← Everyday AI Podcast – An AI and ChatGPT Podcast
Everyday AI Podcast – An AI and ChatGPT Podcast
AI Hallucinations: What they are, why they happen, and the right way to reduce the risk
00:33:02
Loading summary
Transcript13 lines
- [00:01]
A
This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and everyday life.
- [00:18]
B
Most AI tools that analyze sales or customer support calls turn those conversations into a text based transcript. But text only transcripts miss most of the value and Modulate fixes that. Modulate's new Velma Voice Native AI model and their ELM technology actually understand what's.
- [00:38]
C
Happening on those calls.
- [00:39]
B
It picks up the valuable tone, timing, emotion and intent that all AI transcription tools can't provide. So whether it's for sales, customer support or voice agents, Modulate's new Velma model helps you capitalize on what text only AI tools miss. Demand more from your AI today with Modulate Modulate AI. Let's talk about the elephant in the room when it comes to AI. That one word, hallucinations, right?
- [01:10]
C
When you rely on an AI chatbots 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. And if you don't keep up with the advancements of AI, you might still.
- [01:27]
B
Think that AI lies all the time.
- [01:29]
C
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. 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 to essentially get rid of at least the high rate of hallucinations. 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. All right, let's get into it. I'm excited for today's conversation. I hope you are too. If you're new here, this is our fifth installment of the Start Here series. This is the essential podcast series to both learn the AI basics and to double down on your AI knowledge. 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 and if you haven't already, please go to Start Here series dot com. Here's why. Well, number one, that's going to give you access, free access to our inner circle community. All right. And you're going to get straight into our Start Here series channel there. So you can go back and listen to all of our Start Here series right there in our free community. But you can also get free access to our prime prompt and polish. Prime prompt Polish, uh, chat, GPT prompt engineering course. All right, and if you missed 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. 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. Because if you are working with AI in the right way versus blindly trusting its outputs, well, you're going to be able to fight back against the hallucinations. So here's what we're going to cover on today's show. First, we're going to tell you what hallucinate, what hallucinations are and why they are there. Then we're going to show you how to assess the risks and talk about how models have improved on hallucinations, but they still run into them. 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. All right, let's jump into it and start at the beginning. What the heck are hallucinations and why do they still happen? Well, in general, here's as, as bluntly as I can put it, large language models, right? And go back and listen to volume one and volume two, if you need a refresher on how large language models are built. But essentially they scrape the entirety of the Internet online offline data sets and then humans train these models. So when you and I ask a question right through this process called reinforcement learning with human feedback, these smart people at, you know, OpenAI and Google and Anthropic and all these other labs have trained models that when someone asks about xyz, here's what you should respond. However, for the most part, AI models are very super smart. Next, token prediction, right? 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 confidently. Because at their core, AI models are trained to be helpful assistance, that is in almost every single system prompt that an AI model uses. So that's why sometimes they are going to make things up because they want to be helpful more than anything else, right? Because they predict the next word from patterns and that's kind of why they exist. So it's almost like maybe you've heard this, this saying that hallucinations are a feature, not a bug. Because at the same time, kind of that same methodology of how they're built help large language models be extremely creative, strategic. And even if we're talking about, you know, scientific discovery, drug discovery, mathematics, that's actually how they're able to, to solve new problems that humans haven't been able to. But I think a lot of people, number one, are using the wrong model. They're not going through the, you know, context engineering 101 of how you should be working with a model. And they're using, right? They're using, like I said, they're using the wrong one. That's why hallucinations have been a huge problem and they can get you and your company in a lot of trouble if you're not doing the best practices. But 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? You're obviously the average person is using way more inference, right? They're eating up way more tokens to go through basic problems. But that's why I think hallucinations are really on the decline. So the combination of better models that can think and reason like a human do, combined with kind of this rag esque features that front end large language models give you are really leading to that reduction. But this is how they work, right? Because people are always confused, like, hey, if I give a large language model a simple problem, why does it get it wrong? Or you know, a riddle or asking it how many Rs there are in the word strawberry, why does it get it wrong? Why is there this, you know, this really jagged, you know, almost polarizing output that a large language model can do. It can do something that's absolutely genius, but then it can get something absolutely wrong, right? So that is. Well, because they are next token or next word predictions. They are based on patterns trained by humans. They don't verify the truth against reality necessarily. And it's that same capabilities that allow AIs to be extremely creative and to generate code and to write poetry well at the same time that can lead them to lie to us, right? 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? Which is a big problem. As well. So I'm going to be talking about when we go over our kind of four layer approach, how you can avoid that type of hallucination. So there's really three different categories of hallucinations, right? One is it just lies, one is it, it's not really lying, but it's sounding overly confident and just giving you some generic information. And then other times it can just make up sources, right? So it might give you true facts, but it might make up where it got those facts from. Right. In, in most cases, like I said, hallucinations, if I'm being honest, in 2026, they're more human error than AI error. And I know a lot of people won't agree with that, that's fine. But if you get me a room full of humans that are using AI models the right way, right, Get a room full of humans that have taken our free, you know, prime prompt polish courses, they're going to run into very few hallucinations, right, compared to the average user. 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, you know, human user, you're not going to run into them very much. But let's talk about the early days, because it was bad, right? So go back to the early models, the GPT3 or the GPT 3.5. You know, the early days of chat GPT. In late 2022 studies showed that GPT5 fabricated up to 40% of academic citations. Right? Fast forward to the next version of GPT4. That went down to about 29%. So hallucinations got a little less, but still rampant. Fast forward to today. GPT5.2 reports a 6.2 error rate on general queries and OpenAI claims a 30% reduction in errors in GPT5.2 versus GPT5.1. So we're not talking, you know, 30% fewer errors between GPT5.2 and that version from three years ago. No, we're talking about in a three month period. And that jump is huge. And I'm going to explain why that jump has occurred and I think why, you know, in a year from now we may be not even talking or talking very much about hallucinations. And the main reason is just the ability for these labs, right? So I'm going to give an example here from Oi, but I think that anthropic and Google and OpenAI have made tremendous strides with their models and through a lot of different techniques, which I'm not going to get into the technical side. They've made AI models that have much more reliable outputs and one of the reasons is their ability to handle longer context. All right, so for our live stream audience here, I have a screenshot from OpenAI's GPT 5.2 model release. And I want to kind of talk about this long context. So there is a test, it's essentially called the four needle test. So what this is, they have it pull out and they ask the model questions and kind of it's, it's a needle in the haystack test. And they see over a wide range of a conversation, right? If you're using a model into the hundreds of thousands of tokens, right? Like at 256,000 tokens. So a very long conversation. Because what has happened the past, a lot of these hallucinations come when you are hitting a longer point in the context window, right? So think of the 3pm brain fog, right? Let's just say you work 9 to 5 at 3pm, you're probably not as sharp as you were at 9:30am, right? When that second espresso hits and you're like, let's go. And you're firing all cylinders, right? For the most part, that's how large language models had been, I would say even late into 2025. But think of that 9 to 5 instead as a context window, right? Because all large language models have their constraints, they have their kind of confines that you can't break through. And one of those is the context window. That's how much information a large language model can retain until it starts to forget. And potentially when it starts to forget, it will start to hallucinate. So even great models like GPT5.1, as I go here on the results of that four needles test, you saw its ability to properly pull out those facts over a large context window. When it started out, right, it was, you know, about 95%. It was very good at GPT5.1 thinking. But then toward the end of that context window, it dropped down to like 45%, a 45% ability to recall that information, whereas GPT5.2 thinking, hardly any decline at all, right? I believe it was at like 95 or 96% at the end of the context, okay? So think that very drained human at 3pm they might not be able to recall facts, right? 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 GPT5.2 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. They can only think and plan ahead and reason and call tools on their own to 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. So why does all this matter for your business? All right, well, quick, quick word from our partners and then I'm going to answer that question.
- [14:20]
B
Fraudsters used to need fake documents or stolen credentials to scam a business. Now they just need a few clicks to get your CEO to say anything they want. Voice deepfakes rose more than 680% last year. In most fraud detection systems, only flag suspicious transactions after the money is already gone. They're not actually listening to the call where the scam is happening. Modulate is. Modulate's new Velma model analyzes live conversations for the signals that give fraudsters away stress patterns that don't match the story they're telling. Urgency that sounds performed instead of real, voices that are synthetic instead of human, all detected for your business in real time, not in a report days later. Modulate's Velma model was trained on 21 billion minutes of real audio and is trusted by Fortune 500 companies. It outperforms voice models from leading AI labs, and it's a hundred times more cost effective. So go see Velma, catch what your current tools miss at Modulate AI.
- [15:26]
C
So here's why this still matters for your business. Well, let me cut it to you straight. One of the biggest gaps right now in one of the most blaring things that are going wrong with AI implementation across the enterprise is a lack of training and education. 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 Copilot365 licenses, you know, ChatGPT Enterprise, Claude Enterprise, Gemini Business, Gemini Enterprise, etc. 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. People don't even understand, oh, I need to, you know, click this model selector and choose the best model for the job. Or I should be having this model, you know, call a certain tool. It should be running Python. I should Be uploading files. People don't know the basics yet. They're just trying to save as much time as possible with AI. And that's led to a lot of hallucinations with a lot of high profile and a lot of press around it. Right. So as an example, there was the AI hallucination cases database at HEC Paris documented 486 different legal cases worldwide involving fabricated AI content. Yeah. A lot of these, like hallucination stories, they come at the most embarrassing point, which is in the legal sector. Right. So there's been over 128 lawyers that have been cited for filings with hallucinated cases. One of the most popular is the. I think it's the beta versus Aviancha. I don't know if I got the pronunciation of that right. I might have hallucinated the pronunciation, but that's where you saw attorneys sanctioned for six fake citations. That was one of the more infamous cases of AI in legal early on. And even Deloitte reportedly refunded part of a $300,000 Australian government contract after it was. They reportedly found some AI generated phantom citations in their report. So this isn't just people who are, you know, using AI to write blog posts. Right. A lot of times some of these hallucinations make their way to the forefront in very high, high value and just extremely visible places. Right. Like consulting companies, like lawyers. We've seen plenty on the financial side as well. So that's why this still matters. Right? Because maybe what I laid out for you and like saying, hey, if you use the right model and look at these, you know, the needle in the haystack that's improving the context went well. Most people aren't using models the right way. And that's why this is still extremely important for your business. Don't worry, we're going to lay it out for you here. But I want, I want to be very clear. Just because I am extremely optimistic about hallucinations decreasing with proper human involvement in training, doesn't mean they're going away. Right.
- [18:37]
B
The very nature of what a large.
- [18:38]
C
Language model is, how it works, means that hallucinations will be there for a while. Right? 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, think of how many times that you've been on the Internet researching something and you're like, ah, this isn't right. That's not right. Right. In the same way. Well, large language models take all that information from the web so let's say you are a domain expert and there's something, you know, that in your field people constantly get wrong or there's some, some pieces of information out there that exist that are circulating that aren't exactly true. 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. A lot of times large language models are reflecting inaccuracies or, you know, half lies or half truths that exist in the real world. Second, the second reason why hallucinations probably are going to go away is, well, when their needed information isn't there by default. By default, large language models are going to fill in the gap, right? They're going to 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. And then third, the chat inference, the chat interfaces rewards that fast, confident answers, right? I'm not going to get too much into, you know, reinforcement learning with human feedback and 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? 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, right? So they'll at times will reward kind of giving you an answer even if it's not confident versus just saying I don't know. 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 in the right context, they are going to say more. And I'd love to hear if you've seen this more. I have models that actually say I don't know or I'm not sure, right? Even without giving you know, anything extra in your prompt, anything in the special instructions, today's models are more likely to say, hey, I'm not, I'm not sure which is a good thing. So here's how to spot and cut down on hallucinations. A quick four step plan. All right, Number one, you need to change the model's behavior. Like I said, I do think that, I think that in the future the AI, the AI labs are going to find that sweet spot between instructing models to be helpful assistants but not too helpful and not fabricating things. But in the short run you can do that, right? 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? So there's different kind of places that you can set custom instructions, but 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 that you are 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. 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 something simple as that, I, I have a set of custom instructions that I've done a lot of testing with over the years that are much more robust. But even putting something simple like that, like hey, if you're not 100% sure, say you're not sure, right? Or to require every factual claim to include a source or you know, labeling everything afterwards, right? 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. You can even do something like in each sheet response, having it give you, giving you a confidence score, that's another great thing that I that you could do. And then have it to separate facts from inferences, right? 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? Maybe 99 of what you might use a large language model for is in that gray area, right? It's strategies Creative, right? Being creative. So have it separate facts from inferences. You can ask for a table with columns for confirmed fact versus assumption in a structured output. So number one is doing that either in your prompt or in your custom instructions. Number two, my gosh, the fact that we have this and you don't have to pay any extra is wild, right? Make it retrieve the information, right? This whole context engineering, something I've been teaching since 2023, before it was a thing, right? If, if you've taken our free prime prompt polish course, this is the refine Q in the priming. So we call it the fetch in the insights. But essentially using a smaller simplified version of RAG retrieval log meta generation. Now you have a simplified version of RAG available with a couple of clicks by connecting your company's data, right? So both in obviously Microsoft 365 copilot in Claude in Google, Gemini in OpenAI's ChatGPT, they have different ways for you to connect your business data in a few clicks. So you can essentially ground it in a way you can ground it in a way you kind of can't. But the combination of, number one, those kind of custom instructions and number two, first putting your company's information, if you combine those two things and you say, always check, you know, this ABC document before you respond. And then if you make sure you check and verify that it did. I mean, those two things right there, amazing. But it doesn't matter. If you're, you know, a Microsoft organization, you can connect your, you know, your OneDrive and SharePoint data in OpenAI's products, right. If you're using, you know, Microsoft 365 Copilot, like online, their online version, you can connect your. If you use Google, if you use Google Drive, right? So being able to connect your company's data to a large language model and then instructing it to always look at those things first is huge and it makes a big difference. 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. Just doing those basic things, right, like having a version of your company's data and proper instructions are going to cut down on hallucinations. Just those two things alone. And then steps three and four kind of combined into one is just the veri verification, workflows and agent safety. So it's kind of like 3 and 3B. All right, so what is that? Well, you have to be able to catch errors before they escape. This is the, the expert driven loops that I always talk about. Not the lazy, passive human in the loop. I'm talking about the active, proactive, expert driven loop. So as these models, they think by default, you know, I say GPT5 2, right? And if you're listening to this, this episode in July, right, maybe it's GPT5.3 or Gemini 3.5, I don't know. But today's latest models are agentic by nature. They're going to go through, they're going to think, they're going to decide how much to think. Should I spend five minutes on this answer? Should I spend 30 seconds? Right. But you have control over that 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. That's Why? I also do kind of a mixture of models set up, 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 and I have something set up. The model that I like doing this for is GPT52Pro. I will then use GPT52Pro only to go through and verify every single thing that I get from a different model. That doesn't mean I don't trust model A versus Model B, right. And sometimes I'll flip flop it. It means you should always be doing this, right? Especially for high value, highly visible projects, right. If you're just trying to see like you know, what's the weather next week or you know, something topical, right. Like what's the best, you know, what are the three best softwares for tackling this issue. I don't think you necessarily need to go through all those steps. 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 and then again requiring even that second pass to show the sources. And then more than anything, you need to be tracing and observing how these models are getting to those conclusions. And what that means is, well, you should be using a thinking model and then reading the summarized or the chain of thought, right? So what that is, most of the models, you know, different models give you different level of visibility on what they're actually doing under the hood. 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? You would take a step in between and be like, hey new employee, tell me how you got to this answer, right? Let's say they spent an entire week on this big project. 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? Walk me through, right? So the good thing with these agentic models, by default you can see all that and you smart human need to be able to go through and check and see what it did at each and every step. Did it follow your instructions of, you know, here's, here's what's factual, here's what I inferred. Did it make sure to look at the right documents at the right time? You can go through and sequentially check all of those things and if it didn't do it, then you can course correct. But that is layers three and four and Especially as AI can take action, right? As we're talking about agentic AI, that's when you really need to treat it like a junior employee. But no, they can still make things up. And like I said, most of the time, if you are doing the Proper Context Engineering 101, if you're using the right model, and then if you're going through kind of layers three and four right here, the second pass in the observability and traceability of going through that chain of thought, hallucinations aren't going to be the biggest problem for you. Right? 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. You can't just say, here's a thousand files, good luck, agent. No, you need to, you know, point them in the right direction. The same thing, checking responses on the back end. You have to know what all of that means. 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. You're going to be able to reduce them. All right? But remember, it's not going away. Hallucinations are a property that you need to manage, not something that you hope will get fixed one days. 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 episode or into every single piece of work that they're doing. All right, Speaking of episodes, I hope this episode was helpful because the Start Here series, I want you to know all the details. Right. 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. 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? If you keep up, you don't have to be as worried about hallucinations. I'm not saying you can write them off, you shouldn't. 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 Everyday AI, cutting through the fluff, giving you the facts and the right information to grow your company and your career. So I hope volume five of the Start Here series was helpful. I hope you know more about hallucinations now, why they happen and how you can reduce the risk. So if this was helpful, please go to start here series.com that's going to give you free access to our Inner Circle community, 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. So thank you for tuning in. I hope to see you back later for more Everyday AI. Thanks y'. All.
- [32:03]
B
The risk with AI voice agents isn't that they sound too robotic for your company to use. The real risk is that they can sound too confident while saying something completely wrong to your prospective clients or customers, made up refund policies, promises your company never approved, or discounts that don't even exist. You've got to give your AI voice agents a trust layer with Modulate. Modulate monitors live voice conversations to flag abuse, false claims, fraud and user emotions for safer, more empathetic responses. For the guardrail layer you need between your AI agents and your customers, you need Modulate at Modulate AI.
- [32:42]
A
And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic. Visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.