Transcript 0:00 Welcome to We The People in Tech or Not: The Debate. 0:03 This is an AI-generated podcast based on a post from the We The People in Tech or Not newsletter by Alex Rosa, and its goal is to help listeners learn new ways to explain technical topics from a beginner, non-tech understanding up to an expert level. 0:18 So imagine you pour your expertise into a highly important document. Oh, yeah. We've all been there. Right. 0:23 You refine your thoughts, you structure your arguments, and then, you know, just to be safe, you run it through an artificial intelligence assistant. Just for a quick polish. Exactly. 0:32 But it immediately hands you a list of four or five things that change. So you diligently apply them. You hit review again, expecting a clean bill of health, and, uh, it gives you another list. 0:41 Right, and it just keeps going. It does. If you do it a third time, you get yet another list. 0:46 So today, we're exploring this, this infinite loop of AI optimization and why knowing when to stop might actually be the most critical professional skill you can develop. 0:56 I'll be arguing today that we absolutely must put a hard stop to this endless feedback cycle, uh, if we want to protect our unique expertise and, frankly, our sanity. 1:06 And I am coming at this from the perspective of, say, a busy manager or perhaps a smart reader who doesn't necessarily work in technology. 1:14 I hear the warning about the infinite loop, I do, but we have to look at the reality of the modern workplace. Sure. I want to question whether capping AI feedback is actually realistic. 1:24 Like, I want to ask whether non-experts might genuinely, desperately need that endless machine polish just to survive. You know, just to communicate effectively in a really fast-paced environment. 1:34 Well, the tension between efficiency and expertise is exactly why we need to dig into this today, because the core issue we're facing is that AI is fundamentally designed to suggest changes. 1:43 I mean, it, it is what it does. Right. But we have to internalize a very difficult truth here, which is that different does not always mean better. 1:51 When we allow an AI to endlessly review and revise a document, it systematically strips away the human voice. Okay, but- And worse, this unchecked feedback loop, it triggers this massive wave of imposter syndrome. 2:03 You know, the feeling that you're a fraud even when your work is objectively excellent. 2:07 I hear you on the imposter syndrome, but I think we need to define what the AI is actually doing beneath the surface before we totally condemn it, right? Okay, fair enough. 2:15 We are dealing with large language models which are, at their core, probabilistic prediction engines. Yes. 2:21 So when you hand a document to an AI and ask for a review, it's not reading your text looking for some objective mathematical truth. 2:27 It's calculating the most likely stylistic variations based on, well, millions of data points it was trained on. Exactly. It's designed to offer alternatives. 2:36 And for a lot of businesses, that homogenization, that standardization of text, is exactly the efficiency they are paying for. 2:43 Yeah, I agree with your definition, but that probabilistic nature is the exact mechanism that makes blindly trusting it so dangerous. How so? The algorithm is programmed to fulfill a prompt. 2:53 If you ask it, "How can I improve this?" It mathematically cannot output, "Looks perfect. Take the rest of the day off." Well, true. It will find something to change. 3:03 The danger isn't just that your text becomes standardized. The real danger is the psychological fallout from that. I understand that getting a lot of feedback can be jarring, sure, but let's contextualize it a bit. 3:13 If you're writing a complex technical document, you're basically drafting the blueprints of a house but for computer systems. Right. That's a good analogy for the text. It's an assessment, often for a customer. 3:24 So if I run my blueprints through an AI, and it flags 50 potential issues, isn't that a good thing? I mean, if an AI tells me to move a load-bearing wall, shouldn't I listen to it? 3:32 Well, if it is truly a load-bearing wall, like a critical factual error, then yes, absolutely. Right. But that's the illusion of the AI critique. 3:40 The massive lists of, quote-unquote, "errors" that come back aren't usually structural failures. They are almost entirely stylistic choices about how you explain the concept. 3:49 Okay, but 50 stylistic choices can still be overwhelming. They are. And look at the speed of delivery. A human being spends four hours carefully crafting a three-page technical opinion. 3:59 They click a button, and within four seconds, the AI spits back a massive bulleted list of flaws. Yeah, it's instant. 4:06 The human brain just cannot intuitively process that a machine didn't actually read and evaluate the text. 4:12 The immediate psychological reaction is, "Wow, if it found this many mistakes in four seconds, I must be terrible at my job." I, I hear that, but does that experience translate universally? 4:23 You're talking about seasoned experts writing complex system blueprints. Right. Let's step out of the tech expert bubble for a second. 4:29 Imagine a beginner or a junior employee writing a high-stakes technical email to a major client. Okay. They don't have decades of experience. 4:37 When they get that long list of AI suggestions, those might actually be catching real critical mistakes. They might be preventing a disaster. 4:45 That's a fair point, but- How does a non-expert know the difference between a stylistic variation and a load-bearing wall? I mean, if you don't know the rules of the road, shouldn't you follow the GPS implicitly? 4:55 Following the GPS off a cliff because you stopped looking at the road is exactly what happens when you build that dependence. Off a cliff? That seems extreme. But it's true. 5:04 If you accept everything the AI suggests without understanding why, you fundamentally surrender your authority over the text. Let's look at a concrete example from the source material. Okay, let's hear it. 5:15 Imagine you write an opening hook for a technical post. You talk about being an optimist specifically, learning how to keep bad moments small so they don't infect your personal life. Right. 5:24 It sets a very specific human tone before we're diving into the technical weeds. Which is, frankly, a pretty bold stylistic choice for a technical document. It is. 5:33 And the AI's very first recommendation will almost universally be to cut that opening entirely. Interesting. It will tell you to go straight to the technical problem. 5:42 The AI wants to remove the personal element because it doesn't fit the statistical average of a standard technical document. But for a beginner, isn't clarity more important than their burgeoning unique voice? 5:53 If the AI says the personal hook is distracting, I mean, it probably is. Well, you can make it shorter, sure. You can refine it. 6:00 But if you strip it out entirely just because the machine told you to, you lose your ability to defend the work. Defend it how? This is where professional credibility is lost. 6:10 If you hand in a document that the AI essentially rewrote and a colleague or a client questions a specific paragraph, you have no rationale for it. Oh, I see. 6:19 You cannot explain why that paragraph is there because you didn't write it, the machine did. Your credibility takes a massive hit when people realize you're just a passenger in your own work. 6:29 You don't even understand the very blueprint you're presenting. Okay, nobody wants to be a passenger, and I can definitely see how being unable to defend your own document is a huge career liability. Exactly. 6:40 But let's evaluate the practical way out of this because this is where the friction is immense for me. 6:45 The proposed solution from the newsletter to stop this infinite loop and preserve your authority is this, uh, highly structured five-step validation process? 6:54 Yes, you have to install a hard stop in the review loop, and the methodology is actually quite brilliant. Brilliant? Okay, walk me through it. Step one, you ask for a single critique from the AI. Just one. Got it. 7:06 Step two, you read it yourself first because deep down, you usually know which suggestions genuinely make sense. Sure. Step three is the crucial one, a cross-check. 7:15 You take the text and use a second AI tool to validate the first one. Wait. Then step four, you apply only what is strictly needed. 7:22 And finally, step five, you run one final round purely to learn what the machine sees, but you do not apply it. Okay, hold on. Stop right there at step three, the cross-check. 7:32 Let's talk about the time cost of this process. The time cost is worth it. But we're talking about everyday workers here. 7:38 Say a busy manager is using Microsoft Copilot, which is built right into their Word document or their Outlook email. They get a critique. 7:44 You are saying they then need to read it, copy their text, open up a completely different ecosystem like Google's NotebookLM, paste it in, prompt it, and wait to get a second opinion? Yes, that is the process. 7:55 Then filter the result and then do a final round just for educational purposes? That time cost is astronomical. I have 50 emails to clear before lunch. 8:03 I cannot run a five-step multi-platform diagnostic on my daily communications. Is this actually realistic? Now, I know it sounds high up front, but you are miscalculating the true cost of time if you don't do this. 8:14 How am I miscalculating it? This five-step loop is an investment that actually prevents a massive time drain. Let's revisit the probabilistic nature of the AI you mentioned earlier. 8:23 Every single round of AI review generates four to five new recommendations. Right, endlessly. Endlessly. Endlessly. 8:29 If you do not install a hard stop, if you just rely on Copilot and keep hitting review, you fall into the trap of endless tinkering. You fix five things, you run it again, it gives you five more tweaks. 8:39 You will spend an hour on a document that was finished forty-five minutes ago. Okay, I see the danger of the endless tinkering, but why a second tool? Why NotebookLM after Copilot? 8:49 Why not just ask the same tool to double-check its work? Because asking the same algorithm to check its own bias just doesn't work. The source material uses a perfect analogy here. 8:57 It's exactly like asking a second doctor for an opinion. A second doctor. Right. If the first doctor says you need surgery, you don't ask that same doctor if they're sure. 9:05 You go to a different hospital with different diagnostic criteria. That makes sense. Different AI models are built on different underlying architectures. 9:12 If Copilot and NotebookLM both flag the exact same structural issue, you validated that it's a real problem, not just a quirk of one model's training data. I understand the logic of the second doctor, I do. 9:24 But you're still asking people to run a final fifth step just to learn while deliberately ignoring the advice. Yes. Is the second round of AI feedback really just arbitrary tinkering? 9:36 I mean, doesn't the machine find legitimately new ways to improve the text? The data shows an incredible drop in actual value. The diminishing returns are staggering. 9:45 Let's look at the real world test case from the source text. Okay. First round of review, the AI suggests cutting the optimist opening, explaining terms with clearer analogies, and breaking up long sentences. 9:56 You apply the ones that make sense. Seems reasonable. Right. Then you run it for a second round. What does the machine do? Well, it analyzes the new text. 10:05 Yes, and because it must generate a response to fulfill its programming, it gives four suggestions, but three of them are literally the exact same advice from the first round, just dressed up in new vocabulary. 10:15 Repackaged how? Like, what does it actually say? Well, it will look at the analogies you just added, like the house blueprints or the second doctor we've been talking about. 10:24 It will praise them and then suggest you just add more of them. Oh. It might say, "Consider calling the AI a study buddy," or maybe an over-the-shoulder editor. It isn't offering structural critique anymore. 10:35 It's just spinning its wheels offering synonyms. Just churning out text. Exactly. Historically, across dozens of complex posts, only two or three recommendations from these secondary rounds ever actually get applied. 10:47 You are wasting precious hours chasing an algorithm's mandate to simply generate text. That is a very compelling breakdown of the diminishing returns. 10:55 I can concede that chasing the algorithm into round two and three is a waste of time. It really is. But there has to be a distinction in the workload here. 11:04 You cannot apply this five-step second doctor cross-check to every piece of communication. And you shouldn't. The nuance here is crucial. This rigor is reserved for important work. Like what? 11:15 High stakes blueprints, major technical assessments, complex architectural documents where your specific technical opinion is on the line. Routine work does not need this heavy review process. 11:25 The point is having the discipline to put a stop in the loop when your professional authority is at stake. Reserving it for high stakes work makes it much more practical, I'll give you that. 11:34 But I want to shift the focus away from the individual writer for a moment and look at how this tool operates between people. Okay, sure. 11:41 We need to talk about interpersonal dynamics and workplace authority because this is where the psychological impact of AI scales dangerously. 11:48 We talked about imposter syndrome in isolation, but what happens when AI feedback is passed between colleagues? It becomes a massive cultural problem. 11:57 There is this troubling trend where people are taking a colleague's document, feeding it into an AI tool, and sending the raw, unedited result back to the author as a mandate. Just forwarding the AI output? Yes. 12:08 They just copy the bullet points, paste it into an email, and say, "This is what I found that needs to change." Well, I wanna push back on classifying that as inherently problematic. 12:17 Let's go back to my role as a busy manager. Okay. 12:20 If I use an AI to review a team member's work, and the AI flags a real structural issue, say, a violation of a strict corporate publishing rule, why should I tiptoe around the fact that an AI found it? 12:31 I mean, I need to enforce that standard. Efficiency is key. We have to draw a very sharp line here between compliance and style. The author completely agrees with you on hard rules. Right. 12:41 If a text violates a strict corporate policy, enforcement is mandatory, regardless of who or what found it. Let's look at the Microsoft publishing rules example from the post. Okay. 12:51 If you write a public post as an employee, you are required to include a disclaimer at the bottom stating that the views are your own and not Microsoft's. 12:58 If a colleague misses a critical binary rule like that on a mission critical team blog, the reviewer has to enforce it explicitly. Makes sense. You tell them, "This is not me asking. This is a Microsoft rule." 13:09 That is compliance, and it keeps the employee out of trouble. Right. So in those cases, the AI is functioning perfectly as an automated compliance checker. Exactly. 13:18 But the breakdown happens with general stylistic feedback, which, by the way, makes up 95% of what the AI spits out. Sure. 13:25 If you share an AI stylistic review with a colleague, the absolute rule must be this: you must explicitly and clearly state that it is an AI suggestion, not a human mandate. 13:36 Because of the emotional weight a human mandate carries. Yes, exactly. Passing off a machine's subjective stylistic preferences as your own direct orders places a profound psychological burden on the recipient. 13:49 I can see that. 13:50 When a human boss reviews a document and finds five mistakes, the employee knows the boss spent an hour carefully reading their work, but when you pass off an AI's massive bulleted list of 50 minor tweaks as your own, the employee thinks, "Wow, my boss found 50 things wrong with this?" 14:06 Ouch. Imagine opening that email. You read it fast, you miss the nuance, and you immediately react by thinking, "Is my work really that bad?" 14:14 And that feeds directly into the imposter syndrome epidemic we discussed earlier. It actively creates it. People are coming to their reviewers genuinely distressed, asking, "Is it that bad?" 14:25 Because they only saw a wall of text and assumed they were failing at their job. Oh, wow. The reviewer then has to backtrack, calm them down, and explain, "No, no, it's just an AI suggestion. 14:35 I didn't see anything structurally wrong with what you wrote." But the damage to their confidence is already done. It is fascinating because it exposes how much we project human intent onto machine output. 14:45 We do, constantly. We assume there's a conscious mind judging us. 14:50 I can see how, especially outside of tech companies or in roles where technology isn't the primary focus, receiving a machine-generated list of flaws could be highly demoralizing. 15:00 If you don't understand the underlying mechanics of an LLM, that it's just predicting stylistic variations, you assume it's a mathematically proven grade of F on your paper. It spares no one. 15:12 The text notes that even seasoned tech workers fall victim to this psychological trap. Think about the anxiety of just waiting for the machine to process your work. Oh, the loading screen. 15:22 There is a great confession in the newsletter about using Google's Notebook LM on the smartphone. You upload a document to generate an audio overview, and the screen just says, "Generating. Come back in a few minutes." 15:33 Yeah, the waiting is the hardest part. Even an expert admits that if it takes more than five minutes, human insecurity flares up. You start thinking, "Oh, boy, this one is bad. 15:43 The machine is struggling to make sense of my garbage text." That is a highly relatable, albeit completely irrational fear. 15:49 We know logically that it's just server load or network latency or just normal processing time for audio generation, yet we internalize the loading screen as a judgment on our competence. Exactly. 16:01 And if a seasoned expert feels a momentary flash of panic from a loading screen, imagine what a junior employee feels when they are handed a three-page uncontextualized AI critique from their boss. 16:11 That is a very sobering image. The sheer volume of AI polish can move a text so far away from the human being who wrote it, and in the process, alienate the humans who are reading the feedback. Right. 16:22 I think this brings us to a really important point of convergence in our discussion today. Okay. 16:26 Despite my initial skepticism about the time cost of a five-step cross-check method, I have to agree that the psychological cost of an endless uncontextualized AI critique is simply unacceptable. 16:37 We cannot let a probabilistic text engine destroy human confidence. Absolutely. We have to protect the human element in our work. But at the same time, we have to acknowledge the immense power of the tool. 16:47 As the AI itself suggested in that second round of feedback we discussed, it really is a fantastic study buddy. Yes, it is. 16:54 It provides different analogies, allowing us to explain a technical topic one way to a group of engineers today and a completely different way to a group of marketing executives tomorrow. 17:03 It's an invaluable tool for reaching different audiences and preventing us from becoming stagnant in our explanations. 17:09 It is a tool of immense value, provided we remain the master of the tool and not the other way around. And that requires situational awareness, which is where the nuance lies. 17:19 We both agree that how heavily you rely on AI depends entirely on your context. Definitely. 17:25 A busy manager clearing administrative hurdles or a beginner trying to draft a standard professional update might safely accept a higher degree of AI polish simply for safety and efficiency. 17:35 They might not need a second doctor's opinion for a routine Tuesday update. I completely agree with that nuance. Routine work does not require the full rigor of the loop. 17:44 But anyone writing a complex blueprint document where their specific technical opinion, their unique expertise, and their professional authority are on the line must actively guard their voice. Right. 17:54 You cannot afford to apply every single recommendation, or you will lose your authority over the text and damage your professional credibility when you're asked to defend it. So where does this leave our listeners today? 18:05 What is the practical takeaway for the person who is about to open up their word processor right after this debate? The most practical next step is to run an experiment on yourself. 18:14 The next time you have to write an important email or a critical document, try the five-step loop. 18:19 Send it for one critique, read it yourself, get a cross-check from a second tool, apply only what you truly think is needed, and run it one last time strictly to learn what the machine sees. Just to observe it. Exactly. 18:31 See how many suggestions you actually keep and how many you realize are just the machine doing what it was built to do, finding something, anything to change. 18:39 And equally important, if you are a manager or a peer reviewing someone else's work and you use an AI to generate that review, you must clearly label it. Yes, please. Do not pass it off as your own human disappointment. 18:51 Say clearly, "This is an AI suggestion." It costs you absolutely nothing, and it might save your colleague from an entirely unnecessary bout of imposter syndrome. 19:00 Which leads us perfectly into a call to action for our audience. There is a fascinating one-click poll series running right now called Human versus AI. 19:08 The core question is simple but profound: Has an AI review ever made you doubt your own work? That's a great question. We highly encourage you to participate in that poll. Have you experienced it more than once? 19:19 Have you seen it happen to a coworker? Or has it never happened to you at all? It is a question that forces us to look in the mirror rather than just staring at the screen and waiting for the machine's judgment. 19:29 Well said. We invite all our listeners to explore this topic further and to read the full post, which is available for free at intechornot.com. That is intechornot.com. 19:41 As we wrap up today, I keep returning to that vivid image of the infinite loop. Yeah. You fix the text, you feed it back into the machine, and the machine hands you another list of changes. 19:51 It's so easy to see that jagged list of critiques and feel like your professional foundation is broken. But as we've explored today, perhaps the machine isn't diagnosing a fracture at all. 20:01 Perhaps it's just offering you a different color cast for an arm that is already perfectly strong. And knowing when to say enough might be the most human skill of all. Thank you for joining us for this debate.