ChatGPT can confidently give you false information because it predicts the next word probabilistically, not by understanding truth. Use the Devil's Advocate method: demand evidence, search for contradictions, and check if information has expired.
A few weeks ago, a judge based an entire ruling on cases provided to him by an artificial intelligence. The point is that they were completely made up. This is not an isolated case, and we are seeing it more and more. But the important thing is that if a judge with years of experience falls into the trap, imagine you and me, if we had to prepare an important proposal for a client or simply help our children with their homework. If you are already using artificial intelligence, be it chat, GPT, or any of the others that we have available, you will have noticed that sometimes it lies to you, falsifies data, and even makes things up. And it does this to all of us. It is not your fault. But it is not its fault either, because its intention is not to lie to you as a human would. Today, I will explain the real reason why AI hallucinates, and I will also share my methods so that you can turn that defect into a superpower and never fall into its trap again. To understand it, I want you to forget about technology for a second. AI does not know things. Its sole purpose is to find the next most likely word to construct a coherent sentence that makes sense. To help you understand this perfectly, we are going to give you two examples that are super graphic. The first is the incomplete mosaic. Imagine that you give the AI a book with missing pages and you ask it questions about those pages. The AI will never answer. I don't know what it will do is look at the pages before and after and build a story in between. It will organize the words in the form of a mosaic with the aim of constructing a coherent but completely invented story. That is what the AI does. Or the second example, my favorite, the jazz musician. Imagine a jazz musician with years of experience. If you ask him for a well known song that he knows by heart, he will be able to replicate it perfectly. But his genius does not lie in repetition. It lies in improvisation. His brain does not look for the true note. It looks for the note that can fit into the harmony, the one that sounds good. ChatGPT is that musician, and his only truth is probability. He doesn't care if what he says is a fact. What he cares about is that the song he's constructing word by word, makes sense and sounds coherent. So if his goal is to sound good and not to tell the truth, what do we do? Do we stop using it? Absolutely not. What we need to do is change our mindset. And that's the most important idea you're going to take away today. ChatGPT is not an encyclopedia. ChatGPT is the fastest, smartest intern in the world at your disposal. And what do you do with an intern who is very intelligent but who doesn't yet know everything about your method? Do you fire him? No, you be patient with him, you explain to him, you teach him and you practice with him. And that's exactly what we're going to do. Now, look, to guide our intern, we're going to use my method, which I've called the Devil's Advocate. Three questions, three counter prompts that will help you increase the reliability of artificial intelligence and turn it into a real super tool. So the next thing we're going to do is go to our screen. In this case, it's ChatGPT, but you can use it with any artificial intelligence you want. In this case, I've asked him to write me a complete report on the culinary culture of Mexico. You can ask him for absolutely anything you want. Imagine it applied to your work. Any type of economic report from any sector, anything. Once he's given us our complete report, we're going to use the first of these tricks within the Devil's Advocate method, which is the following. We're going to tell him, great, now give me three reliable and verifiable online sources where I can contrast this information. We're going to send it to him. And now he'll start searching the Internet. Notice that he tells us about UNESCO and gives us some links here. Then he also cites Wikipedia and then Mexican cuisine from the Maya to McDonald's. Well, here we have another link. In this case, what we would do is go to these links and verify that the information he's giving us is true. If so, great. We know that he is not hallucinating. And if he is, then we have already certified that what he is giving us is an invention. It is a hallucination and this information would not be useful to us. Okay, now we are going to move on to the second step of this Devil's Advocate method. And this is really wonderful for me. And we are going to make Chat GPT question himself. In this case, I have asked him for a report on climate change today and once he has given me this entire report, look what I am going to tell him. Act like a skeptical expert. Give me three arguments or points of view that contradict the answer that you just gave me. And with this, what we are doing is, as I was saying, making Chat GPT question himself and give us a 360 degree view on this subject. We are going to send it to him and look, he is going to put it to the test like a good skeptic, right? And here he gives us three arguments that question or contradict key parts of the report that he gave us earlier. So here we would analyze each of these three counterarguments and we could see if it is something that fits us or not. Also, I'll leave you with another tip. If it is a report that you have done and that you have to present and share with people at work in a meeting with investors, you can share it with Chat GPT and tell it dismantle any of the arguments that I am giving in this report. Give me all the objections that any investor, teammates or my boss could make to me, try to destroy the work that I have just done. That way you'll have a very broad view of those things that they can tell you that perhaps you're not considering or valuing at the moment and it helps you to be even better prepared. But having said that, with this second trick of the devil's advocate method, and by making ChatGPT or any artificial intelligence self question, we're also going to increase the reliability of its responses. And now we're going to move on to that last trick, that last leg of the Devil's Advocate method. And now we're going to move on to this third way to increase the reliability of the information that artificial intelligence gives us, which is checking its expiration date. We'll leave you with this last report that we did on climate change today. And one of the problems or handicaps that artificial intelligence has is that it doesn't update in real time, but rather the teams in this case, for example, the OpenAI team, the parent company of Chat GPT, or in the case of Google for Gemini, or in the case of Antropic for Cloud. These teams periodically update their knowledge to bring it up to date of the latest events, news and everything that has been happening. So we run the risk that he is giving us information that is no longer valid at that time. So we are going to do that expiration check. What we are going to tell him is what part of the information that you have given me could be out of date or have changed recently. We are going to send him this and look, good question. I am pointing out the points of the report that could be out of date or subject to change recently. So here he cites some of the points that he considers could be out of date here we would come to our discretion and we would start, for example, we can tell him to go look for web pages, obviously public ones, which he can consult in order to update the information or simply take that report and put our human touch trying to go look for that data that he is providing us and in that way update it to today's date. Okay, we have just seen this Devil's advocate method, very simple and effective. We have seen that it has three parts which are the demanding evidence, searching for contradictions and checking for expiration dates. Now you have understood that AI lies are not a flaw, but rather part of its nature of that musician jazz. And by using the devil's advocate method, that is by questioning it, by searching for those contradictions, by asking for evidence, by checking that the information it is giving us is up to date, you stop being a user who believes what you are told and you become the person who conducts the orchestra. You are the one with the criteria, the experience and the final say. I hope you put this method to the test and that you let me know in the comments if it has been useful to you and especially if it has increased the reliability and reduced those lies and hallucinations of artificial intelligence. And I remind you that every day, every afternoon, I send an email to my newsletter explaining many interesting things about the world of AI and the latest news. I'll leave the link in the description so you can subscribe. And pay attention to next week's video because we are going to see if that is true that AI is going to take our jobs. See you next week, by the way. And as I always tell you, remember that you don't have to be an expert in artificial intelligence. You just have to want to stop wasting time and have criteria. Yeah.
