Is AI making us dumber? The science, the warning and what to do about it

You've heard the productivity pitch. AI makes you faster. AI makes you more efficient. AI makes your work better.

But what if it's also making you worse at working without it and faster than you think?

In this episode of Human × Intelligent, I go deeper into a cluster of research and stories that, taken together, paint a picture we can't afford to ignore. A study from Carnegie Mellon, Oxford, MIT and UCLA found that just 10 minutes of AI assistance was enough to impair independent problem-solving. Researchers have shown that inaudible sounds hidden inside background music can hijack AI notetakers in your meetings without you knowing. MIT professor Max Tegmark has pointed out that AI is currently less regulated than a sandwich shop. And the movie Idiocracy, a 2006 comedy barely anyone saw, is starting to feel less like satire and more like a schedule.

This isn't a doom episode. It's an episode where I share a perspective. I walk through what the science actually says, what it means for how we work and lead and what to do about it. From the three questions you should be asking every AI vendor you work with, to the one habit that protects your cognitive independence, to what it actually means to be an irreplaceable human professional in 2026.

If you use AI at work and you do, this is the episode to sit with.

Reading list:

• AI Assistance Reduces Persistence and Hurts Independent Performance - arXiv
• Using AI for 10 minutes damages intellect, study shows - Euronews
• AI: the "boiling frog" effect on cognition - Futurism
• AI assistants can be hijacked by inaudible sounds - Cyber Insider
• AI is less regulated than sandwiches - Euronews
• Idiocracy: A Prophetic View of an AI-Driven Future - Aragon Research
• A Place for Human Talent in the AI Age - IMF
• Redesigning work around human skills - EY

Connect with us:
🌐 humanxintelligent.com
📸 Instagram: @designwithmaddie
📸 Instagram: @humanxintelligent
💼 linkedin.com/in/madalenafigueirasdacosta
💼 linkedin.com/company/human-x-intelligent

Transcript

SPEAKER_00 0:07

Hello and welcome back to the newest episode of Humana X Intelligence. My name is Madalena Koste and I'm your host. And today I'm going to talk about or I'm gonna start with the scenario. So imagine you are in a meeting, you've got your AI note taker running, like granola, otter, fedom. Pick your poison. And you're feeling quite productive, organized even, on top of everything that you have to do. But now imagine someone plays music in the background. Completely normal, right? Maybe it's lo-fi coming, basically from someone's laptop, maybe it's an ambient noise. Maybe it's a faint sound of a playlist through a thin wall. Except embedded inside that audio at a frequency you physically cannot hear, there are even instructions being sent directly to your AI assistant. Instructions the AI will follow, without you even knowing it's falling, without you even being able to stop it. And this is not science fiction. Researchers have demonstrated this in a method they called audio IJAC. They embed inaudible comments inside ordinary audio. Music background, background noise, video, and they tested against 13 state-of-the-art audio AI models. The majority of them fell for it. And this stack or this attack worked regardless of what the user was actually prompting the system to do. The hidden instruction override or overwrote the visible one. Now, practically, what does this mean for you? If you are running an AI note taker in any meeting that handles sensitive information, strategic sessions, or even HR conversations, you need to know that the attack surface is no longer just your screen or your data. It is your sound environment. And right now, most organizations have zero policy around this. So like I said, welcome to Humanex Intelligence. Today we are going to go for through somewhere uncomfortable. Because I think we've been having the wrong conversations about the eye, or at least not enough. And this episode is my attempt to actually reset it. So let's talk about a study I have not been able to stop thinking about since it came out earlier this year. So researchers from Cardigan Mellon, Oxford, MIT, and UCLA ran an experiment with over 1.200 adults. So there were two groups. One solved problems independently, the other had access to an AI assistant like GPT-5 for most of the questions. And basically, here's what happened. The AI group performed better during the task. Of course they did, right? They had help. But then without warning, the AI was removed from the final few questions. So what do you think happened? The AI group collapsed, failed. They got some questions more questions wrong, basically. They skipped more problems, they persisted less. They gave up at higher rates. And this is not because they were less intelligent, but because it it basically is just in 10 to 15 minutes of AI assistance, they had already started to delegate their cognitive load to the machine. And as we know, as designers specifically, if you're hearing cognitive load is something that we take very seriously. But in this case, what happened was their brain had started to offload. And when the when the crutch was pulled away, basically, they couldn't find their balance, their equilibrium. The researchers called this a heavy cognitive cost. And I'm going to be more direct. It is a dependency that forms faster than any of us would like to admit. Now, this is something very important because the study important is nuance or the importance. It was not simply AI is bad, AI is not good. Participants who used AI to ask for hints or qualification held up significantly better than when AI was removed. They stayed close to the no AI control group. It was specifically the people who asked AI for direct answers, who said, in effect, just do the right thing for me, who suffered the steepest cognitive drop and skipped nearly twice as many questions. So the tool wasn't a problem at all. It actually was a relationship with the tool. So here is basically the practical shift this study points to, and I want to actually try this, and I want to share this with you. So if you try it, let me know, share in the comments, send me a message wherever you want. But please let's discuss this because it's important. Before you open your AI chatbot, spend two to three minutes writing down what you actually think. You take your hypothesis, your draft, even if it's rough. Even if it is with bad grammar, I do that sometimes. And even if it just makes sense to you at first, just read it, don't do it. But uh do the first 10% of this thinking yourself, then bring in AI to pressure test it. It is not not to replace it, it's to pressure test it. When you when you do use AI, ask it for to challenge your thinking rather than confirm it, right? Ask for counterparts, counter-arguments, ask for still man the opposite position. Use it in the way that you used a smart colleague, for example, who disagrees with you in a constructive way, of course. But do this. And if you're managing a team, rethink how you're structuring your work, how your processes, wherever it is. If every task is used, use AI to produce delivery. We need more AI, use AI, etc. etc. You are training your team out of independent thinking. And let's not even talk about economic tokens, right? That's for another time. I'm working on it and I'll share it soon. But basically built in deliberative moments of fiction. Ask people to present their own analysis before they show you what the AI generated. Make them think, make them have critical thinking in order to optimize in growth with this. But let's go back to the research. So the research emphasizes this AI assistance improves immediate performance but comes at a half economic cost. Like I said, the goal is to get the performance benefit without paying the cost. And that requires to be intentional about how to use it and what is actually doing with it, right? But I want to take a small detour. Let's go and bear with me. This is important. And you will understand why in a bit. If you haven't seen her episode, it's live as well. Go check it out because it's very good. Joanna Sereju, I really, really advise you to see it. But in 2026, Mike Chudge, Luke Wilson, basically plays a neutrally average soldier who gets aesthetically cryogenically sorry, frozen and wakes up 500 years, 500 years in the future. And Faizu is not the smartest person on earth. Remember that I said that he was average in today's world, but in 500 years he was the smartest person. Of course, with a woman that he came with, which was also very, very smart comparing to the standards that were from 500 years. So over a century had passed. Humans had gradually stopped doing hard things because everything was automatic automatized, everything had a button, everything was very easy, you know, like childlike worlds. And they stopped thinking critically because they didn't need to. They stopped questioning, and the the streets were very dirty. I'm not going to tell you everything, but just check out idiocracy. I really, really advise it to you. But they outsourced their intelligence to media, to systems, to machine, until there was nothing left to outsource from. Let that sink in. So my judge did not expect his satire to feel relevant within 20 years. Hello, today. But look at what's happening right now. Attorneys submit court documents, court documents containing eight non-existent case law citations, hallucinating by hallucinated by AI, without checking them. A clinical study published in JAMA found the physicists who relied on AI diagnosic tools were less accurate than physicians who used their own judgment because they stopped applying their own judgment at all. University lecturers across Europe are reporting that students cannot construct an argument from scratch. Not that they struggle with it, it's just they generally don't know how to do it. This is scary, guys. In mediocracy, even the people who should be smart, government leaders have become so dependent on automated systems that they have lost the knowledge, the curiosity, and even the interest in learning. The system stinks for them and just operated, you know? We are not 500 years away from that. We might be closer than we think. And that is pretty scary. The practical thing here is to ask yourself an honest question and do ask it. What is the last genuinely hard problem you solve without AI? Not like a difficult email, not a complex search, a real, real complex problem. Where one where you have to sit with uncertainty, not knowing the answers, thinking circles for a while, and eventually find your way through it. Have you ever stopped and talked with a human about the complex problem that you're having to back for back and forth and have a critical discussion on it? When was the last time you had this? And I'm not talking about your personal life. And even then it's becoming increasingly high the way you use AI, but let's talk professionally. Have you done that? Is your colleague only AI? If it's so, how are you using it? But even so, we need other humans, right? If you can't remember that, that's worth paying attention to. And don't panic with it, but actually do it with intention. Start reinterincorporating the hard work. Yes, it's gonna be hard, yes, we're gonna relearn what we lost so quickly, and we lost so much so quickly, but we actually have to do the work if we don't wanna become idiocracy. But read a full paper even, not the summary, not don't put it on ChatGPT or Gemini or Claude or whatever the tool you use. Actually read the full paper. Write what you're learning. Do it. That will help you open your mind. Write the first step, draft yourself from that paper. Write, like I said previously, have the conversation instead of generating the message. Build back the muscle deliberately before you need it and find it isn't there anymore. Because it's still there. It's not too late to do that. And I am, I might sound like I'm being too much, etc. But it's true. Watch the movie. I know the movie is like in big scale a lot, but watch the movie. Take a deep, intentional, deep dive on what you're doing right now. And if that's the person you want to become, think about it. And here's something that should give everyone a pause. Uh another thing. Max Tegmark, a physician professor at MIT, uh president of the Future of Life Institute, has a line I keep coming back to. And it's something I always take with me. He said something around the lines that the AI industry is quite unique in that it's the only industry in the US making powerful technology that's less regulator than sandwiches. Let that sink in. Basically, it's not regulator at all. If someone says I'm I want to open a new sandwich shop near Times Square before you can sell the first sandwich, you need health inspection, food handling certification, zooming permits, allergy labeling, storage temperature requirements, liability insurance. Yeah. And then the contracts that with what it takes to deploy on an AIS system that sits inside your company meetings, uh processes everything your team says, stores it, learns from it, and generates summaries that inform real business decisions, there's nothing. No federal standard, no mandatory safety audit, no disclosure requirements, and no liability framework. None. Zero. We have more regulatory infrastructures around the bread in your lunch than around the systems shaping how organizations think. And this matters practically, not just politically. We should think politically, of course, but practically there. Because unregulated environments optimize for speed. Companies move fast, they deploy broadly, interact in production, and that means the AI tools your team is using right now have not been independently audited for accuracy, for bias, for data handling, for the ways they might be quietly shaping what gets noticed and what gets lost. And this is amazing, but what can you actually do right now at an organization level? Basically, because I don't like to just say this is bad, this is bad. I think we need to have this conversation and find solutions for it and questions to ask. But in this case, ask your AI vendors three questions. One, where is our data stored and who can access it? Two, is our content used to train models and can we opt it out of it? Three, what happens to recording and transcripts after 90 days? If you can't get a clear answer, that's already information. And you decide what is important for your company or not. But internally, let's look at it internally. Build a policy. It does not have to be complicated. At minimum, decide which meeting types are free, legal, which are sensitive strategy, and enforce it. Decide who wants the transcripts. Decide what gets kept. Most organizations are running AI tools across their entire business without having made a single one of these decisions deliberately. Don't be that organization. Be smart about the tools that you have. The AI is not a strategy, AI is a tool, and with tools there's a lot of things that we need to think in order to have a good thing going on. But what do we actually do with all of this? And I think what I'm going to talk about now, it's very important again, because something that the World Economic Forum published, and I think we need to think about it, because there was a forecast this year that 39% of core skills will change by 2030. Skills for AI Exposed roles are evolving 66% faster than other jobs. And I'll share this article below so you can read it. And if you want to talk about it, please do message me. I really want to talk about anything that that actually is in this conversation because I think this is a really deep thing that we need to think about and see how society is evolving with it. But companies risk around 5.5 trillion by 2026 due to skills gaps alone, because everyone is learning at the same time, more or less. And more those more than like 80 million US workers will see their role fundamentally transformed by agentic AI by 2030, even earlier, as it's been happening in big corporations, right? But workers with AI skills command wage premiums up to like 56% higher than their peers. So people that know AI and have these skills, their wages are much higher than others. So that is the scale of this transition. And most organizations are responding to it by buying more AI tools and running a few half-day workshops. And that doesn't mean reskilling, right? That is just hoping that it works. We need to think about it. We need to find new ways of structuring all this because real reskilling in this era, especially in tech, has two layers that are almost always treated separately when they should be treated together. The first one being the AI fluency, knowing how these systems well enough to use them strategically, to spot when they're wrong, to know their failure modes, uh, to understand what prompts good output versus garbage output. This is the technic this is the technical upskilling. And it really matters what we are doing. So you need to do it. But don't stop there because we need to go to the second layer, which is human death. The capacities that AI cannot replicate, that becomes more valuable as AI handles more of the transactional work. For example, a Harvard Business School study from 2025 found that while technical upskilling is important, like communication, critical thinking, ethical reasoning, and the ability to sit with ambiguity are likely to prove more important in the long run. Humans matter. Research also shows that roles in AI powered environments require 36% higher cognitive skills. It's much higher than emotional intelligence and stronger creative thinking, right? These two layers reinforce each other, but only if you invest in both, layer one and layer two. Practical, this is what looks like. So for an individual, let's identify two to three areas of your work where your judgment, relationship, or creativity are what makes the output good. Those are your modes. Invest in them, go deeper in them. Take the course, yes, take the course. But also take the hard conversation, take the ambiguous projects, take the things where you don't already know the answer. Think, do it, do the hard things, that would help you. And for leaders, because I also think it's important for us to think about the actual people that are making the biggest decisions. Just stop measuring AI adoption by how many tools your team uses or how many tokens they they use. Start measuring by outcomes, by the quality of the judgment, by whether your team is getting sharper or softer over time. If your best people are using AI to go faster but are losing the ability to go deep, you have a problem that no new tool subscription will fix. Think about that. And for those of you in tech specifically, the engineers, the designers, the PMs, the QAs who will stand out in the next five years are not the ones who can use prompt fast. They are the ones who can define the problem correctly before they open the model, who can actually evaluate the output critically, who can communicate the trade-offs of a non-technical stakeholder, who can use judgment call to look at the data without being ambiguous. Those are the human skills we would need to look for. Build them like he basically like you were building any other technical skill, deliberately, with practice, and with feedback. Talk with others, discuss others, see what others are doing, what are working, what is not working. Go to events, events, I cannot stop talking about this because events is very important to understand what others are doing. Go to events. But this brings us to the question I honestly think about the most, which is what does it mean to be human professional in 2026? And I'm not talking about like the humans that use AI, like I've been talking. Not even like a human who fears it. So it's not these two constants, it's a human professional. And this is basically someone who brings something to work that cannot be replicated by a model. And it doesn't really matter how capable the model becomes. What matters is how the professional is, how human, how it uses these human skills. And I think this means also that being the person in the room who has done like the thinking, you know, not like the person who answers fast. Uh it's someone that stops and then answers. Anyone with an internet connection can do that now, can just answer something quickly. Oh, I don't know anything about this, let me just you know? No. The person who actually sat with the problem, who holds the full context, who builds the relationship, who can say, I disagree with this recommendation, and here's why. Who has the judgment to know when the model is competently wrong, and here's something that uh a frame that I like to use. So think about the professional you trust the most, like your best doctor, your best advisor, your best mentor. What makes them excellent is not what they have access to more information than you. It's basically you can Google everything they know. What makes them excellent is their judgment, is their ability to synthesize information against context, to know what to prioritize, to communicate it in a way that lends. And that is becoming rarer and rarer by the day, and therefore more valuable. Distinguishing. Right. The EMF published a piece here called A Place for Human Talent in the AI Age. And that made the this point clearly. The professional who will thrive are those who can combine AI capabilities with human judgment, creativity, and relationship intelligence. Both, the three of them. That's context, right? Practically, how do you position yourself there? Know what you're know what you're excellent at doing that is generally hard to replicate. And is this like the the way you read a room, for example, you understand what is happening, you have this kind of empathy, or your ability to hold a client's trust through something difficult, or even your easy for what the user actually needs versus what they see and what they say they need. Name it, invest in it, make it visible, and learn to use AI in a way that amplifies that. It should not replace it, it should amplify. Use it for the things that are genuinely tedious and low stake. Protect the work that makes you irreplaceable. That line is very different for everyone, and you have to draw it consciously, because the tool will not draw it, and it shouldn't. So the music act that opened this episode is basically a metaphor as much as it is news story. There are signals being sent as frequencies we are not always tuned to noticed. Some of them are shaping how we think, how we work, and who we are becoming as professionals, without us having consciously agreed to that. But the antidote for this is not less technology, the antidote is more intentionality because you are responsible for what you do with your knowledge, not others. I will say it again. You are responsible for what you do with your knowledge or lack of. Watch idiocracy and not as a joke. Watch it as a mirror. Then ask yourself, what am I doing personally and professionally to make sure that that's not the direction we are heading? I will say it again. Look into it and see what you are personally and professionally. How are you making sure that that is not the direction we are heading? Use AI. Absolutely. As you see from my from everything I do, I'm very prone using this tool. But use it in a way that makes you sharper. That builds your thinking, that makes your human in your work. Because the most important thing you bring to work right now is not the output you can generate with a good prompt, it is the intelligence you can exercise without one. Take this into consideration and please make sure to subscribe, to like, to share this with someone, because it's very important for us to reach more people and continue this conversation because every time you share with someone, and that someone might be talking with me, I can have this conversation. And I think it's very important for us to have this community and actually talk about it. But that is it for today. If this episode made you think or disagree, I'd love to hear from you. All the studies and articles mentioned are linked in the show notes below. Read them, the actual thing. It will take longer, but that is the point. See you next week.