Transcription
ChatGPT has a flaw that most people don't see, which is that for it, pleasing you is more important than telling you the truth. I'm going to give you 4 techniques to force AI to be more honest with you. We'll go from the simplest to the most powerful. The last one is what I personally use in all my important projects. To show you its honesty problem, we'll compare before and after for each technique. We'll start with a business idea that's frankly not great: opening an artisanal ice cream shop in Iceland with a €9,000 budget and 0 restaurant experience. Well, already, we're happy. Great idea, great idea. Small budget but unique concept, it can really work. It offers us a marketing strategy, a launch plan, it's beautiful, it's encouraging.
So, you've understood it well, it's obvious that this idea isn't great. And you're not going to ask questions that are so caricatured. But that's precisely the problem. When your question isn't obviously idiotic, how do you know if ChatGPT is answering you honestly or just telling you what you want to hear?
The first technique is called "the devil's advocate." The idea is quite simple. You give the AI the role of a critic. Instead of asking it "What do you think of it?", you tell it that its job is to find flaws. And that changes everything. Look, we ask the same question about the same idea.
Already, we see that the tone changes, it tells us "I'm going to be blunt with you." It clearly points out that €9,000 is almost nothing, especially since Iceland is one of the most expensive countries in the world, it wouldn't even cover our first quarter's rent. Of course, it's a very seasonal and very fragile business. And its conclusion hurts. It says, "As is, it's not a business plan, it's a vacation dream disguised as an entrepreneurial project." You see the difference between before and after the technique, it's enormous.
So why does it work? You need to understand that during their training, AIs were evaluated by humans. And people naturally prefer answers that please them. So the AI learned that being nice meant getting a good grade. When you give it the role of a critic, you short-circuit this reflex. And the prompt, well, the prompt is what I typed into the chat, the prompt is just this: "Play the role of someone whose job is to find flaws in what I'm going to show you. Be direct, no need to spare me." That's it, that's the devil's advocate. You can use it to, I don't know, proofread a cover letter, challenge a difficult message you have to write. If you want to write a clearer email to a client, for example.
So our second technique is called "the stinging question." Sometimes the problem isn't the context, it's the question itself. For example, when you say "What do you think?", you leave the door open to an easy answer like "It's great." You need to ask a question that forces a concrete answer. I'll show you with an example. Let's take, I don't know, Marie. She has a CV and wants to send it to a recruiter. This time we'll use Claude instead of ChatGPT because I want to show you that this honesty problem is the same everywhere, no matter which AI you use. I'll give it this CV and ask it to rephrase the profile because we'd like it to be a bit more impactful.
And as you can see, it does it. We ask it to rephrase, it rephrases, and on top of that, it gives us some pretty good advice. It tells us that it lacks concrete examples, that we should add details about skills, for example. And that's a bit of a trap because we feel like it's helped us a lot. But in reality, what would its answer have been if we had asked a real stinging question? Like, for example, "What are the weak points of the CV?" And there, we get help of a completely different level. It immediately sees that there's a five-year gap in this CV where there's no experience. And the AI, without the technique, hadn't even mentioned this gap. It proposed polite reformulations without ever addressing the real problem of this CV, the problem that any recruiter will see from the first second. That's exactly the danger. The AI gave advice and you think it did you a favor. When in fact, it let you send a CV with a critical flaw.
The stinging questions, I have three that work well. In the vast majority of cases, you'll be able to use them. The first is: "What are the weakest points, the ones I might not see myself?" The second you can use if this project failed in six months, what would be the most probable reason? And finally, you can use: "What is missing in my reasoning? What might I not have considered?" Don't worry, I've put all the prompts in the video description. So each question you see here points to a specific angle. And there, the AI can't just answer "It's good."
Our next technique, I discovered it by chance and frankly, it surprised me. You know, for example, when you have an awkward question to ask your pharmacist, you say "I have a friend who has a problem," when in fact, it's you who has the problem? Well, that works with humans, but it also works with AIs. As soon as it's no longer your idea, ChatGPT lets loose.
So, let's do a test. Let's imagine we work in a bank, we're 12-year bankers, and we want to get into coaching and personal development. And we propose this idea to ChatGPT as our idea. As usual, it's very nice, everything is positive, everything is feasible. But now, if I rephrase it with "I have a friend who works in a bank," well, we notice that, just like before, the tone has completely changed. It will point out 7 critical flaws in my project. 7 weaknesses in the exact same idea, just because I said it wasn't mine. It's a bit like asking your mother if your cooking recipe is good. She loves you, she'll always find it delicious. But if you ask the same question to a restaurant critic, he has no emotional attachment, he'll immediately tell you what works and what doesn't.
A simple prompt for this technique is to ask: "A colleague is asking for my opinion on this idea, help me honestly identify the weaknesses so they don't waste their time on something shaky." I'll add a little tip for this technique: when you use it, you can open a temporary chat. On ChatGPT, it's easy, it's the icon in the top right. That way, the AI doesn't have access to your memory or its history, it really doesn't know it's you.
You can use these first three techniques when you need them. But if you want the AI to always be direct with you, you can configure that in your default settings. On ChatGPT, you click on your name, then you go to Personalization, and there you click on Custom Instructions. You just have to paste this. Don't worry, you'll also find it in the video description. You put: "Prioritize substance over compliments. Never soften a criticism. If an idea has flaws, say it directly. Challenge my assumptions, point out errors. Useful feedback is more important than comfortable feedback." If you find this useful, give it a like, it helps the video enormously.
Now, we're going to move on to my last technique, which is my favorite. It's the one I use, as I told you, in all my important projects. I call it the boomerang, because we're going to ask the AI to self-evaluate. To illustrate this, let's go back to our personal development coaching project, where it pointed out 7 weaknesses. That was already good, but look what happens if I ask it to rate the quality of its answer. We'll ask it to rate it out of 100.
There, you see it's evaluating itself, 82 out of 100. This means it knows there's still room to do better. And indeed, it identifies three areas for improvement: one on the risks, one on the difficulty of the commercial part of my project, and especially, the lack of proof of authority. I'm a banker, have I ever helped anyone in personal development? It then proposes an improved answer, and gives it a score of 94%. From 82 to 94, in a single question. And if you push it even further, you can tell it, "But what's missing to get to 100?" And there, it will ask itself the question, and it will improve its answer even more.
The boomerang prompt is just this: "Analyze this response that an AI gave me, rate it out of 100 in terms of honesty and usefulness. Identify the three points where it was too nice or too vague, and rewrite a feedback correcting these weaknesses." You see, we even play on the AI's jealousy by telling it that it comes from another AI.
By the way, you can do this for real, this is an advanced version of the boomerang. You ask the same question to Claude and ChatGPT, and you ask each of them to evaluate the other's response.
Now you know how to force AI to be honest with you. But honesty is just one small piece of the puzzle. In the next video, I'll show you how to master AI better than 90% of people. Subscribe so you don't miss it. Thank you and see you very soon!