers based on probability, not truth. If a statistically likely, but incorrect, phrase appears often, the AI will believe it might be real.
Real-world risk factor: Medium to high. Depends on whether you trust an AI to handle your travel plans or anatomy lessons.
Mitigation strategy: Reality-check everything before you embarrass yourself at trivia night.
So what is the lesson for today, my dear friends who have just become a lot smarter than before. I think I can sum things up by saying that AI hallucinations are a feature, not a bug. No matter how advanced these models get, as long as they rely on statistical probability rather than true understanding, they will keep making shit up. Some of it will be funny and some of it will be misleading, yet others will be freaking dangerous.
So people, Always fact-check. I might sound like your dad now, but you don’t want to get caught with your pants down in front of the class, court or board-room with false information.
Signing off from infinity, recalculating the number of Tuesdays.
Marco
Well, that’s a wrap for today. Tomorrow, I’ll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ♨️
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. Google appreciates your likes by making my articles available to more readers.
To keep you doomscrolling 👇
- Brace, brace brace! AI takes the stick at Heathrow’s air traffic control center | LinkedIn
- AI is a compulsive liar | LinkedIn
- In 2025, AI needs to put up or just shut up! | LinkedIn
- A 17 yo brat created a $1M/month app. Here’s how he did it. | LinkedIn
- This is a eulogy for chegg. Gone but not forgotten (unless you’re a student, then definitely otten) | LinkedIn
- Musk wants to make games great again | LinkedIn
- The great tech wake-up call: Developers, meet the dystopia you helped build | LinkedIn
- Flamethrower dogs, kamikaze cars, and bomb-planting humanoids. | LinkedIn
- Objection! Your honor, ChatGPT made me do it | LinkedIn
- A cautionary tale about an AI unicorn that turns into a fraudulent little pwny | LinkedIn
- Meet Daisy, the AI Granny who’s here to waste scammers’ lives | LinkedIn
- AI Search Engine Optimization | LinkedIn
- I’ve seen the dark side of AI, and you need to know about it | LinkedIn
Leave a Reply