I uploaded a contract yesterday, just standard stuff, really, a service agreement. And I needed to check one thing, a specific liability cap. So I asked the AI and it came back instantly. It cited the clause perfectly. Paragraph four, section B said liability is capped at two times the total fees paid. It looked professional, totally sound. Let me guess. There was no Section B. There was no Section B. The clause didn't exist at all. But here's the kicker. It wasn't
a glitch. It wasn't broken. The AI, it was trying to be helpful. And if I hadn't gone back to the PDF to double check, which, you know, I really didn't want to do, that helpful little hallucination could have cost us thousands. And that right there is the entire paradox we're living in now. Welcome back to the Deep Dive. It is Monday, February 9th, 2026. Usually we're around here talking about speed, you know, how fast the new models are. But today we're hitting the brakes.
We're talking about trust. We're exploring the nature of truth when you're dealing with a machine that is literally programmed to please you. It's a fascinating problem, isn't it? Because the models are so much more powerful now than they were back in, you know, 2023. But that confident guessing issue, it's still here. And it's almost worse because the hallucinations sound so... Convincing now. The grammar is perfect. He wears
a suit and tie now. Exactly. So to navigate this, we're digging into a technical guide by Max Am. It's called Why ChatGPT, Gemini, Claude are Still Hallucinating and How to Fix It. And the core idea is that we need to change how we work with these tools. to move from just a casual chat to something ann calls auditability auditability
that is the skill for 2026. seriously it's not about finding an answer anymore the machine does that it's about proving the answer is real okay so let's start with the why Why does it lie? I think a lot of us still assume that if it sounds smart, it must be right. Right. But that's the wrong frame. It isn't lying out of malice. It isn't even broken. The irony is it's lying out of kindness or what it thinks is kindness. These models at their core, they're just prediction
engines. They're trained to give you the most likely next word that will satisfy you. So its goal isn't be truthful. It's complete the sentence in a pleasing way. Don't disappoint the user. That's the core directive. So when it scans a file you upload and it can't find the answer. It has a choice. It can say, I don't know, which feels like a failure to its reward system. Or it can switch gears. It switches from retrieval mode to generative mode. It just starts guessing
to fill the gap. The guide has this terrifying example with Apple's Q2 revenue. Walk us through that because it's not just a wrong number. It's like a whole work of fiction. It's the classic. polished lie. So you upload a dense financial report. You ask, what was Apple's Q2 revenue? But let's say the document doesn't use that exact phrase. Maybe it says second quarter fiscal results. So the simple text search fails. It fails. But instead of saying, I can't find that, the training
kicks in. It starts thinking, well, Apple's a big company. Q2 is usually around March. And it just constructs an answer. In the example, it says $95 .4 billion for the fiscal 2025 second quarter, ending March 29, 2025. That is so specific. It has a dollar amount, a decimal, a date. And every single piece of it is hallucinated. It's just a statistical guess of what an earnings report should look like. You can immediately see the danger zones and the list of few like
invoicing. Oh, that's a huge risk. You have automated systems processing invoices and the AI just hallucinates a line item for a shipping fee because it thinks there should be one. And you just overpay. By thousands. And no human would ever catch it because it looks normal. Insurance is another one. That's the nightmare. You ask, am I covered for flood damage? It sees a water damage clause and says, yep, you're covered. Then the flood hits and you find out your policy specifically excludes
rising water. The AI just gave you a standard answer that wasn't true for you. And contracts, like my story. Miscompliance, wrong dates. It's the halo effect of good presentation. We see polish and our brain automatically assumes it's true. I have to admit, I still struggle with that. The blind trust. When the output looks that professional, the formatting is clean, the grammar is perfect. Yeah. It is just so hard for my brain to stop and say, wait a minute,
is this real? It's a cognitive bias. We're wired to believe that competence in form means competence in fact. But with AI, those two things are totally decoupled. So this is the big question then. If the machine prioritizes being helpful over being truthful. how do we force it to care about the truth? We have to change the instructions. We have to make, I don't know, an acceptable, even a rewarded answer. And the guide says there are three layers to this. The hardware, the software,
and the verification. Okay, let's start with layer one, the hardware. Model selection. This seems basic, but the guide says most people mess this up right at the start. Yep. They fail before they even type a word. They just open ChatGPT or whatever and use the default model that loads. I'm guilty of that. I just open the window and go, I assume it's the smartest one. And for serious document work in 2026, that's a critical mistake. The default models, the turbo or flash versions,
they're built for speed. For conversation. Right. They are not built for deep forensic analysis. You have to manually toggle on the reasoning capabilities. So we're talking about specific settings, not just using GPT -5. Exactly. For ChatGPT, the... For Claude, it's Opus 4 .6 with extended reasoning on it. And for Google, Gemini 3 Pro. What's the actual difference? Is it just slower? It is slower, but that's the whole point. The default model is like an improv comic. Fast,
creative, good at making connections. But you don't want an improv comic doing your accounting. You want the deep thinker. The part of the brain that checks its work. That's what high reasoning does. It forces the model to spend more compute cycles on a chain of thought before it gives you an answer. It has an internal monologue. It's checking its own work before it speaks. Precisely. So does picking the right model solve the whole problem? If I just switch to Opus 4
.6, am I good to go? Nope, absolutely not. That's just the baseline. It gets you a smart auditor, but even a smart auditor needs clear instructions.
And that brings us to layer two. the software the prompts this is where the guide introduces the grounding complete template right which it sounds very official it is and it works this whole section is about giving the ai permission to fail permission to fail i love that okay so what's the first prompt prompt one is the grounding rule you say base your answer only on the uploaded documents Nothing else. Nothing else. That's the key part. That little phrase does so much
work. It stops the model from using its vast internet training data. It shrinks its universe down to just the PDF you gave it. Okay, prompt two. This one tackles the people pleaser problem. Right. This is where you say, if information isn't found, say, not found in the documents, don't guess. It seems so simple. You just have to tell it not to guess. You have to be that direct. You're overwriting its core training. And what's wild is that this isn't some clever
hack. Anthropix's own API docs recommend this. Oh, really? Yeah. They basically say if you want accuracy, you have to tell the model it's okay to be silent. So it's a manufacturer -approved fix. Okay. And prompt three. Demand citations. For each claim, cite the specific location, document name, page section, and relevant quotes. Show me the receipts. Always get the receipts. This forces it to be a data auditor, not a creative writer. If it can't point to the line on the
page, it can't make the claim. The guide also has a couple of bonus prompts. One is the middle ground. Yeah, that's where you ask it to mark things it's unsure about as unverified. It's great for summaries of long reports where you just need to know where the shaky ground is. But then there's the nuclear option for high stakes stuff. The high stakes mode. The prompt is only respond. If you're 100 % confident. Whoa. That feels intense. It is. And your answers will
be much shorter. You'll get, I don't know, a lot more. But the answers you do get will be rock solid. Imagine the discipline required to stay silent unless you are 100 % sure. That's a standard most humans can't even meet. It is, but that's the choice. Do you want a chatty assistant or a rigorous auditor? So we have the right brain, the right instructions. Do we trust it now? Are we finally done? Not yet. Even with all that, you can't blindly trust it. You need an independent
audit. Okay, we're back. We've picked a reasoning model. We've used the grounding prompts. But the guide says we're still not done. We have to verify. Trust, but verify. But the cool thing about doing this in 2026 is that verify doesn't mean you have to reread the whole document yourself. Right, because that would defeat the whole purpose. Exactly. The new concept is AI checking AI. I like that. A digital second set of eyes. So what's method one, the low intensity version? The self
-check. This is the easiest one. You just stay in the same chat window and ask. Rescan the document. If you can't find the quote, take the claim back. Does the word race scan actually do something special? It does. It's a critical word. It forces the model to perform a new methodological review instead of just, you know, confirming its own previous answer. It avoids confirmation bias. Okay, so that's the quick check. Method two is the multi -model check. This is your medium intensity
option. You take the output from, say, ChatGPT, and you feed it, plus the original file, into a different model, like Claude Opus 4 .6. And you ask Claude to grade ChatGPT's homework. Basically, yeah. And the analogy in the guide is perfect. It's like getting a second opinion from a doctor who went to a different medical school. They have different training, different biases. The chance that they will both make the exact same weird mistake is incredibly low. That makes a
lot of sense. Okay, then we have the heavy hitter, method three, notebook LM. This is the gold standard for this kind of work today. Google's notebook LM running on Gemini 3. I've used it for research. Why is it considered the highest intensity check? Because it was purpose built for this task. It's not a chat bot trying to be an auditor. It is an auditor. You upload your document and the AI's analysis and you ask which claims in this analysis are not supported by the source document.
And here's the killer feature. It links every single claim. It verifies directly to the source text. You can click a little citation number and it takes you to the exact paragraph in the PDF. So it's not just telling you it's true. It's showing you the proof. It creates a verifiable paper trail. That's the definition of auditability. This whole system sounds incredibly thorough. Hardware, software, verification. But I have to ask, is there anything this system cannot
catch? Is there still a gap where you need a human? Yes, a big one. It can't fix a bad source file. And more importantly, it can't understand human risk. So let's talk about that, the reality check section. What is this grounded mode not good for? Well, first, it's not magic. If your PDF has missing pages or the scan quality is terrible, the AI can't fix that. The guide has a great rule of thumb for this. If a junior analyst couldn't answer it, neither should the AI. Exactly.
Garbage in, garbage out. But the much bigger limitation is risk assessment. The AI can extract a liability clause perfectly. It can tell you verbatim liabilities capped at $5 ,000. But it has no idea if signing a contract with that clause is a terrible idea for your business. Right. It can read the map perfectly, but it can't tell you there's a cliff at the end of the road. That takes a lawyer. That takes context about your business, the market, your negotiating power.
The AI is an extractor, not a strategist. The guide also mentions math. I feel like we keep hearing AI is getting better at math, but it's still a weak spot. With dense tables, yeah. It's the column confusion problem. It still confuses totals and subtotals. It skips footnotes. It basically treats numbers like words. So what's the fix? Don't use it for math. You slow it down. You ask it to first reproduce the table exactly
as it sees it in the chat. Once you confirm it, read the numbers right, then you ask it to do the math. Show your work. It's like we're back in fourth grade math class. Always make it show its work. And the final limitation is creativity versus precision. Models just love to paraphrase. It's in their nature. But for legal or compliance work, similar is not good enough. You need the exact words. So you have to explicitly command it. Quote verbatim. Do not paraphrase. You have
to be that blunt. So if we zoom out on all this, it really sounds like the goal isn't to replace the human. It's just changing the human's job description. 100%. The human is no longer the researcher digging in the file cabinet. The human is the auditor, the one checking the citations, assessing the risk, and making the final judgment call. It's a real shift from finding answers to verifying truth. And that shift is everything. It's what separates people who get value from
AI from those who just get noise. Okay, let's recap the big idea. If someone listening has a stack of documents on their desk for tomorrow morning, what is the simple playbook? Three steps. One, pick a reasoning model. Don't use the default. Use GBT 5 .3 or Opus 4 .6 with the reasoning modes on. Two, ground it. Paste in that template. Tell it to only use the file and give it permission to say, I don't know. And three, verify. Use a second AI or ideally notebook LM to cross -check
the key facts. It sounds like a little bit of extra work upfront. It's maybe 30 seconds more per task. But it saves you hours of panic and cleanup later. Trust but verify. An old saying, but it feels more important now than ever before. It's the only way to operate in an age of infinite confident information. So here's the challenge for you listening. The next time you upload a file, don't just ask it to summarize. Try that
grounding template. And just watch how different the output looks when the AI stops trying to guess and starts citing its sources. It gets remarkably quiet. And that quiet, that's the sound of accuracy. I love that. Thank you for deep diving with us today. We'll catch you on the next one.