AI with Kyle Daily Update 182

The Chicago Sun-Times printed a 15-book summer reading list last year. Ten of them did not exist.

Fake titles. Fake plots. Andy Weir's The Last Algorithm… Andy Weir didn't write The Last Algorithm. Isabel Allende, Min Jin Lee, Maggie O'Farrell, Percival Everett - all credited with books that have never existed and never will. Ironically the model included a real Ray Bradbury one. Bradbury famously hated computers and the internet - he would have HATED generative AI! Fitting.

The journalist did not check. Nobody at the Sun-Times checked. It went to print.

Embarrassing.

And it's not just them. Deloitte (in)famously shipped two AI-hallucinated reports last year. One to the Australian government. One to a provincial government in Canada. Fake footnotes, fake citations, the lot. Sold to actual governments. By a Big Four firm. They got bullied for it, rightly…and a fair amount of that was me. Sorry!

Grab the full slide deck + resources here https://aiwithkyle.com/resources/ai-hallucinations

The Numbers Are Mad

According to an Exploding Topics poll 92% of people never bother to verify what their AI tells them.

Out of every 100 people prompting ChatGPT, 92 just take the answer. Paste it into the doc. Send it off to the client. Print it in the paper. Just ship it out the door. Whatever.

About 45% of AI responses have significant problems (BBC/EBU tested 3,000 of them). MIT measured something even nastier - the model is 34% more confident when it's wrong than when it's right. So when it sounds the most certain… that's exactly when you should be the most suspicious.

And there are 486+ court cases globally that involve AI-fabricated citations. That we've caught. The actual number is far higher. Lawyers submitting briefs with cases the AI invented.

This is the world we're in. Nobody checks. The AI is most confident when it's wrong. Hallucinations are seeping into the published record.

Hallucinations Aren't A Bug

Go and read Andrej Karpathy’s “dream” tweet as its important:

Here’s the most important part:

Hallucinations is all that Large Language Models (LLMs) do.

Remember that.

The model is not looking anything up. There's no database. No filing cabinet of facts with True/False stickers next to each one. It does not know things.

When you type a prompt, it predicts the next most probable token, then the next, then the next. It's been trained on roughly the entire internet - all the books it could scrape, every YouTube transcript, the lot - and from that mountain of words it learned what confident, correct-looking answers tend to look like.

That's all. Tbf that’s still SUPER useful and we can do amazing things with it. But that’s the underlying mechanism.

So technically… every output is a hallucination. We just call it intelligent when the output lands on something true and call it a hallucination when it doesn't. But it’s the same mechanism either way. We're the ones adding the value judgement, after the fact.

This is why the "AI lied to me" thing is a category error. A liar knows the truth and hides it. The AI has no concept of truth. It cannot lie.

Type 1 vs Type 2

There are two flavours of hallucination and only one of them is easy to spot.

Type 1 - Invents something that does not exist. Fake book. Fake court case. Fake statistic. The Sun-Times list is the textbook case. These are the obvious ones - go to fact-check, the thing isn't there, five seconds on Google and you're sorted.

Type 2 - Drifts from a document you actually gave it. This one's nasty. You hand the model a real PDF, a real report, a real source. It then misquotes it. Changes a number. Reorders a paragraph. Summarises a thing that isn't quite what the document said. This is verisimilitude (what a Wow word!) - the semblance of truth.

Type 1 you can catch.

Type 2 is the insidious one that’ll catch you unaware.

Will This Get Fixed? OpenAI Says Nope.

September 2025, OpenAI published a paper called Why Language Models Hallucinate. The headline finding: hallucinations are mathematically inevitable with current training methods.

Why? Because models are trained to answer. Abstaining gets penalised in training. So when the model doesn't know, it guesses… confidently. That's the optimal strategy under the current reward structure.

It’s kinda like if you take an exam. If you write NOTHING you will get 0 points. If you write something resembling an answer (even if you have NO idea what you are talking about) you’ll get more than 0. So: write something. That’s basically what the models are trained to do.

What Actually Helps (In Order Of Effort)

So: hallucinations happen. And will continue to happen. What can we do?

Here are 4 easy modifications.

Turn on web search. This is the single biggest free win. ChatGPT, Claude, Gemini all have a toggle. On the free plans it's sometimes off by default. Flip it on. GPT-5 with web search makes around 45% fewer factual errors. Easy win.

Use the stronger model for high-stakes work. Claude Opus 4.6 or 4.7. GPT-5 Pro with thinking on. Gemini 2.5 Pro. They're slower. They cost more - or have limited usage on free plan (this is one of the biggest reasons to oay). But when the output is going somewhere public… use them! I built today's slide deck on Opus and the research alone was 20 minutes of thinking time before it even started drawing. That's the price of getting it right. Probably the best $20/month most people can spend, tbf.

Deep research / extended thinking. Toggle it on for anything where the answer needs to be correct, not just fast. The model goes off, checks citations, pulls real sources, comes back with something defensible. Every chatbot has its own name for this and they all do roughly the same thing. Use it.

NotebookLM for grounded learning. This is the bonus one. NotebookLM gives you a personal RAG system - you upload PDFs, transcripts, books, and the model is forced to answer from those sources only. Citation hallucination drops dramatically when the model is grounded in your documents. Great for studying. Great for working with research papers. Great when you need this answer from this document… not the model's vibey “memory” of it.

These are more about the harness and tools around the model than the model itself. This is where most improvements in accuracy occur now.

Remember it’s not Google

The reason all this is hard is that we've been trained for 25 years to think "computer + question = answer". You typed it into Google. Google found the page. The page had the answer. Find the thing, show you the thing.

That mental model does not apply here. It’s over.

AI is a writing tool. A drafting tool. A thinking tool. It's excellent at working with information you give it. It's unreliable at sourcing information from its own memory - because it doesn’t have an indexed database like Google does. Treat it accordingly.

The people getting the most out of AI right now aren't treating it as a smarter Google. They realise it’s not just Google 2.0. It’s a tool for different types of work. Use AI to work with information. Don't trust it to source information from memory.

Resources: you can get the full slide deck from the Livestream and a PDF quickstart guide here https://aiwithkyle.com/resources/ai-hallucinations

Kyle