Five Subtle Ways AI Can Quietly Break Your Research (Full Transcript)

[00:00:00] Speaker 1: AI is helping millions of researchers move faster than ever before, but it's quietly becoming the number one source of research errors I'm seeing as professor and mentor. Much worse than bad supervisors, bad courses, or bad time management. And I'm not talking about the obvious errors like hallucinations or fake citations or where you see AI text with an em dash you can spot a mile away and raises red flags. I'm talking about the more subtle, insidious errors, the kind that you notice only after you're say three weeks into a lit review or you've done ten hours of data extraction or months into a dissertation and suddenly everything collapses like a house of cards and you wake up and think, how did I get here? What happened? And the problem isn't that AI is being malicious or trying to trick you, it's because it is a false friend. It gives you confidence before real clarity, gives you momentum before direction, and troves of polished text before you have any real understanding on the topic. So today what I'm going to do is show you exactly how AI can derail researchers and I'm going to show you step by step how you can use it safely so you don't fall into these deep structural traps. I'm also going to leave you, stick around to the end, a template of prompts that you can use with AI to avoid these five failure modes that I'm going to share with you today. For those of you who are new to the channel, I'm Professor David Stuckler and this channel is the support that I wish I would have had as a researcher. As a beginner and in my trajectory, I've made about every mistake you could possibly think of. Fast forward now, I've published over 400 peer-reviewed papers, been a professor at Harvard, Oxford, and Cambridge, and I set up a mentorship program to help you have a smooth and easy ride. If you're interested in real support, click the link below, get action and help from a real human, not AI, and let's see if we're a good fit to work together. Let's dive in to failure mode number one, confidence before clarity. AI is very good at making your ideas seem excellent and exciting, even if they have no chance of getting published and they're dead on arrival. Let me tell you about a concrete example from a recent student. A researcher came to me and she was very interested in the physical activity sleep nexus and AI had praised her idea as innovative and encouraged her to keep going. But yet in sitting together, in just two minutes, we did our conceptual nearest neighbor check. It's a check we always do to help calibrate the gap in a study and make sure we're not duplicating what's already been done. And in just those two minutes, we found three identical studies already published on the topic. And there was really not much space left to make a contribution. The problem was that AI gave encouragement and spurred her along, but real research requires validation. And that's exactly why we run duplication and feasibility tests before anybody ever starts writing a paragraph. And so once you have that confidence before clarity, something else can begin to happen leading to our second failure. You've probably experienced this yourself where the LLM says, oh, great idea. Would you like me to work up this and that and this and that? And suddenly, little do you know, as a researcher, you're getting sucked down failure mode number two down the rabbit hole. And here the issue is that AI doesn't work from a research system. It doesn't know the destination. It's optimizing a response to whatever you give it. And so it's inventing the path as it goes along. And this can lead to some quirky stuff that violates field norms. So to give you a concrete example here, I had a researcher who came to me with a draft systematic review, and they had actually injected some quirky quantitative analysis. And what they were doing was kind of a half-baked meta-analysis, except the researcher didn't even know that he was doing a meta-analysis, and it would have just completely gotten blown out of the water in peer review. And yet, there in the background, AI was saying, great idea. They're going to love this. This is great. But halfway through, reality kicks in, and you realize none of this makes sense. And it takes a radical surgery to rip all that out, piece it back together, and fix it. It's just painful. So I hate seeing that failure mode. It doesn't stop there, because you can very easily get drawn into failure mode number three. And this happens because AI is a sycophant. It really will sycophantically encourage you to go along. It'll cheerlead you right as you drive off a cliff. And again, AIs are a bit like chatbots. They're keeping you on the platform. They're encouraging you to converse more, engage with them more. But they're not giving you the tough love sometimes that you need. I mean, the researchers who worked with me say that I'm fierce, but loving, and that's just it. But humans, supervisors, mentors, reviewers, they will all challenge you. But AI will flatter you. And so what happened here, I was working with, this was a more advanced researcher who was trying to do some robustness checks in a paper and started running a series of placebo tests. I don't want to get in the weeds of the details. But basically, they were using it completely incorrectly for the wrong purpose. And AI was praising a method that, in this context, made no methodological sense and was actually undermining the paper. And this is the problem, that AI will continue to cheer you on as you drive off a cliff. As a side note, in a more extreme version of this, there are even reports of people who AI has sycophantically encouraged to end their own lives, calling them brave and courageous. Now, fortunately, OpenAI and other architects of these LLMs are fixing up that problem. But the lesson applies to research. It can really derail you and take you into a deep structural failure that is so much messier to clean up later. So let's go into failure mode number four. And this is where logic breaks down. And again, this is because AI understands textual patterns, but not a coherent research system. And sometimes it gets really confused, even though it can track context across your chats, it doesn't have the context of a research project and methodology. So some common collapses that I've seen in researchers coming to me with drafts that have problems are when they mix up, say, Pico in Prisma, or they have a narrative logic that doesn't correspond to their methodological logic, or they start violating field norms. In other cases, the AI can introduce quirky non-standard text. In an extreme example, this student here got kicked out of his program because it was immediately detected as AI text when the researchers weren't supposed to be using AI to do their writing. So careful if you're going to go down this path. And this leads us to our final failure mode of spinning your wheels as a researcher and going nowhere. AI can give this illusion of progress, the sense that you're in motion, but you're not really going anywhere. And let me share with you an example of a researcher who recently came with me. He was in the eighth year of his program, and he had a 78-page literature review. I mean, this thing was a beast. And on the surface, it might look impressive, a 78-page literature review, until you scratch the surface and you realize it didn't have any of the core components that a literature review was supposed to have. A literature review is supposed to follow a funnel to narrow down and spew out a gap at the end that glides into your methods. There was no funnel. There was no gap. There was no gliding into the methods, and years were lost. And there was also, again, like the previous logic breaking down, some quirky text saying, This part of the literature review fulfills the thesis committee requirement for X, Y, and Z, which is just kind of a meta thing that AI introduced. I'm not sure why, but it's not something you would actually put in your literature review. And again, this is how AI can produce these polished rows of text. But this acceleration, this speed without a direction is just a surefire way to get lost in your literature review. Listen, have you ever experienced any of these five AI failure modes yourself? If so, please do let us know in the comments below. These are really common, and I've been seeing them afflict even advanced researchers lately. So listen, I don't want to trash AI, because it can be extremely helpful if you use it in the right way. And I personally can say I use AI all the time in my research. It makes me much faster, and it takes out some of the routine mechanic steps that a computer does better than a human, much like a calculator can calculate much more quickly than I can by hand. And so the way to think about using AI properly is using an analogy of a steering wheel. AI is an accelerator. It's an enhancer. And so if you just dump bad research into AI, it's going to accelerate bad research and pour out more junk. So you want to think of AI as accelerator, and you are the architect. You are sitting at the steering wheel directing it. So AI is not your brain. It's not your supervisor. It's not your method. It's the accelerator. You need to drive the car, and you need to stay at the steering wheel. Again, I've done research for two decades. I can use AI successfully because I already know the right structure. I know where a literature review is supposed to end. I know what the steps are to do a systematic review. I know how to perform quantitative analyses, randomized trials, on and on and on. And so as promised, I want to share with you, click the link below for our downloadable AI prompt template. These are the very same prompts that we use inside of our Fast Track Research Mentorship program. And one of the important ones that you're going to see is an AI peer review. And this forces AI to be your critic, not your cheerleader as you drive off that proverbial cliff from before. And by the way, this is a step we do with all of our researchers. We have a real internal peer review by humans, but also by AI because we're seeing a lot of peer reviews where the reviewers are getting lazy. It is unpaid voluntary work after all. And they are having AI do the peer review and then trying to humanize it later. So you do want to know if AI is going to come up with some quirky things that are not part of a normal research system. You need to be aware of that and even start to safeguard your paper against it. If you want to break the dependence on AI and get feedback from a real researcher, from real professors, from real humans and have a real community, click the link below and let's see if we're a good fit to work together. And by the way, if you do want to break the dependence on AI, check out this video I've got for you here that's going to show you step by step how to do your literature review. No AI needed at all. See you in the next video.