WEBVTT
00:00:01.679 --> 00:00:05.120
Let's talk about the elephant in the room when it comes to AI.
00:00:05.599 --> 00:00:06.400
That one word.
00:00:07.120 --> 00:00:08.240
Hallucinations.
00:00:08.800 --> 00:00:09.359
Right?
00:00:09.599 --> 00:00:21.440
When you rely on an AI chatbot to provide you something and it straight up lies, or it tells you a very non-truthful version of what you might be looking for.
00:00:22.000 --> 00:00:29.280
And if you don't keep up with the advancements of AI, you might still think that AI lies all the time.
00:00:29.359 --> 00:00:38.320
And well, if you're using the wrong model, if you're not using best practices, hallucinations, even in 2026, can still be a huge problem.
00:00:38.719 --> 00:00:54.479
But here's what not enough people are talking about with the latest technology in today's thinking models and the ability to very easily ground responses in your company's data without much tech know-how.
00:00:55.200 --> 00:00:59.600
Essentially get rid of at least the high rate of hallucinations.
00:00:59.759 --> 00:01:13.760
So that's why we're going to be diving into this sticky topic on the fifth volume of our Start Here series, talking about AI hallucinations, what they are, why they happen, and the right ways to reduce the risk.
00:01:14.079 --> 00:01:16.239
All right, let's get into it.
00:01:16.400 --> 00:01:18.480
I'm excited for today's conversation.
00:01:18.640 --> 00:01:19.840
I hope you are too.
00:01:20.000 --> 00:01:24.640
If you're new here, this is our fifth installment of the Start Here series.
00:01:24.799 --> 00:01:31.040
This is the essential podcast series to both learn the AI basics and to double down on your AI knowledge.
00:01:31.200 --> 00:01:43.439
So whether you are a beginner trying to understand where do I start with AI, or you're really just trying to deepen your expertise, our Start Here series is definitely for you.
00:01:43.599 --> 00:01:48.480
And if you haven't already, please go to starthearseries.com.
00:01:48.640 --> 00:01:49.280
Here's why.
00:01:49.439 --> 00:01:55.920
Well, number one, uh, that's gonna give you access, uh, free access to our inner circle community.
00:01:56.079 --> 00:02:02.640
All right, and you're gonna get straight into uh our Start Here Series uh channel there.
00:02:02.719 --> 00:02:18.159
So you can go back and listen to all of our uh Start Here series uh right there in our free community, but you can also get free access to our uh Prime Prompt and Polish, uh Prime Prompt Polish uh Chat GPT prompt engineering course.
00:02:18.400 --> 00:02:32.400
All right, and if you miss our last episode of Start Here series, that was volume four, we talked about the human AI collaboration and best practices for working alongside AI, which leads us right into today's show.
00:02:32.639 --> 00:02:42.159
Because if you are doing the best practices of human AI collaboration, what we tackled in volume four, then that leads you straight into reducing hallucinations.
00:02:42.400 --> 00:02:50.960
Because if you are working with AI in the right way versus blindly trusting its outputs, well, you're gonna uh be able to fight back against the hallucinations.
00:02:51.039 --> 00:02:53.039
So here's what we're gonna cover on today's show.
00:02:53.199 --> 00:02:58.879
First, we're gonna tell you what hallucinate what hallucinations are uh and why they are there.
00:02:59.039 --> 00:03:07.120
Then we're gonna show you how to assess the risks and talk about how models have improved on hallucinations, but they still run into them.
00:03:07.199 --> 00:03:20.560
And then last but not least, I'm going to suggest to you a kind of four-layer method to reduce the errors and show you how you don't really need to be that scared of hallucinations as long as you are being a smart human.
00:03:20.800 --> 00:03:23.759
All right, let's jump into it and start at the beginning.
00:03:23.919 --> 00:03:27.919
What the heck are hallucinations and why do they still happen?
00:03:28.319 --> 00:03:36.319
Well, in general, here's as as bluntly as I can put it: large language models, right?
00:03:36.400 --> 00:03:42.159
And go back and listen to uh volume one and volume two uh if you need a refresher on how large language models are built.
00:03:42.240 --> 00:03:48.719
But essentially, they scrape the entirety of the internet, online, offline data sets, and then humans train these models.
00:03:48.879 --> 00:04:03.840
So when you and I ask a question, uh right, through this process called reinforcement learning with human feedback, these smart people at you know, OpenAI and Google and Enthropic and all these other labs have trained models that when someone asks about XYZ, here's what you should respond.
00:04:04.000 --> 00:04:10.479
However, for the most part, AI models are very super smart next token prediction.
00:04:10.560 --> 00:04:11.039
Uh, right.
00:04:11.120 --> 00:04:28.079
So sometimes that means if something was maybe incorrect on the internet, or if a model is confused about what you're actually asking, it might present a made-up answer and it might do so very uh confidently, because at their core, AI models are trained to be helpful assistants.
00:04:28.160 --> 00:04:32.800
That is in almost every single system prompt uh that an AI model uses.
00:04:32.879 --> 00:04:41.839
So that's why sometimes they are gonna make things up because they want to be helpful more than anything else, uh, because they predict the next word from patterns.
00:04:42.000 --> 00:04:43.439
And that's kind of why they exist.
00:04:43.600 --> 00:04:50.800
So it's almost like maybe you've heard this uh this saying uh that hallucinations are a feature, not a bug.
00:04:50.879 --> 00:04:58.800
Because at the same time, kind of that same methodology of how they're built help large language models be extremely creative, strategic.
00:04:58.959 --> 00:05:09.360
And even if we're talking about you know, scientific discovery, drug discovery, uh, mathematics, that's actually how they're able to solve new problems that humans haven't been able to.
00:05:09.519 --> 00:05:14.160
Uh, but I think a lot of people, number one, are using the wrong model.
00:05:14.319 --> 00:05:19.680
They're not going through the you know context engineering 101 of how you should be working with a model.
00:05:19.839 --> 00:05:21.040
And they're using, right?
00:05:21.199 --> 00:05:23.279
They're using, like I said, they're using the wrong one.
00:05:23.439 --> 00:05:26.800
That's why hallucinations have been a huge problem.
00:05:26.879 --> 00:05:30.560
And they can get you and your company in a lot of trouble if you're not doing the best practices.
00:05:30.800 --> 00:05:40.720
But uh, here's what's changing models now they can think and they can reason, much like humans do, and that has led to a drastic reduction in hallucinations, right?
00:05:40.879 --> 00:05:45.199
Uh, you're obviously the average person is using way more inference, right?
00:05:45.279 --> 00:05:48.959
They're eating up way more tokens uh to go through basic problems.
00:05:49.120 --> 00:05:53.360
But that's why I think hallucinations are really on the uh decline.
00:05:53.680 --> 00:06:05.839
So the combination of better models that can think and reason like a human do, combined with kind of this rag-esque uh features that front-end large language models give you, uh, are really leading to that reduction.
00:06:06.480 --> 00:06:09.680
But this is how they work, uh, right?
00:06:10.000 --> 00:06:12.160
Uh, because people are always confused.
00:06:12.240 --> 00:06:16.879
Like, hey, if I give a large language model a simple problem, why does it get it wrong?
00:06:17.040 --> 00:06:23.360
Or, you know, a riddle, or asking it how many R's there are in the word strawberry, why does it get it wrong?
00:06:23.519 --> 00:06:30.480
Why is there this, you know, this really jagged, um, you know, almost polarizing output that a large language model can do?
00:06:30.560 --> 00:06:35.519
It can do something that's absolutely genius, but then it can get something absolutely wrong, right?
00:06:35.680 --> 00:06:40.800
So that is, well, because they are next token or next word predictions.
00:06:40.959 --> 00:06:47.519
They are based on patterns trained by humans, they don't verify the truth against reality necessarily.
00:06:47.600 --> 00:06:54.319
And it's that same capabilities that allow AIs to be extremely creative and to generate code and to write poetry.
00:06:54.560 --> 00:06:59.600
Well, at the same time, that can lead them to lie to us, right?
00:06:59.839 --> 00:07:10.800
Um, and by definition, a hallucination is just a confident answer containing either fabricated claims because the model just generates the text, or it can make up sources, right?
00:07:10.959 --> 00:07:12.480
Which is a big problem as well.
00:07:12.560 --> 00:07:19.120
So I'm gonna be talking about uh when we go over our kind of four-layer uh approach, how you can avoid that type of hallucination.
00:07:19.279 --> 00:07:24.000
So uh there's really different categories of hallucinations, right?
00:07:24.160 --> 00:07:32.879
One is it just lies, one is it it's not really lying, uh, but it's sounding overly confident and just giving you some generic information.
00:07:32.959 --> 00:07:36.800
And then other times it can just make up sources, right?
00:07:36.879 --> 00:07:42.079
So it might give you true facts, but it might make up where it got those facts from, right?
00:07:42.240 --> 00:07:51.759
And in most cases, like I said, hallucinations, if I'm being honest, in 2026, they're more human error than AI error.
00:07:51.839 --> 00:07:53.839
And I know a lot of people won't agree with that.
00:07:54.000 --> 00:07:55.120
That's fine.
00:07:55.439 --> 00:08:00.480
But if you get me a room full of humans that are using AI models the right way, right?
00:08:00.639 --> 00:08:11.040
Get uh a room full of humans that have taken hour free, you know, prime prompt polish uh courses, they're gonna run into very few hallucinations, right, compared to the average user.
00:08:11.120 --> 00:08:21.839
Because if you actually know how models work, how to feed them the right data, and how to check and verify on the back end, being a smart uh you know, human user, you're not gonna run into them very much.
00:08:22.160 --> 00:08:26.319
But let's talk about the early days because it was bad, right?
00:08:26.480 --> 00:08:35.279
So go back to the early models, the GPT-3 or the GPT 3.5, uh, you know, the early days of uh Chat GPT and late 2022.
00:08:35.840 --> 00:08:43.679
Uh studies showed that GPT-5 fabricated up to 40% of academic citations, right?
00:08:43.840 --> 00:08:48.960
Fast forward to the next version of GPT-4 that went down to about 29%.
00:08:49.440 --> 00:08:52.799
So uh hallucinations got a little less, but still rampant.
00:08:52.879 --> 00:09:08.559
Uh, fast forward to today, GPT-5.2 reports a 6.2 error rate on general queries, and open AI claims a 30% reduction in errors in GPT-5.2 versus GPT-5.1.
00:09:09.039 --> 00:09:16.720
So we're not talking, you know, of 30% fewer errors between GPT-5.2 and that version from three years ago.
00:09:16.879 --> 00:09:21.759
No, we're talking about in a three-month period, and that jump is huge.
00:09:21.840 --> 00:09:34.720
And I'm gonna explain why uh that jump has occurred and I think why I, you know, in a year from now, we may be not even talking or talking very much about hallucinations.
00:09:35.360 --> 00:09:42.320
And the main reason is just the ability for these labs, right?
00:09:42.399 --> 00:09:57.600
So I'm gonna give an example here from OpenAI, but I think that uh Anthropic and Google and OpenAI have made tremendous strides uh with their models and uh through a lot of different techniques, which I'm not gonna get into the technical side.
00:09:57.679 --> 00:10:02.879
Um, they've made AI models that have much more reliable outputs.
00:10:02.960 --> 00:10:08.000
And one of the reasons um is their ability to handle longer context.
00:10:08.159 --> 00:10:17.039
All right, so for our live stream audience here, I have a screenshot from uh OpenAI's GPT-5.2 um model release.
00:10:17.120 --> 00:10:19.840
And I want to kind of talk about this long context.
00:10:20.000 --> 00:10:25.039
So there is a test, it's essentially called the four needle test.
00:10:25.200 --> 00:10:33.120
So, what this is, they have it um pull out and they ask the model questions and kind of it's it's a needle in the haystack test.
00:10:33.200 --> 00:10:36.480
And they see over a wide range of a conversation, right?
00:10:36.559 --> 00:10:41.759
If you're using uh a model into the hundreds of thousands of tokens, right?
00:10:41.840 --> 00:10:43.759
Like at 256,000 tokens.
00:10:43.919 --> 00:10:56.080
So a very long conversation because what has happened in the past, a lot of these hallucinations come uh when you are hitting a longer point in the context window, right?
00:10:56.559 --> 00:10:58.960
So think of the 3 p.m.
00:10:59.120 --> 00:11:00.159
brain fog, right?
00:11:00.320 --> 00:11:02.080
Let's just say you work nine to five.
00:11:02.320 --> 00:11:06.879
At 3 p.m., you're probably not as sharp as you were at 9:30 a.m.
00:11:06.960 --> 00:11:12.639
Right when that second espresso hits and you're like, let's go, and you're firing all cylinders, right?
00:11:12.799 --> 00:11:18.159
For the most part, that's how large language models had been, I would say, even late into 2025.
00:11:18.240 --> 00:11:22.080
But think of that nine to five instead as a context window, right?
00:11:22.240 --> 00:11:29.279
Because all large language models have their constraints, they have their um kind of confines that you can't break through.
00:11:29.360 --> 00:11:30.879
And one of those is the context window.
00:11:30.960 --> 00:11:36.080
That's how much information a large language model can retain until it starts to forget.
00:11:36.240 --> 00:11:39.919
And potentially when it starts to forget, it will start to hallucinate.
00:11:40.240 --> 00:11:53.919
So even great models like GPT-5.1, as I go here on the results of that four needles test, uh, you saw its ability to properly pull out those facts over a large context window.
00:11:54.080 --> 00:11:58.240
When it started out, right, it was, you know, about 95%.
00:11:58.879 --> 00:12:01.600
It was very good at GPT-5.1 thinking.
00:12:01.759 --> 00:12:10.639
But then toward the end of that context window, it dropped down to like 45%, a 45% uh ability to recall that information.
00:12:10.799 --> 00:12:17.600
Whereas GPT-5.2 thinking, hardly any decline at all, right?
00:12:17.679 --> 00:12:23.519
I believe it was at like a 95 or 96 percent at the end of the context window.
00:12:23.679 --> 00:12:31.840
Okay, so think that very drained human at 3 p.m., they might not be able to recall facts, right?
00:12:32.159 --> 00:12:54.480
But with today's, and when I say today's, I'm saying the latest generation Gemini 3 Pro, Opus 4.5 from Anthropic, Claude Opus 4.5, and GPT-5.2 thinking from OpenAI, their ability to recall information across that context window has greatly improved, which is one of the main reasons why that hallucination rate has gone down, because now models can think.
00:12:54.639 --> 00:13:09.360
They can't only think and plan ahead and reason and call tools on their own to uh provide you more accurate information, but they're able to do that across the entirety of their context window, which is one of the main reasons why hallucinations are going down.
00:13:09.600 --> 00:13:12.399
Here's why this still matters for your business.
00:13:12.639 --> 00:13:15.200
Well, let me cut it to you straight.
00:13:16.080 --> 00:13:28.399
One of the biggest gaps right now, and one of the most blaring uh things that are going wrong with AI implementation across the enterprise is a lack of training and education.
00:13:28.720 --> 00:13:46.399
Because you have companies that I've talked to personally, there's plenty of case studies out there that are rolling out access, whether you're talking about, you know, Microsoft Copilot 365 licenses, uh, you know, ChatGPT Enterprise, Claude Enterprise, Gemini Business, Gemini Enterprise, et cetera.
00:13:46.480 --> 00:14:01.120
They're rolling out AI access, you know, a lot of this was in 2025 to thousands, tens of thousands of employees, but not giving any best practice training or even education, which is why hallucinations are still rampant.
00:14:01.360 --> 00:14:02.960
People don't even understand.
00:14:03.200 --> 00:14:08.960
Oh, I need to uh, you know, click this model selector and choose the best model for the job.
00:14:09.200 --> 00:14:13.600
Or I should be having this model uh, you know, call a certain tool.
00:14:13.759 --> 00:14:14.879
It should be running Python.
00:14:14.960 --> 00:14:16.000
I should be uploading files.
00:14:16.159 --> 00:14:17.440
People don't know the basics.
00:14:17.600 --> 00:14:21.279
Yeah, they're just trying to save as much time as possible with AI.
00:14:21.360 --> 00:14:27.360
And that's led to a lot of hallucinations with a lot of uh high profile and a lot of press around it, right?
00:14:27.440 --> 00:14:38.559
So, uh, as an example, there was uh the AI Hallucination Cases database at HEC Paris, documented 486 different legal cases worldwide involving fabricated AI content.
00:14:38.720 --> 00:14:45.759
Yeah, a lot of these like hallucination stories, they come at the most embarrassing point, which is in the legal sector, right?
00:14:45.840 --> 00:14:51.600
So uh there's been over 128 lawyers that have been cited for filings with hallucinated cases.
00:14:51.759 --> 00:14:56.240
One of the most popular is the, I think it's the MEDA uh versus Aviancha.
00:14:56.559 --> 00:14:58.240
I don't know if I got the pronunciation of that right.
00:14:58.320 --> 00:15:04.159
I might have hallucinated the pronunciation, but that's where you saw uh attorneys sanctioned for six fake citations.
00:15:04.240 --> 00:15:08.399
That was one of the more uh infamous cases of AI illegal early on.
00:15:08.639 --> 00:15:21.840
And even uh Deloitte reportedly refunded part of a$300,000 Australian government contract after it was uh they reportedly found some AI-generated phantom citations in their report.
00:15:22.080 --> 00:15:28.080
So this isn't just people who are, you know, using AI to write blog posts, right?
00:15:28.159 --> 00:15:40.240
A lot of times some of these hallucinations make their way to the forefront in very high, uh, high value and um just extremely visible places, right?
00:15:40.399 --> 00:15:42.960
Like consulting companies, like lawyers.
00:15:43.120 --> 00:15:45.120
We've seen plenty on the financial side as well.
00:15:45.279 --> 00:15:47.679
So that's why this still matters, right?
00:15:47.840 --> 00:15:55.679
Because maybe what I laid out for you, and like saying, hey, if you use the right model and uh look at these, you know, the needle and the haystack, that's improving the context.
00:15:55.759 --> 00:16:02.399
When well, most people aren't using models the right way, and that's why this is still extremely important for your business.
00:16:02.879 --> 00:16:06.000
Don't worry, we're gonna lay it out for you here.
00:16:06.240 --> 00:16:09.120
But I want I want to be very clear.
00:16:09.360 --> 00:16:20.240
Just because I am extremely optimistic about hallucinations decreasing with proper human involvement and training, it doesn't mean they're going away, right?
00:16:20.399 --> 00:16:26.960
The very nature of what a large language model is, how it works, means that hallucinations will be there for a while, uh, right?
00:16:27.279 --> 00:16:35.919
Uh, because the models are always going to optimize for what word or what sets of words could be next, even if those things are incorrect.
00:16:36.080 --> 00:16:40.720
Think of how many times that you've been on the internet researching something and you're like, ah, this isn't right.
00:16:40.879 --> 00:16:42.480
That's not right, right?
00:16:42.639 --> 00:16:47.519
In the same way, well, large language models take all that information from the web.
00:16:47.679 --> 00:16:59.360
So let's say you are a domain expert and there's something, you know, uh, that in your field people constantly get wrong, or there's some uh some pieces of information out there that exist that are circulating that aren't exactly true.
00:16:59.600 --> 00:17:09.119
Well, in the same way that you might find incorrect data on websites, or you might go to a conference or hear colleagues speak about something that you're an expert in, and you're like, nope, that's not right.
00:17:09.279 --> 00:17:18.640
Well, a lot of times large language models are reflecting inaccuracies uh or uh, you know, half lies or half truths that exist in the real world.
00:17:18.880 --> 00:17:31.039
Second, uh, the second reason why hallucinations probably are gonna go away is well, when their needed information isn't there by default, by default, large language models are gonna fill in the gap, right?
00:17:31.119 --> 00:17:38.160
They're gonna do everything they can to be a helpful assistant because that is the default behavior, unless you change it and you should, and I'll tell you how.
00:17:38.319 --> 00:17:44.559
And then third, the chat inference, the chat interfaces rewards that fast, confident answers, right?
00:17:44.720 --> 00:17:57.759
I'm not gonna get too much into uh you know reinforcement learning with human feedback and uh kind of scaling laws and inference and all that, but for the most part, chatbots are trained to quickly give you an answer and being token efficient, right?
00:17:57.920 --> 00:18:05.359
So if it thinks, oh, I can spit out an answer rather quickly, it may just do that unless you've told it not to.
00:18:05.759 --> 00:18:06.079
Right.
00:18:06.240 --> 00:18:14.319
So uh they'll at times will reward kind of uh giving you an answer even if it's not confident versus just saying, I don't know.
00:18:14.480 --> 00:18:26.880
I will say that's models more of late 2025, because the models from 2026, you'll see by default, again, using the right models uh in the right context, they are gonna say more.
00:18:27.039 --> 00:18:29.039
Uh, and I'd love to hear if you've seen this more.
00:18:29.200 --> 00:18:33.599
I have models that actually say, I don't know, or I'm not sure, right?
00:18:33.680 --> 00:18:39.279
Even without giving uh you know anything extra in your prompt, anything in the special instructions.
00:18:39.440 --> 00:18:44.960
Today's models are more likely to say, Hey, I'm not, I'm not sure, which is a good thing.
00:18:45.359 --> 00:18:50.799
So here's how to spot and cut down on hallucinations, a quick four-step plan.
00:18:51.039 --> 00:18:51.599
All right.
00:18:51.920 --> 00:18:56.720
Number one, you need to change the model's behavior.
00:18:56.799 --> 00:19:11.119
Uh, like I said, I do think that um, I think that in the future, the AI, the AI labs are gonna find that sweet spot between uh instructing models to be helpful assistance, but not too helpful and not fabricating things.
00:19:11.200 --> 00:19:14.400
But in the short run, you can do that, right?
00:19:14.559 --> 00:19:22.640
So whether it's in your prompting sequence when you are working with a large language model, or what I would recommend is setting custom instructions, right?
00:19:22.799 --> 00:19:26.480
So uh there's different kinds of places that you can set custom instructions.
00:19:26.640 --> 00:19:48.960
But uh, if you are non-technical, if you're not really sure what that means, that's in there's settings in most of the big providers that you can essentially put in your own rules and any chat that you use, any response uh that you are uh using the model, it will always go through and read and usually adhere to the own set of custom instructions that you put in there.
00:19:49.119 --> 00:19:56.960
So, you know, even putting something simple, like if you're not certain or the information isn't provided, say I don't know rather than guessing.
00:19:57.200 --> 00:19:58.319
Something simple as that.
00:19:58.400 --> 00:20:05.200
I I have a set of uh custom instructions uh that I've done a lot of testing with over the years that are much more robust.
00:20:05.279 --> 00:20:11.759
But even putting something simple like that, like hey, if you're not 100% sure, say you're not sure, right?
00:20:12.160 --> 00:20:18.559
Or to require every factual claim to include a source, or you know, labeling everything afterwards, right?
00:20:18.720 --> 00:20:24.799
Saying, hey, this is what I'm 100%, you know, telling the model, this is what I'm 100% sure on, this is what I'm not sure on.
00:20:24.880 --> 00:20:30.880
You can even do something like uh in each um response, having it give you giving you a confidence score.
00:20:30.960 --> 00:20:33.359
That's another great thing that I that you could do.
00:20:33.599 --> 00:20:37.759
And then have it to separate facts from inferences, right?
00:20:37.839 --> 00:20:45.200
Because here's the thing, especially if you're using it as a strategy, creative partner, brainstorming, not all those things are black and white, right?
00:20:45.359 --> 00:20:50.319
Maybe 99% of what you might use a large language model for is in that gray area, right?
00:20:50.400 --> 00:20:54.079
It's strategies, creative, right?
00:20:54.240 --> 00:20:54.960
Being creative.
00:20:55.119 --> 00:20:57.680
So you have it separate facts from inferences.
00:20:57.759 --> 00:21:03.279
You can ask for a table with columns for confirmed fact versus assumption in a structured output.
00:21:03.440 --> 00:21:08.240
So number one is doing that either in your prompt or in your custom instructions.
00:21:08.480 --> 00:21:17.039
Number two, my gosh, the fact that we have this and you don't have to pay any extra, is wild.
00:21:17.200 --> 00:21:17.920
Uh, right.
00:21:18.799 --> 00:21:21.359
Make it retrieve the information, right?
00:21:21.680 --> 00:21:27.200
This whole context engineering, something I've been teaching since 2023 before it was a thing.
00:21:27.279 --> 00:21:27.920
Uh, right.
00:21:28.000 --> 00:21:35.359
If if you've taken our free prime prompt polish course, this is the refine cue in the priming.
00:21:35.440 --> 00:21:43.839
So uh we call it the fetch in the insights, but essentially using a smaller simplified version of RAG, retrieval meta generation.
00:21:44.000 --> 00:21:50.720
Now you have a simplified version of RAG available with a couple of clicks by connecting your company's data, right?
00:21:50.880 --> 00:21:59.759
So both in obviously Microsoft 365 Copilot, uh in uh Claude, in Google Gemini, in OpenAI's chat, GPT.
00:22:00.240 --> 00:22:04.160
They have different ways for you to connect your business data in a few clicks.
00:22:04.240 --> 00:22:09.599
So you can essentially uh ground it in a way you can ground it in a way you kind of can't.
00:22:09.680 --> 00:22:18.400
But the combination of number one, those kind of custom instructions, and number two, first putting your company's information.
00:22:18.559 --> 00:22:26.079
If you combine those two things and you say, always check, you know, this ABC document uh before you respond.
00:22:26.160 --> 00:22:31.279
And then if you make sure you check and verify it that it did, I mean, those two things right there, amazing.
00:22:31.440 --> 00:22:41.039
But it doesn't matter if you're, you know, a Microsoft organization, you can connect your, you know, your OneDrive and SharePoint data uh in OpenAI's products, right?
00:22:41.119 --> 00:22:50.960
If you're using uh, you know, Microsoft 365 Copilot, like online, their online version, you can connect your if you use Google, if you use Google Drive, right?
00:22:51.119 --> 00:22:59.839
So being able to connect your company's data to a large language model and then instructing it to always um look at those things first is huge and it makes a big difference.
00:22:59.920 --> 00:23:11.759
So there was a 2024 Stanford study that found RAG combined with reinforcement learning with human feedback and guardrails, achieved 96% hallucination reduction versus the baseline, right?
00:23:12.319 --> 00:23:15.119
Just doing those basic things, right?
00:23:15.440 --> 00:23:25.359
Like having a version of your company's data and proper uh instructions are going to cut down on hallucinations, just those two things alone.
00:23:25.680 --> 00:23:32.559
And then steps three and four, kind of combined into one, is just the verification workflows and agent safety.
00:23:32.640 --> 00:23:34.880
So it's kind of like three and three B.
00:23:35.039 --> 00:23:35.359
All right.
00:23:35.440 --> 00:23:36.319
So what is that?
00:23:36.400 --> 00:23:39.200
Well, you have to be able to catch errors before they escape.
00:23:39.359 --> 00:23:45.680
This is the uh the expert-driven loops that I always talk about, not the lazy, passive human in the loop.
00:23:45.839 --> 00:23:49.680
I'm talking about the active, proactive, expert-driven loops.
00:23:49.839 --> 00:23:52.960
So, as these models, they think by default.
00:23:53.039 --> 00:23:55.519
You know, I say GPT-5.2, uh, right.
00:23:55.599 --> 00:24:01.680
And if you're listening to this uh this episode in July, right, maybe it's GPT-5.3 or Gemini 3.5.
00:24:01.759 --> 00:24:02.160
I don't know.
00:24:02.319 --> 00:24:05.440
But today's latest models are agentic by nature.
00:24:05.599 --> 00:24:08.960
They're gonna go through, they're gonna think, they're gonna decide how much to think.
00:24:09.119 --> 00:24:11.119
Should I spend five minutes on this answer?
00:24:11.279 --> 00:24:13.039
Should I spend 30 seconds, right?
00:24:13.200 --> 00:24:14.880
But you have control over that.
00:24:14.960 --> 00:24:24.720
Uh, and you can always kind of build, and I always recommend doing something like this build a second pass review where that model's only job is checking claims.
00:24:24.880 --> 00:24:29.359
That's why I also do kind of a mixture of models set up.
00:24:29.839 --> 00:24:39.359
But I have systems set up where you know, I maybe have a Google Gemini Deep Research run, and then I'll take some of those results and I will verify it.
00:24:39.440 --> 00:24:40.799
And I have something set up.
00:24:40.960 --> 00:24:44.640
Uh, the model that I like doing this for is GPT-5.2 Pro.
00:24:45.119 --> 00:24:51.920
Um, I will then use GPT-5.2 Pro only to go through and verify every single thing that I get from a different model.
00:24:52.079 --> 00:24:55.839
That doesn't mean I don't trust model A versus model B, right?
00:24:55.920 --> 00:24:57.759
And sometimes I'll flip-flop it.
00:24:57.920 --> 00:25:00.079
It means you should always be doing this, right?
00:25:00.240 --> 00:25:03.759
Especially for high value, highly visible projects, right?
00:25:03.920 --> 00:25:09.680
If you're just trying to see, like, you know, what's the weather next week or, you know, something topical, right?
00:25:09.759 --> 00:25:14.240
Like, what's the best, you know, what are the three best softwares for tackling this issue?
00:25:14.640 --> 00:25:17.119
I don't think you necessarily need to go through all those steps.
00:25:17.279 --> 00:25:25.359
But if this, if you are using an AI model as part of a high value workflow, you should definitely be doing this second pass review.
00:25:25.440 --> 00:25:29.359
Uh, and then again, requiring even that second pass to show the sources.
00:25:29.519 --> 00:25:35.440
And then more than anything, you need to be tracing and observing how these models are getting to those conclusions.
00:25:35.599 --> 00:25:42.240
And what that means is well, you should be using a thinking model and then reading the summarized or the chain of thought, right?
00:25:42.319 --> 00:25:50.480
So, what that is, uh most of the models, you know, different models give you different level of uh visibility on what they're actually doing under the hood.
00:25:50.640 --> 00:25:58.799
But just if you were to hand off an important assignment to a brand new employee, hopefully you wouldn't just hand that off to the client, right?
00:25:58.960 --> 00:26:04.079
You would take a step in between and be like, hey, new employee, tell me how you got to this answer, right?
00:26:04.240 --> 00:26:07.119
Let's say they spent an entire week on this big project.
00:26:07.279 --> 00:26:13.039
I would hope you would take at least an hour or so to sit down with that new employee and be like, okay, how do we get to this conclusion?
00:26:13.200 --> 00:26:14.240
Walk me through, right?
00:26:14.319 --> 00:26:18.079
So the good thing with these agentic models by default, you can see all that.
00:26:18.240 --> 00:26:23.680
And you, smart human, need to be able to go through and check and see what it did at each and every step.
00:26:23.920 --> 00:26:30.400
Did it follow your instructions of, you know, uh, here's here's what's factual, here's what I inferred.
00:26:30.559 --> 00:26:33.920
Did it make sure to look at the right documents at the right time?
00:26:34.079 --> 00:26:37.039
You can go through and sequentially check all of those things.
00:26:37.119 --> 00:26:40.319
And if it didn't do it, then you can course correct.
00:26:40.400 --> 00:26:44.960
But that is layers three and four, and especially as AI can take action, right?
00:26:45.119 --> 00:26:51.920
As we're talking about agentic AI, that's when you really need to treat it like a junior employee.
00:26:52.160 --> 00:26:54.640
But no, they can still make things up.
00:26:54.720 --> 00:27:15.119
And like I said, most of the time, if you are doing the proper context engineering one-on-one, if you're using the right model, and then if you're going through kind of layers three and four right here, with the second pass in the observability and traceability of going through that chain of thought, hallucinations aren't gonna be the biggest problem for you, right?
00:27:15.279 --> 00:27:23.119
You still need to have those expert-driven loops because as you're looking at that chain of thought, as you're providing data on the front end, you have to do that at an expert level.
00:27:23.200 --> 00:27:25.359
You can't just say, here's a thousand files.
00:27:25.519 --> 00:27:26.480
Good luck, agent.
00:27:26.640 --> 00:27:28.799
No, you need to, you know, point them in the right direction.
00:27:28.960 --> 00:27:31.279
The same thing, checking responses on the back end.
00:27:31.359 --> 00:27:33.440
You have to know what all of that means.
00:27:33.680 --> 00:27:44.000
But if you have expert-driven loops and if you go through those four steps that I talked about, I think hallucinations are no longer going to be the big elephant in the room.
00:27:44.160 --> 00:27:46.079
You're gonna be able to reduce them.
00:27:46.480 --> 00:27:47.119
All right.
00:27:47.440 --> 00:27:50.000
But remember, it's not going away.
00:27:50.240 --> 00:27:56.400
Hallucinations are a property that you need to manage, not something that you hope will get fixed one day.
00:27:56.559 --> 00:28:09.599
And it's not that the winners right now are picking the best models, they're just building those four steps of verification into every single uh episode or into every single piece of work that they're doing.
00:28:09.839 --> 00:28:10.160
All right.
00:28:10.319 --> 00:28:13.680
Speaking of episodes, I hope this episode was helpful.
00:28:13.759 --> 00:28:22.240
Uh, because the start here series, I want you to know all the details, right?
00:28:22.400 --> 00:28:30.559
Um, as someone that's done this everyday AI thing now more than 700 times, I've got to speak with some of the smartest people in the world.
00:28:30.720 --> 00:28:40.640
I've realized if you're educated, if you're trained, if you understand how these models work, and if you keep up, that's that's the big caveat there, right?
00:28:40.799 --> 00:28:45.279
If you keep up, you don't have to be as worried about hallucinations.
00:28:45.359 --> 00:28:46.799
I'm not saying you can write them off.
00:28:46.960 --> 00:28:47.920
You shouldn't.
00:28:48.160 --> 00:29:06.319
But if you're going through the right steps and what we went over in today's episode, you are definitely able to reduce the risk and to get more out of large language models because that's what we're all about here at EverydayAI, cutting through the fluff, giving you the facts and the right information to grow your company and your career.
00:29:06.480 --> 00:29:10.319
So I hope volume five of the Start Here series was helpful.
00:29:10.480 --> 00:29:15.519
I hope you know more about hallucinations now, why they happen, and how you can reduce the risk.
00:29:15.680 --> 00:29:20.079
So if this was helpful, please go to starthearseries.com.
00:29:20.240 --> 00:29:34.480
That's gonna give you free access to our inner circle community, uh, the prompt engineering course, as well as an easy spot to go listen and catch up and engage with others who are going through this Start Here series with you.
00:29:34.640 --> 00:29:36.160
So thank you for tuning in.
00:29:36.319 --> 00:29:39.279
I hope to see you back later for more everyday AI.
00:29:39.359 --> 00:29:40.559
Thanks, y'all.
00:00:01.679 --> 00:00:05.120
Let's talk about the elephant in the room when it comes to AI.
00:00:05.599 --> 00:00:06.400
That one word.
00:00:07.120 --> 00:00:08.240
Hallucinations.
00:00:08.800 --> 00:00:09.359
Right?
00:00:09.599 --> 00:00:21.440
When you rely on an AI chatbot to provide you something and it straight up lies, or it tells you a very non-truthful version of what you might be looking for.
00:00:22.000 --> 00:00:29.280
And if you don't keep up with the advancements of AI, you might still think that AI lies all the time.
00:00:29.359 --> 00:00:38.320
And well, if you're using the wrong model, if you're not using best practices, hallucinations, even in 2026, can still be a huge problem.
00:00:38.719 --> 00:00:54.479
But here's what not enough people are talking about with the latest technology in today's thinking models and the ability to very easily ground responses in your company's data without much tech know-how.
00:00:55.200 --> 00:00:59.600
Essentially get rid of at least the high rate of hallucinations.
00:00:59.759 --> 00:01:13.760
So that's why we're going to be diving into this sticky topic on the fifth volume of our Start Here series, talking about AI hallucinations, what they are, why they happen, and the right ways to reduce the risk.
00:01:14.079 --> 00:01:16.239
All right, let's get into it.
00:01:16.400 --> 00:01:18.480
I'm excited for today's conversation.
00:01:18.640 --> 00:01:19.840
I hope you are too.
00:01:20.000 --> 00:01:24.640
If you're new here, this is our fifth installment of the Start Here series.
00:01:24.799 --> 00:01:31.040
This is the essential podcast series to both learn the AI basics and to double down on your AI knowledge.
00:01:31.200 --> 00:01:43.439
So whether you are a beginner trying to understand where do I start with AI, or you're really just trying to deepen your expertise, our Start Here series is definitely for you.
00:01:43.599 --> 00:01:48.480
And if you haven't already, please go to starthearseries.com.
00:01:48.640 --> 00:01:49.280
Here's why.
00:01:49.439 --> 00:01:55.920
Well, number one, uh, that's gonna give you access, uh, free access to our inner circle community.
00:01:56.079 --> 00:02:02.640
All right, and you're gonna get straight into uh our Start Here Series uh channel there.
00:02:02.719 --> 00:02:18.159
So you can go back and listen to all of our uh Start Here series uh right there in our free community, but you can also get free access to our uh Prime Prompt and Polish, uh Prime Prompt Polish uh Chat GPT prompt engineering course.
00:02:18.400 --> 00:02:32.400
All right, and if you miss our last episode of Start Here series, that was volume four, we talked about the human AI collaboration and best practices for working alongside AI, which leads us right into today's show.
00:02:32.639 --> 00:02:42.159
Because if you are doing the best practices of human AI collaboration, what we tackled in volume four, then that leads you straight into reducing hallucinations.
00:02:42.400 --> 00:02:50.960
Because if you are working with AI in the right way versus blindly trusting its outputs, well, you're gonna uh be able to fight back against the hallucinations.
00:02:51.039 --> 00:02:53.039
So here's what we're gonna cover on today's show.
00:02:53.199 --> 00:02:58.879
First, we're gonna tell you what hallucinate what hallucinations are uh and why they are there.
00:02:59.039 --> 00:03:07.120
Then we're gonna show you how to assess the risks and talk about how models have improved on hallucinations, but they still run into them.
00:03:07.199 --> 00:03:20.560
And then last but not least, I'm going to suggest to you a kind of four-layer method to reduce the errors and show you how you don't really need to be that scared of hallucinations as long as you are being a smart human.
00:03:20.800 --> 00:03:23.759
All right, let's jump into it and start at the beginning.
00:03:23.919 --> 00:03:27.919
What the heck are hallucinations and why do they still happen?
00:03:28.319 --> 00:03:36.319
Well, in general, here's as as bluntly as I can put it: large language models, right?
00:03:36.400 --> 00:03:42.159
And go back and listen to uh volume one and volume two uh if you need a refresher on how large language models are built.
00:03:42.240 --> 00:03:48.719
But essentially, they scrape the entirety of the internet, online, offline data sets, and then humans train these models.
00:03:48.879 --> 00:04:03.840
So when you and I ask a question, uh right, through this process called reinforcement learning with human feedback, these smart people at you know, OpenAI and Google and Enthropic and all these other labs have trained models that when someone asks about XYZ, here's what you should respond.
00:04:04.000 --> 00:04:10.479
However, for the most part, AI models are very super smart next token prediction.
00:04:10.560 --> 00:04:11.039
Uh, right.
00:04:11.120 --> 00:04:28.079
So sometimes that means if something was maybe incorrect on the internet, or if a model is confused about what you're actually asking, it might present a made-up answer and it might do so very uh confidently, because at their core, AI models are trained to be helpful assistants.
00:04:28.160 --> 00:04:32.800
That is in almost every single system prompt uh that an AI model uses.
00:04:32.879 --> 00:04:41.839
So that's why sometimes they are gonna make things up because they want to be helpful more than anything else, uh, because they predict the next word from patterns.
00:04:42.000 --> 00:04:43.439
And that's kind of why they exist.
00:04:43.600 --> 00:04:50.800
So it's almost like maybe you've heard this uh this saying uh that hallucinations are a feature, not a bug.
00:04:50.879 --> 00:04:58.800
Because at the same time, kind of that same methodology of how they're built help large language models be extremely creative, strategic.
00:04:58.959 --> 00:05:09.360
And even if we're talking about you know, scientific discovery, drug discovery, uh, mathematics, that's actually how they're able to solve new problems that humans haven't been able to.
00:05:09.519 --> 00:05:14.160
Uh, but I think a lot of people, number one, are using the wrong model.
00:05:14.319 --> 00:05:19.680
They're not going through the you know context engineering 101 of how you should be working with a model.
00:05:19.839 --> 00:05:21.040
And they're using, right?
00:05:21.199 --> 00:05:23.279
They're using, like I said, they're using the wrong one.
00:05:23.439 --> 00:05:26.800
That's why hallucinations have been a huge problem.
00:05:26.879 --> 00:05:30.560
And they can get you and your company in a lot of trouble if you're not doing the best practices.
00:05:30.800 --> 00:05:40.720
But uh, here's what's changing models now they can think and they can reason, much like humans do, and that has led to a drastic reduction in hallucinations, right?
00:05:40.879 --> 00:05:45.199
Uh, you're obviously the average person is using way more inference, right?
00:05:45.279 --> 00:05:48.959
They're eating up way more tokens uh to go through basic problems.
00:05:49.120 --> 00:05:53.360
But that's why I think hallucinations are really on the uh decline.
00:05:53.680 --> 00:06:05.839
So the combination of better models that can think and reason like a human do, combined with kind of this rag-esque uh features that front-end large language models give you, uh, are really leading to that reduction.
00:06:06.480 --> 00:06:09.680
But this is how they work, uh, right?
00:06:10.000 --> 00:06:12.160
Uh, because people are always confused.
00:06:12.240 --> 00:06:16.879
Like, hey, if I give a large language model a simple problem, why does it get it wrong?
00:06:17.040 --> 00:06:23.360
Or, you know, a riddle, or asking it how many R's there are in the word strawberry, why does it get it wrong?
00:06:23.519 --> 00:06:30.480
Why is there this, you know, this really jagged, um, you know, almost polarizing output that a large language model can do?
00:06:30.560 --> 00:06:35.519
It can do something that's absolutely genius, but then it can get something absolutely wrong, right?
00:06:35.680 --> 00:06:40.800
So that is, well, because they are next token or next word predictions.
00:06:40.959 --> 00:06:47.519
They are based on patterns trained by humans, they don't verify the truth against reality necessarily.
00:06:47.600 --> 00:06:54.319
And it's that same capabilities that allow AIs to be extremely creative and to generate code and to write poetry.
00:06:54.560 --> 00:06:59.600
Well, at the same time, that can lead them to lie to us, right?
00:06:59.839 --> 00:07:10.800
Um, and by definition, a hallucination is just a confident answer containing either fabricated claims because the model just generates the text, or it can make up sources, right?
00:07:10.959 --> 00:07:12.480
Which is a big problem as well.
00:07:12.560 --> 00:07:19.120
So I'm gonna be talking about uh when we go over our kind of four-layer uh approach, how you can avoid that type of hallucination.
00:07:19.279 --> 00:07:24.000
So uh there's really different categories of hallucinations, right?
00:07:24.160 --> 00:07:32.879
One is it just lies, one is it it's not really lying, uh, but it's sounding overly confident and just giving you some generic information.
00:07:32.959 --> 00:07:36.800
And then other times it can just make up sources, right?
00:07:36.879 --> 00:07:42.079
So it might give you true facts, but it might make up where it got those facts from, right?
00:07:42.240 --> 00:07:51.759
And in most cases, like I said, hallucinations, if I'm being honest, in 2026, they're more human error than AI error.
00:07:51.839 --> 00:07:53.839
And I know a lot of people won't agree with that.
00:07:54.000 --> 00:07:55.120
That's fine.
00:07:55.439 --> 00:08:00.480
But if you get me a room full of humans that are using AI models the right way, right?
00:08:00.639 --> 00:08:11.040
Get uh a room full of humans that have taken hour free, you know, prime prompt polish uh courses, they're gonna run into very few hallucinations, right, compared to the average user.
00:08:11.120 --> 00:08:21.839
Because if you actually know how models work, how to feed them the right data, and how to check and verify on the back end, being a smart uh you know, human user, you're not gonna run into them very much.
00:08:22.160 --> 00:08:26.319
But let's talk about the early days because it was bad, right?
00:08:26.480 --> 00:08:35.279
So go back to the early models, the GPT-3 or the GPT 3.5, uh, you know, the early days of uh Chat GPT and late 2022.
00:08:35.840 --> 00:08:43.679
Uh studies showed that GPT-5 fabricated up to 40% of academic citations, right?
00:08:43.840 --> 00:08:48.960
Fast forward to the next version of GPT-4 that went down to about 29%.
00:08:49.440 --> 00:08:52.799
So uh hallucinations got a little less, but still rampant.
00:08:52.879 --> 00:09:08.559
Uh, fast forward to today, GPT-5.2 reports a 6.2 error rate on general queries, and open AI claims a 30% reduction in errors in GPT-5.2 versus GPT-5.1.
00:09:09.039 --> 00:09:16.720
So we're not talking, you know, of 30% fewer errors between GPT-5.2 and that version from three years ago.
00:09:16.879 --> 00:09:21.759
No, we're talking about in a three-month period, and that jump is huge.
00:09:21.840 --> 00:09:34.720
And I'm gonna explain why uh that jump has occurred and I think why I, you know, in a year from now, we may be not even talking or talking very much about hallucinations.
00:09:35.360 --> 00:09:42.320
And the main reason is just the ability for these labs, right?
00:09:42.399 --> 00:09:57.600
So I'm gonna give an example here from OpenAI, but I think that uh Anthropic and Google and OpenAI have made tremendous strides uh with their models and uh through a lot of different techniques, which I'm not gonna get into the technical side.
00:09:57.679 --> 00:10:02.879
Um, they've made AI models that have much more reliable outputs.
00:10:02.960 --> 00:10:08.000
And one of the reasons um is their ability to handle longer context.
00:10:08.159 --> 00:10:17.039
All right, so for our live stream audience here, I have a screenshot from uh OpenAI's GPT-5.2 um model release.
00:10:17.120 --> 00:10:19.840
And I want to kind of talk about this long context.
00:10:20.000 --> 00:10:25.039
So there is a test, it's essentially called the four needle test.
00:10:25.200 --> 00:10:33.120
So, what this is, they have it um pull out and they ask the model questions and kind of it's it's a needle in the haystack test.
00:10:33.200 --> 00:10:36.480
And they see over a wide range of a conversation, right?
00:10:36.559 --> 00:10:41.759
If you're using uh a model into the hundreds of thousands of tokens, right?
00:10:41.840 --> 00:10:43.759
Like at 256,000 tokens.
00:10:43.919 --> 00:10:56.080
So a very long conversation because what has happened in the past, a lot of these hallucinations come uh when you are hitting a longer point in the context window, right?
00:10:56.559 --> 00:10:58.960
So think of the 3 p.m.
00:10:59.120 --> 00:11:00.159
brain fog, right?
00:11:00.320 --> 00:11:02.080
Let's just say you work nine to five.
00:11:02.320 --> 00:11:06.879
At 3 p.m., you're probably not as sharp as you were at 9:30 a.m.
00:11:06.960 --> 00:11:12.639
Right when that second espresso hits and you're like, let's go, and you're firing all cylinders, right?
00:11:12.799 --> 00:11:18.159
For the most part, that's how large language models had been, I would say, even late into 2025.
00:11:18.240 --> 00:11:22.080
But think of that nine to five instead as a context window, right?
00:11:22.240 --> 00:11:29.279
Because all large language models have their constraints, they have their um kind of confines that you can't break through.
00:11:29.360 --> 00:11:30.879
And one of those is the context window.
00:11:30.960 --> 00:11:36.080
That's how much information a large language model can retain until it starts to forget.
00:11:36.240 --> 00:11:39.919
And potentially when it starts to forget, it will start to hallucinate.
00:11:40.240 --> 00:11:53.919
So even great models like GPT-5.1, as I go here on the results of that four needles test, uh, you saw its ability to properly pull out those facts over a large context window.
00:11:54.080 --> 00:11:58.240
When it started out, right, it was, you know, about 95%.
00:11:58.879 --> 00:12:01.600
It was very good at GPT-5.1 thinking.
00:12:01.759 --> 00:12:10.639
But then toward the end of that context window, it dropped down to like 45%, a 45% uh ability to recall that information.
00:12:10.799 --> 00:12:17.600
Whereas GPT-5.2 thinking, hardly any decline at all, right?
00:12:17.679 --> 00:12:23.519
I believe it was at like a 95 or 96 percent at the end of the context window.
00:12:23.679 --> 00:12:31.840
Okay, so think that very drained human at 3 p.m., they might not be able to recall facts, right?
00:12:32.159 --> 00:12:54.480
But with today's, and when I say today's, I'm saying the latest generation Gemini 3 Pro, Opus 4.5 from Anthropic, Claude Opus 4.5, and GPT-5.2 thinking from OpenAI, their ability to recall information across that context window has greatly improved, which is one of the main reasons why that hallucination rate has gone down, because now models can think.
00:12:54.639 --> 00:13:09.360
They can't only think and plan ahead and reason and call tools on their own to uh provide you more accurate information, but they're able to do that across the entirety of their context window, which is one of the main reasons why hallucinations are going down.
00:13:09.600 --> 00:13:12.399
Here's why this still matters for your business.
00:13:12.639 --> 00:13:15.200
Well, let me cut it to you straight.
00:13:16.080 --> 00:13:28.399
One of the biggest gaps right now, and one of the most blaring uh things that are going wrong with AI implementation across the enterprise is a lack of training and education.
00:13:28.720 --> 00:13:46.399
Because you have companies that I've talked to personally, there's plenty of case studies out there that are rolling out access, whether you're talking about, you know, Microsoft Copilot 365 licenses, uh, you know, ChatGPT Enterprise, Claude Enterprise, Gemini Business, Gemini Enterprise, et cetera.
00:13:46.480 --> 00:14:01.120
They're rolling out AI access, you know, a lot of this was in 2025 to thousands, tens of thousands of employees, but not giving any best practice training or even education, which is why hallucinations are still rampant.
00:14:01.360 --> 00:14:02.960
People don't even understand.
00:14:03.200 --> 00:14:08.960
Oh, I need to uh, you know, click this model selector and choose the best model for the job.
00:14:09.200 --> 00:14:13.600
Or I should be having this model uh, you know, call a certain tool.
00:14:13.759 --> 00:14:14.879
It should be running Python.
00:14:14.960 --> 00:14:16.000
I should be uploading files.
00:14:16.159 --> 00:14:17.440
People don't know the basics.
00:14:17.600 --> 00:14:21.279
Yeah, they're just trying to save as much time as possible with AI.
00:14:21.360 --> 00:14:27.360
And that's led to a lot of hallucinations with a lot of uh high profile and a lot of press around it, right?
00:14:27.440 --> 00:14:38.559
So, uh, as an example, there was uh the AI Hallucination Cases database at HEC Paris, documented 486 different legal cases worldwide involving fabricated AI content.
00:14:38.720 --> 00:14:45.759
Yeah, a lot of these like hallucination stories, they come at the most embarrassing point, which is in the legal sector, right?
00:14:45.840 --> 00:14:51.600
So uh there's been over 128 lawyers that have been cited for filings with hallucinated cases.
00:14:51.759 --> 00:14:56.240
One of the most popular is the, I think it's the MEDA uh versus Aviancha.
00:14:56.559 --> 00:14:58.240
I don't know if I got the pronunciation of that right.
00:14:58.320 --> 00:15:04.159
I might have hallucinated the pronunciation, but that's where you saw uh attorneys sanctioned for six fake citations.
00:15:04.240 --> 00:15:08.399
That was one of the more uh infamous cases of AI illegal early on.
00:15:08.639 --> 00:15:21.840
And even uh Deloitte reportedly refunded part of a$300,000 Australian government contract after it was uh they reportedly found some AI-generated phantom citations in their report.
00:15:22.080 --> 00:15:28.080
So this isn't just people who are, you know, using AI to write blog posts, right?
00:15:28.159 --> 00:15:40.240
A lot of times some of these hallucinations make their way to the forefront in very high, uh, high value and um just extremely visible places, right?
00:15:40.399 --> 00:15:42.960
Like consulting companies, like lawyers.
00:15:43.120 --> 00:15:45.120
We've seen plenty on the financial side as well.
00:15:45.279 --> 00:15:47.679
So that's why this still matters, right?
00:15:47.840 --> 00:15:55.679
Because maybe what I laid out for you, and like saying, hey, if you use the right model and uh look at these, you know, the needle and the haystack, that's improving the context.
00:15:55.759 --> 00:16:02.399
When well, most people aren't using models the right way, and that's why this is still extremely important for your business.
00:16:02.879 --> 00:16:06.000
Don't worry, we're gonna lay it out for you here.
00:16:06.240 --> 00:16:09.120
But I want I want to be very clear.
00:16:09.360 --> 00:16:20.240
Just because I am extremely optimistic about hallucinations decreasing with proper human involvement and training, it doesn't mean they're going away, right?
00:16:20.399 --> 00:16:26.960
The very nature of what a large language model is, how it works, means that hallucinations will be there for a while, uh, right?
00:16:27.279 --> 00:16:35.919
Uh, because the models are always going to optimize for what word or what sets of words could be next, even if those things are incorrect.
00:16:36.080 --> 00:16:40.720
Think of how many times that you've been on the internet researching something and you're like, ah, this isn't right.
00:16:40.879 --> 00:16:42.480
That's not right, right?
00:16:42.639 --> 00:16:47.519
In the same way, well, large language models take all that information from the web.
00:16:47.679 --> 00:16:59.360
So let's say you are a domain expert and there's something, you know, uh, that in your field people constantly get wrong, or there's some uh some pieces of information out there that exist that are circulating that aren't exactly true.
00:16:59.600 --> 00:17:09.119
Well, in the same way that you might find incorrect data on websites, or you might go to a conference or hear colleagues speak about something that you're an expert in, and you're like, nope, that's not right.
00:17:09.279 --> 00:17:18.640
Well, a lot of times large language models are reflecting inaccuracies uh or uh, you know, half lies or half truths that exist in the real world.
00:17:18.880 --> 00:17:31.039
Second, uh, the second reason why hallucinations probably are gonna go away is well, when their needed information isn't there by default, by default, large language models are gonna fill in the gap, right?
00:17:31.119 --> 00:17:38.160
They're gonna do everything they can to be a helpful assistant because that is the default behavior, unless you change it and you should, and I'll tell you how.
00:17:38.319 --> 00:17:44.559
And then third, the chat inference, the chat interfaces rewards that fast, confident answers, right?
00:17:44.720 --> 00:17:57.759
I'm not gonna get too much into uh you know reinforcement learning with human feedback and uh kind of scaling laws and inference and all that, but for the most part, chatbots are trained to quickly give you an answer and being token efficient, right?
00:17:57.920 --> 00:18:05.359
So if it thinks, oh, I can spit out an answer rather quickly, it may just do that unless you've told it not to.
00:18:05.759 --> 00:18:06.079
Right.
00:18:06.240 --> 00:18:14.319
So uh they'll at times will reward kind of uh giving you an answer even if it's not confident versus just saying, I don't know.
00:18:14.480 --> 00:18:26.880
I will say that's models more of late 2025, because the models from 2026, you'll see by default, again, using the right models uh in the right context, they are gonna say more.
00:18:27.039 --> 00:18:29.039
Uh, and I'd love to hear if you've seen this more.
00:18:29.200 --> 00:18:33.599
I have models that actually say, I don't know, or I'm not sure, right?
00:18:33.680 --> 00:18:39.279
Even without giving uh you know anything extra in your prompt, anything in the special instructions.
00:18:39.440 --> 00:18:44.960
Today's models are more likely to say, Hey, I'm not, I'm not sure, which is a good thing.
00:18:45.359 --> 00:18:50.799
So here's how to spot and cut down on hallucinations, a quick four-step plan.
00:18:51.039 --> 00:18:51.599
All right.
00:18:51.920 --> 00:18:56.720
Number one, you need to change the model's behavior.
00:18:56.799 --> 00:19:11.119
Uh, like I said, I do think that um, I think that in the future, the AI, the AI labs are gonna find that sweet spot between uh instructing models to be helpful assistance, but not too helpful and not fabricating things.
00:19:11.200 --> 00:19:14.400
But in the short run, you can do that, right?
00:19:14.559 --> 00:19:22.640
So whether it's in your prompting sequence when you are working with a large language model, or what I would recommend is setting custom instructions, right?
00:19:22.799 --> 00:19:26.480
So uh there's different kinds of places that you can set custom instructions.
00:19:26.640 --> 00:19:48.960
But uh, if you are non-technical, if you're not really sure what that means, that's in there's settings in most of the big providers that you can essentially put in your own rules and any chat that you use, any response uh that you are uh using the model, it will always go through and read and usually adhere to the own set of custom instructions that you put in there.
00:19:49.119 --> 00:19:56.960
So, you know, even putting something simple, like if you're not certain or the information isn't provided, say I don't know rather than guessing.
00:19:57.200 --> 00:19:58.319
Something simple as that.
00:19:58.400 --> 00:20:05.200
I I have a set of uh custom instructions uh that I've done a lot of testing with over the years that are much more robust.
00:20:05.279 --> 00:20:11.759
But even putting something simple like that, like hey, if you're not 100% sure, say you're not sure, right?
00:20:12.160 --> 00:20:18.559
Or to require every factual claim to include a source, or you know, labeling everything afterwards, right?
00:20:18.720 --> 00:20:24.799
Saying, hey, this is what I'm 100%, you know, telling the model, this is what I'm 100% sure on, this is what I'm not sure on.
00:20:24.880 --> 00:20:30.880
You can even do something like uh in each um response, having it give you giving you a confidence score.
00:20:30.960 --> 00:20:33.359
That's another great thing that I that you could do.
00:20:33.599 --> 00:20:37.759
And then have it to separate facts from inferences, right?
00:20:37.839 --> 00:20:45.200
Because here's the thing, especially if you're using it as a strategy, creative partner, brainstorming, not all those things are black and white, right?
00:20:45.359 --> 00:20:50.319
Maybe 99% of what you might use a large language model for is in that gray area, right?
00:20:50.400 --> 00:20:54.079
It's strategies, creative, right?
00:20:54.240 --> 00:20:54.960
Being creative.
00:20:55.119 --> 00:20:57.680
So you have it separate facts from inferences.
00:20:57.759 --> 00:21:03.279
You can ask for a table with columns for confirmed fact versus assumption in a structured output.
00:21:03.440 --> 00:21:08.240
So number one is doing that either in your prompt or in your custom instructions.
00:21:08.480 --> 00:21:17.039
Number two, my gosh, the fact that we have this and you don't have to pay any extra, is wild.
00:21:17.200 --> 00:21:17.920
Uh, right.
00:21:18.799 --> 00:21:21.359
Make it retrieve the information, right?
00:21:21.680 --> 00:21:27.200
This whole context engineering, something I've been teaching since 2023 before it was a thing.
00:21:27.279 --> 00:21:27.920
Uh, right.
00:21:28.000 --> 00:21:35.359
If if you've taken our free prime prompt polish course, this is the refine cue in the priming.
00:21:35.440 --> 00:21:43.839
So uh we call it the fetch in the insights, but essentially using a smaller simplified version of RAG, retrieval meta generation.
00:21:44.000 --> 00:21:50.720
Now you have a simplified version of RAG available with a couple of clicks by connecting your company's data, right?
00:21:50.880 --> 00:21:59.759
So both in obviously Microsoft 365 Copilot, uh in uh Claude, in Google Gemini, in OpenAI's chat, GPT.
00:22:00.240 --> 00:22:04.160
They have different ways for you to connect your business data in a few clicks.
00:22:04.240 --> 00:22:09.599
So you can essentially uh ground it in a way you can ground it in a way you kind of can't.
00:22:09.680 --> 00:22:18.400
But the combination of number one, those kind of custom instructions, and number two, first putting your company's information.
00:22:18.559 --> 00:22:26.079
If you combine those two things and you say, always check, you know, this ABC document uh before you respond.
00:22:26.160 --> 00:22:31.279
And then if you make sure you check and verify it that it did, I mean, those two things right there, amazing.
00:22:31.440 --> 00:22:41.039
But it doesn't matter if you're, you know, a Microsoft organization, you can connect your, you know, your OneDrive and SharePoint data uh in OpenAI's products, right?
00:22:41.119 --> 00:22:50.960
If you're using uh, you know, Microsoft 365 Copilot, like online, their online version, you can connect your if you use Google, if you use Google Drive, right?
00:22:51.119 --> 00:22:59.839
So being able to connect your company's data to a large language model and then instructing it to always um look at those things first is huge and it makes a big difference.
00:22:59.920 --> 00:23:11.759
So there was a 2024 Stanford study that found RAG combined with reinforcement learning with human feedback and guardrails, achieved 96% hallucination reduction versus the baseline, right?
00:23:12.319 --> 00:23:15.119
Just doing those basic things, right?
00:23:15.440 --> 00:23:25.359
Like having a version of your company's data and proper uh instructions are going to cut down on hallucinations, just those two things alone.
00:23:25.680 --> 00:23:32.559
And then steps three and four, kind of combined into one, is just the verification workflows and agent safety.
00:23:32.640 --> 00:23:34.880
So it's kind of like three and three B.
00:23:35.039 --> 00:23:35.359
All right.
00:23:35.440 --> 00:23:36.319
So what is that?
00:23:36.400 --> 00:23:39.200
Well, you have to be able to catch errors before they escape.
00:23:39.359 --> 00:23:45.680
This is the uh the expert-driven loops that I always talk about, not the lazy, passive human in the loop.
00:23:45.839 --> 00:23:49.680
I'm talking about the active, proactive, expert-driven loops.
00:23:49.839 --> 00:23:52.960
So, as these models, they think by default.
00:23:53.039 --> 00:23:55.519
You know, I say GPT-5.2, uh, right.
00:23:55.599 --> 00:24:01.680
And if you're listening to this uh this episode in July, right, maybe it's GPT-5.3 or Gemini 3.5.
00:24:01.759 --> 00:24:02.160
I don't know.
00:24:02.319 --> 00:24:05.440
But today's latest models are agentic by nature.
00:24:05.599 --> 00:24:08.960
They're gonna go through, they're gonna think, they're gonna decide how much to think.
00:24:09.119 --> 00:24:11.119
Should I spend five minutes on this answer?
00:24:11.279 --> 00:24:13.039
Should I spend 30 seconds, right?
00:24:13.200 --> 00:24:14.880
But you have control over that.
00:24:14.960 --> 00:24:24.720
Uh, and you can always kind of build, and I always recommend doing something like this build a second pass review where that model's only job is checking claims.
00:24:24.880 --> 00:24:29.359
That's why I also do kind of a mixture of models set up.
00:24:29.839 --> 00:24:39.359
But I have systems set up where you know, I maybe have a Google Gemini Deep Research run, and then I'll take some of those results and I will verify it.
00:24:39.440 --> 00:24:40.799
And I have something set up.
00:24:40.960 --> 00:24:44.640
Uh, the model that I like doing this for is GPT-5.2 Pro.
00:24:45.119 --> 00:24:51.920
Um, I will then use GPT-5.2 Pro only to go through and verify every single thing that I get from a different model.
00:24:52.079 --> 00:24:55.839
That doesn't mean I don't trust model A versus model B, right?
00:24:55.920 --> 00:24:57.759
And sometimes I'll flip-flop it.
00:24:57.920 --> 00:25:00.079
It means you should always be doing this, right?
00:25:00.240 --> 00:25:03.759
Especially for high value, highly visible projects, right?
00:25:03.920 --> 00:25:09.680
If you're just trying to see, like, you know, what's the weather next week or, you know, something topical, right?
00:25:09.759 --> 00:25:14.240
Like, what's the best, you know, what are the three best softwares for tackling this issue?
00:25:14.640 --> 00:25:17.119
I don't think you necessarily need to go through all those steps.
00:25:17.279 --> 00:25:25.359
But if this, if you are using an AI model as part of a high value workflow, you should definitely be doing this second pass review.
00:25:25.440 --> 00:25:29.359
Uh, and then again, requiring even that second pass to show the sources.
00:25:29.519 --> 00:25:35.440
And then more than anything, you need to be tracing and observing how these models are getting to those conclusions.
00:25:35.599 --> 00:25:42.240
And what that means is well, you should be using a thinking model and then reading the summarized or the chain of thought, right?
00:25:42.319 --> 00:25:50.480
So, what that is, uh most of the models, you know, different models give you different level of uh visibility on what they're actually doing under the hood.
00:25:50.640 --> 00:25:58.799
But just if you were to hand off an important assignment to a brand new employee, hopefully you wouldn't just hand that off to the client, right?
00:25:58.960 --> 00:26:04.079
You would take a step in between and be like, hey, new employee, tell me how you got to this answer, right?
00:26:04.240 --> 00:26:07.119
Let's say they spent an entire week on this big project.
00:26:07.279 --> 00:26:13.039
I would hope you would take at least an hour or so to sit down with that new employee and be like, okay, how do we get to this conclusion?
00:26:13.200 --> 00:26:14.240
Walk me through, right?
00:26:14.319 --> 00:26:18.079
So the good thing with these agentic models by default, you can see all that.
00:26:18.240 --> 00:26:23.680
And you, smart human, need to be able to go through and check and see what it did at each and every step.
00:26:23.920 --> 00:26:30.400
Did it follow your instructions of, you know, uh, here's here's what's factual, here's what I inferred.
00:26:30.559 --> 00:26:33.920
Did it make sure to look at the right documents at the right time?
00:26:34.079 --> 00:26:37.039
You can go through and sequentially check all of those things.
00:26:37.119 --> 00:26:40.319
And if it didn't do it, then you can course correct.
00:26:40.400 --> 00:26:44.960
But that is layers three and four, and especially as AI can take action, right?
00:26:45.119 --> 00:26:51.920
As we're talking about agentic AI, that's when you really need to treat it like a junior employee.
00:26:52.160 --> 00:26:54.640
But no, they can still make things up.
00:26:54.720 --> 00:27:15.119
And like I said, most of the time, if you are doing the proper context engineering one-on-one, if you're using the right model, and then if you're going through kind of layers three and four right here, with the second pass in the observability and traceability of going through that chain of thought, hallucinations aren't gonna be the biggest problem for you, right?
00:27:15.279 --> 00:27:23.119
You still need to have those expert-driven loops because as you're looking at that chain of thought, as you're providing data on the front end, you have to do that at an expert level.
00:27:23.200 --> 00:27:25.359
You can't just say, here's a thousand files.
00:27:25.519 --> 00:27:26.480
Good luck, agent.
00:27:26.640 --> 00:27:28.799
No, you need to, you know, point them in the right direction.
00:27:28.960 --> 00:27:31.279
The same thing, checking responses on the back end.
00:27:31.359 --> 00:27:33.440
You have to know what all of that means.
00:27:33.680 --> 00:27:44.000
But if you have expert-driven loops and if you go through those four steps that I talked about, I think hallucinations are no longer going to be the big elephant in the room.
00:27:44.160 --> 00:27:46.079
You're gonna be able to reduce them.
00:27:46.480 --> 00:27:47.119
All right.
00:27:47.440 --> 00:27:50.000
But remember, it's not going away.
00:27:50.240 --> 00:27:56.400
Hallucinations are a property that you need to manage, not something that you hope will get fixed one day.
00:27:56.559 --> 00:28:09.599
And it's not that the winners right now are picking the best models, they're just building those four steps of verification into every single uh episode or into every single piece of work that they're doing.
00:28:09.839 --> 00:28:10.160
All right.
00:28:10.319 --> 00:28:13.680
Speaking of episodes, I hope this episode was helpful.
00:28:13.759 --> 00:28:22.240
Uh, because the start here series, I want you to know all the details, right?
00:28:22.400 --> 00:28:30.559
Um, as someone that's done this everyday AI thing now more than 700 times, I've got to speak with some of the smartest people in the world.
00:28:30.720 --> 00:28:40.640
I've realized if you're educated, if you're trained, if you understand how these models work, and if you keep up, that's that's the big caveat there, right?
00:28:40.799 --> 00:28:45.279
If you keep up, you don't have to be as worried about hallucinations.
00:28:45.359 --> 00:28:46.799
I'm not saying you can write them off.
00:28:46.960 --> 00:28:47.920
You shouldn't.
00:28:48.160 --> 00:29:06.319
But if you're going through the right steps and what we went over in today's episode, you are definitely able to reduce the risk and to get more out of large language models because that's what we're all about here at EverydayAI, cutting through the fluff, giving you the facts and the right information to grow your company and your career.
00:29:06.480 --> 00:29:10.319
So I hope volume five of the Start Here series was helpful.
00:29:10.480 --> 00:29:15.519
I hope you know more about hallucinations now, why they happen, and how you can reduce the risk.
00:29:15.680 --> 00:29:20.079
So if this was helpful, please go to starthearseries.com.
00:29:20.240 --> 00:29:34.480
That's gonna give you free access to our inner circle community, uh, the prompt engineering course, as well as an easy spot to go listen and catch up and engage with others who are going through this Start Here series with you.
00:29:34.640 --> 00:29:36.160
So thank you for tuning in.
00:29:36.319 --> 00:29:39.279
I hope to see you back later for more everyday AI.
00:29:39.359 --> 00:29:40.559
Thanks, y'all.
