Latest / AI Report / Bonus: EU AI-act uitgelegd door Google NotebookLM

Transcript

Bonus: EU AI-act uitgelegd door Google NotebookLM · AI Report

• 0:00
  [alarm] Oh! Zo klinken 165.000 scheerbewegingen per minuut die je helpen om fris voor de dag te komen.
• 0:10
  De flexibele scheerkop van de Philips S9000 Prestige volgt de contouren van je gezicht en scheert tot op nul millimeter van je huid.
• 0:18
  Ervaar een ultiem gladde en comfortabele scheerbeurt, ook wanneer je nog niet helemaal wakker bent. Want dankzij de slimme sensoren past het scheerapparaat zich helemaal aan jou aan.
• 0:29
  Zo kun jij nog extra genieten van je kopje koffie voor je de deur uit gaat. De Philips S9000 Prestige is het ideale scheerapparaat om jouw dag goed mee te beginnen.
• 0:40
  Koop de Philips S9000 Prestige bij jouw favoriete winkel. Koffietijd voor Mannen, de podcast waarin drie onbezonnen gasten je wekelijks een kijkje geven in een zinderende studentenleven.
• 0:51
  We gaan geen onderwerp uit de weg en helpen je op het gebied van liefde. Zou je huilen tijdens de seks een, een, een afknapper vinden? Vriendschap en alle andere problemen die je tegenkomt in je studentenleven.
• 1:00
  Ze miste een beetje vo. [gelach] Elke maandag om drie uur sharp op je favoriete podcastplatform. Ik heb met zoveel meiden getongd. Echt met vier verschillende. Ik zo: bam. Dat was vier keer dezelfde.
• 1:14
  [piep] We luisteren deze podcast op eigen risico. Okay, so get this. Today's deep dive is, uh- It's a big one. Yeah. We're diving headfirst into the EU's Artificial Intelligence Act. Okay. You know, that whole thing.
• 1:29
  Right. And let me tell you, these excerpts you guys sent over from their legislative text- Dense. This thing reads like, honestly- From Pegasus... someone mashed up a sci-fi novel with a law textbook.
• 1:40
  Well, it's groundbreaking stuff, right? It is. The EU's really, I think, trying to get out ahead of the curve on this whole regulating AI thing. Yeah. Which, as we know, is moving faster than- Oh, tell me about it...
• 1:52
  a caffeinated robot. It's true. This AI stuff is just like warp speed, but also it's everywhere. Right. Absolutely. I mean, we're way past just talking about self-driving cars and, you know, like, robots in factories.
• 2:05
  Right. So with this massive document, where do we even begin to unpack it all? Well, I think a good place to start is with the scope. Right. Like, what does the EU even consider an AI system?
• 2:18
  Because it's not as simple as you might think. Yeah, it can't just be like, if it runs on code, it's AI, can it? Definitely not.
• 2:24
  The act actually defines an AI system as one that's designed to operate with a degree of autonomy. Okay.
• 2:32
  Meaning it can make decisions or take actions to achieve objectives without a human having to, like, hold its hand the whole way. Okay, so some level of thinking for itself. Exactly. Got it.
• 2:42
  But where it starts to get really interesting to me is when they get into the examples. Okay. Like self-driving cars, sure, we get it. But then they mention AI for job recruitment. Yeah.
• 2:50
  And I was like, hold on, my resume's being analyzed by an algorithm now. Well, that's exactly it, right? Yeah. It, it highlights how broadly they're thinking about this.
• 2:59
  It's not just the AI that makes the headlines, it's the AI that's sort of silently seeping into all these everyday processes- Right... that affect all of us. Right.
• 3:07
  And speaking of hold on moments, they also specifically call out things like biometric categorization and emotion recognition systems.
• 3:16
  Okay, those both sound like they could go sideways really fast, depending on who's using them and how. Exactly. What's the EU's thinking there? Well, that's where their whole focus on trustworthy AI comes in. Okay.
• 3:26
  They're not just trying to prevent, you know, the Hollywood robot apocalypse scenario. Right.
• 3:31
  It's more about making sure AI is developed and used in a way that respects fundamental rights, democratic values, ethical principles. Mm-hmm. You know, all the important stuff.
• 3:42
  So it's about making sure AI serves us, not the other way around. Exactly. I like it. But how do they actually plan to do that? I mean, we can't just program a moral code into a computer, right?
• 3:52
  Not easily, that's for sure. And that's where it gets even more interesting. Okay. The act takes what they call a risk-based approach to regulating AI.
• 4:02
  Risk-based, so they're acknowledging that not all AI is created equal. Exactly. Okay. They're saying, hey, some AI applications are inherently riskier than others. Makes sense.
• 4:11
  And they actually go so far as to identify certain applications as high risk. High risk, meaning what? Meaning those with the potential to seriously impact people's health, their safety, their fundamental rights.
• 4:23
  We're talking AI used in healthcare, transportation, law enforcement areas where mistakes can have huge consequences. Okay, yeah, high risk AI, now we're getting into it. So give me some examples.
• 4:35
  What kind of AI applications get flagged with this, like, handle with extreme caution label? Well, I mean, some are obvious, right? Yeah. Like self-driving cars, medical devices. Right.
• 4:46
  But the act throws in some curveballs too. Yeah. Did you know that AI systems used for things like job recruitment, credit scoring, even some aspects of law enforcement could fall under this high risk category? Whoa.
• 5:00
  Okay, that is surprising. But I guess when you think about it, the potential for harm in those areas is huge. Absolutely. Imagine an AI system, like, denying someone a job or a loan. Right.
• 5:10
  Not because of their qualifications or their credit score, but because it was trained on, like, biased data that unfairly categorized them based on their background or where they live.
• 5:20
  Exactly, and that's exactly what the EU is trying to prevent. It's not that these applications are inherently bad or dangerous. Right. It's that they recognize the risks, and they wanna put safeguards in place.
• 5:29
  Safeguards. Music to my ears. [laughs] So what kind of safeguards are we talking about?
• 5:36
  Like, what does the act actually require from developers and those using these high-risk AI systems to prove they're playing by the rules? Well, once an AI system is classified as high risk, things get a lot stricter.
• 5:48
  First and foremost, data. Okay. The act stresses that these systems have to be trained on data that's relevant, representative, and as much as possible, free from errors and biases.
• 5:59
  So no more feeding AI algorithms a diet of junk food data and hoping they'll magically spit out healthy, unbiased decisions. Exactly. Got it.
• 6:07
  And it's not just about the quality of the data itselfIt's about how that data's handled.
• 6:12
  The act mandates transparency about how the data is collected, used, stored, 'cause you need to know where this information is coming from, how it's being processed, to even begin to ensure it's not being misused or manipulated.
• 6:26
  Transparency is key, especially when we're talking about decisions that can have a major impact on people's lives. Absolutely. And speaking of control, another crucial aspect is human oversight.
• 6:37
  The EU is very clear that humans need to be kept in the loop, especially when it comes to these high-risk AI systems. Okay, that's reassuring. So no Skynet taking over the world just yet. Yeah.
• 6:46
  Tell me more about this human oversight thing. It sounds crucial. It really is.
• 6:50
  Basically, the act is saying that for these high-risk systems, there needs to be a way for humans to understand how the AI is making its decisions, and crucially, to intervene or even override those decisions if necessary.
• 7:03
  So it's not about blindly trusting the algorithm, but rather having mechanisms in place to monitor and control it, like an emergency brake for AI. You got it.
• 7:14
  And this all ties back into the transparency requirements we talked about. Right, because what good is an emergency brake if you don't even know what the AI is doing or why it's doing it? Precisely.
• 7:23
  The act stresses that people have a right to know how these systems work, especially when those systems are making decisions that directly affect their lives.
• 7:33
  It's about empowering individuals, not making them feel like they're at the mercy of some, like, you know, inscrutable black box algorithm. Exactly. Okay. And it's not just about individual decisions either.
• 7:43
  The act also addresses things like cybersecurity. Oh, right, of course. Developers have to take steps to protect these high-risk systems from being hacked or manipulated. Yeah.
• 7:53
  We don't need AI systems going rogue because someone forgot to update their security protocols. Exactly. It's all connected. Okay. This is already making my head spin a bit.
• 8:01
  It's like trying to map out all the possible scenarios in a game of chess, but with AI instead of pawns and bishops. And we're just getting started, right? We're just getting started. Mm-hmm. This document is a beast.
• 8:11
  There's always more, right? [chuckles] Right. So we've been, like, really getting into the weeds on these high-risk systems. Yeah. But the act doesn't just list them out and call it a day. Yeah.
• 8:21
  It actually goes deeper into how they figure out what deserves that high risk label in the first place. Because it can't be as simple as someone just saying, "Well, that AI seems kinda sketchy.
• 8:31
  Let's slap a warning on it." You'd think so, wouldn't you? Right. But no, they have a whole system. The act lays out these specific criteria, takes a pretty multifaceted approach.
• 8:42
  One of the big ones, unsurprisingly, is the AI's intended purpose. Okay, so, like, what's it actually being used for? Exactly. Okay. Is it being used in a sector where the potential for harm is high?
• 8:54
  Like healthcare, transportation, law enforcement. Bingo. Areas where a mistake could have, like, really serious consequences. Right. You don't want a glitching AI performing brain surgery or, like, landing planes.
• 9:07
  Exactly. Nobody wants that. But it goes beyond just the field, too. They also consider the severity of that potential harm. Okay. Could this AI cause physical injury? Could it lead to financial ruin?
• 9:18
  What about infringing on someone's fundamental rights? So it's kind of a sliding scale then. Like, a slightly biased AI recommending products is bad. Right.
• 9:27
  But an AI misdiagnosing a patient is, like, whole other level of bad. Exactly. Orders of magnitude worse. And the act recognizes those nuances, which is pretty impressive.
• 9:38
  But here's where it gets even more interesting, and this might surprise you. They don't just look at the AI itself. Okay. They also factor in the context of how it's actually being deployed. Hold on.
• 9:50
  Are you saying the same AI could be high risk in one situation? Yes. But not another. Precisely. It's all about the context. Okay, give me an example, 'cause now my brain's doing a logic puzzle. Okay.
• 10:01
  So remember how we talked about AI in job recruitment? Yeah, and how it could be high risk, depending. Exactly. If that AI is just helping, say, sort through resumes based on keywords someone's looking for. Yeah.
• 10:11
  Probably pretty low risk, right? Makes sense. But if that same AI is the one making the final hiring decisions, no human involved. Oh, okay, yeah, I see where you're going with this. Suddenly, very different story.
• 10:23
  Same AI, potentially a vastly different impact depending on how much power it actually has. Exactly.
• 10:28
  It's not just about the tech itself, it's about how it's used, how it's integrated into a system, and how much autonomy it's given.
• 10:36
  And this really highlights a key theme throughout the entire act, which is human control and oversight. Right, especially with systems that can have such big consequences if they go wrong. Absolutely.
• 10:46
  But even with these criteria, right, there are always gonna be those gray areas, situations where it's not so clear-cut if something should be high risk or not.
• 10:54
  I'd imagine, especially since AI is, you know, evolving every single day. Mm. What might seem low risk today could be a huge concern tomorrow.
• 11:04
  Exactly, and that's why they built this really interesting concept into the act called regulatory sandboxes. Okay, yes, the regulatory sandboxes. We touched on that earlier. Yeah.
• 11:13
  That's where they basically get to test out these AI systems in a controlled environment, right? Exactly. Think of it like a safe space for experimentation. Developers can test their AI in real world scenarios. Okay.
• 11:25
  But within a structured framework. Mm-hmm. Under the watchful eye of the regulators. So it's like giving a new driver a learner's permit before handing them the keys to a race car. Right. Right?
• 11:35
  You wanna make sure they can handle the turns before letting them loose on the open road. Perfect analogy.
• 11:40
  And it ties directly into what the EU is trying to do here, which is foster innovation and progress in AI, but also making sure it's done responsibly and ethically. Finding that balance. Yeah.
• 11:50
  Okay, so speaking of responsibility, let's go back to those safeguards for the high risk systemsWhat are the key things that developers and providers actually have to do to prove they're taking this seriously?
• 12:02
  Well, it always comes back to data, doesn't it? Right. The act is incredibly clear high-risk AI must be trained on data that is relevant, representative, and as free from errors and biases as humanly possible.
• 12:17
  They are not playing around with this. It's like that old saying, "Garbage in, garbage out," but with much, much higher stakes. Exactly. And it's not just about the data itself, remember. It's the whole life cycle.
• 12:26
  Right. The act says you gotta be transparent about how you gather that data, how you store it, how you use it.
• 12:31
  So no more shady data harvesting practices, hoping no one will notice if your AI learned some bad habits from a biased data set. Nope. Not if you wanna operate in the EU anyway. The stakes are too high. Rightly so.
• 12:43
  Okay, so high quality data, transparency. Mm. What else? We talked about human oversight being a big deal. What does that actually look like in practice?
• 12:53
  So it's about building in mechanisms for human intervention and control. The act is very clear that for these high-risk systems, humans need a way to understand what the AI is doing and why. Right.
• 13:06
  And critically, there needs to be a way to override those decisions if necessary.
• 13:11
  So it's like having an off switch or a manual override in case the AI starts going rogue or making decisions that just don't align with, you know, our ethical standards. That's a good way to think about it.
• 13:21
  It's about shared control, not complete automation. And of course, this ties back into the transparency requirement. Right.
• 13:26
  Because what good is an off switch if you don't even know what the AI is doing or why it's doing it? Exactly. People have a right to know, especially when these systems are making decisions that affect their lives.
• 13:36
  It's about empowering individuals, not just blindly trusting these complex algorithms to make the right call every time. Absolutely. But here's the thing, the EU didn't stop at just high-risk systems.
• 13:48
  Remember those general purpose AI models we mentioned? Yeah. The ones that can be used for a whole bunch of different things. Yeah, like those AI image generators that everyone's going crazy about these days.
• 13:58
  Those seem pretty harmless. Mostly just good for fun or sparking creative ideas. On the surface, sure. But the EU recognizes that even those models, as fun as they seem, come with their own sets of risks.
• 14:10
  Okay, now you've got me curious. What kind of risks are we talking about here? Are we talking AI-generated deepfakes being used to spread misinformation or AI art infringing on copyright? You're getting warmer.
• 14:22
  Those are definitely concerns. But the EU's taking a broader perspective here, looking at the potential for what they call systemic risk. Systemic risk.
• 14:32
  That sounds kind of ominous, like something out of a disaster movie. It's not about being alarmist.
• 14:36
  It's more about recognizing that these powerful AI models, especially the ones with what the act calls high impact capabilities. Okay. Those could potentially disrupt entire systems, even entire markets, in theory.
• 14:50
  Okay, now we're talking about something with potentially huge consequences. Mm. Can you give me an example, though? Like, how could something like an AI image generator pose a systemic risk? Imagine this.
• 15:02
  What if these image generators get so good- Yeah... so advanced that they can create fake news articles or videos, social media posts that are totally indistinguishable from the real thing? Oh, wow. Yeah.
• 15:13
  Then it's not just a funny meme anymore, right? Right. That's like straight up recipe for chaos. It could undermine trust in all information online. Exactly. Potentially even influencing elections- Yeah...
• 15:25
  or inciting violence. It's like weaponized misinformation, but fueled by AI. It really is a whole new level, and that's exactly why the EU is taking this systemic risk thing so seriously. It's not about being anti-AI.
• 15:39
  It's about recognizing just how powerful these models are- Right... and putting some guardrails in place before they're used in ways that could, you know, really destabilize things on a large scale.
• 15:50
  So how do they plan to do that?
• 15:51
  I mean, do these general purpose AI models, even the ones with high impact capabilities, do they get hit with the same strict regulations as those high-risk systems we talked about earlier?
• 16:00
  It's not quite a one size fits all approach. The act does single out these general purpose models, especially the ones with that systemic risk factor- Yeah...
• 16:07
  for some extra attention, and a lot of it comes back to transparency. There it is again. Remember how we talked about how important the data is for those high-risk systems? Oh, absolutely.
• 16:16
  It's like the foundation, right? Right. The raw material that the AI learns from. Garbage in, garbage out. Exactly.
• 16:22
  Well, for these general purpose models that have that systemic risk potential, the act goes even further, demanding even more meticulous documentation. Okay. They wanna know where that training data came from- Right...
• 16:36
  how it was selected, what steps were taken to address any potential biases. So it's about accountability, basically, but on a much larger scale. Exactly.
• 16:46
  Because if you're training an AI model on, like, a massive data set scraped from who knows where on the internet, I mean, there's so much that could go wrong. A lot that can go wrong.
• 16:57
  The internet's amazing, but it's also, you know... It's a mess. [laughs] Yeah. It's full of misinformation and biases and even just straight up bad stuff.
• 17:05
  And these models are often trained on so much data that it's really tough to control for all of that. Right. But the EU is basically saying, "Hey, even if it's hard, you still gotta try." Okay.
• 17:16
  They're pushing these developers to be a lot more proactive about identifying and mitigating those risks from the get-go.
• 17:23
  So it's not enough to just say, "Oops, our AI model accidentally developed a prejudice against certain groups of people because of the data we fed it." Nope, not anymore.
• 17:32
  They're being held responsible for what their AI creations are learning and how they're using it in the real world. Exactly.
• 17:38
  The act really emphasizes that need for these comprehensive risk assessments, mitigation strategies in place, especially for these big, powerful, general purpose models.
• 17:48
  It's about thinking ahead, anticipating those problems, not just playing catch up after something blows up.
• 17:53
  It's almost like they're trying to write the rule book for a game that's constantly changing its own rules as it goes. Right.
• 18:00
  Because AI is moving so fast, how can you even predict what it'll be able to do a year from now, let alone five or 10 years down the line?But it sounds like the EU is at least making a good effort to stay ahead of the game.
• 18:10
  They're definitely putting their foot down and saying, "This is how we're gonna do things." The EU's AI Act is a big deal, a real landmark piece of legislation, and it's gonna be- Yeah...
• 18:20
  so interesting to watch how this all plays out in practice. But, you know, rules are only as good as their enforcement, right? That's what I was thinking. Yeah.
• 18:27
  We- we've talked about all these rules and requirements, but how does the EU actually make sure companies are actually following them? What happens if an AI system goes rogue, so to speak?
• 18:38
  Well, they have a whole system for that too. The act outlines this multi-layered approach to enforcement. On one level, you've got national authorities in each EU country.
• 18:49
  They're the ones on the ground doing that market surveillance, making sure the AI being used in their backyard is following the rules.
• 18:55
  Oh, they're like the AI police, making sure no one's sneaking in any illegal algorithms. Kind of. They have the authority to request information from providers, conduct audits. Mm.
• 19:04
  They can even order a non-compliant system to be taken off the market. Okay, so they've got some teeth. But what about those bigger general purpose models we've talked about? Mm.
• 19:13
  The ones with that systemic risk potential. I mean, those can have impacts way beyond just one single country. Right, and that's where the AI office comes in. Okay.
• 19:22
  So this is an EU level body that was specifically created to oversee how this act is implemented. Okay. They're there to provide guidance to those national authorities, make sure everyone's on the same page. Right.
• 19:34
  Monitor how the act is actually being applied across the EU, and they even have the power to investigate potential violations themselves, especially when it comes to those really high impact AI models.
• 19:45
  So the AI office is kinda like the AI Supreme Court, stepping in when the stakes are really high to make sure everyone across the EU is playing by the same rules. That's a great way to put it.
• 19:55
  They're really key for making sure there's consistency and cooperation among all the different national authorities, because the last thing you want is one country being super lax about the rules while another one is coming down hard on every little thing.
• 20:09
  Right. That would be a nightmare for companies that are trying to develop and actually use these AI systems across the EU. It needs to be a level playing field. But okay, what about consequences?
• 20:19
  We've talked about the rules, we've talked about who enforces them, but what happens if a company gets caught breaking the AI Act? Let's just say the EU means business when it comes to enforcement. Right.
• 20:30
  They've got a whole tiered system of fines for violations. Hit me with it. What are we talking about here? Well, for those, like, really minor infractions, the fines can be pretty small- Mm-hmm...
• 20:40
  relatively speaking, but for those more serious violations.
• 20:43
  Yeah, like if a company is caught using a prohibited AI practice or just totally disregarding the rules for high-risk systems, those fines can reach tens of millions of euros. Oof, that's gonna leave a mark.
• 20:55
  That's a serious motivator to get it right.
• 20:56
  It's designed to make you think twice, and for those companies that are providing those general purpose AI models, the ones with that huge systemic risk potential, those fines can go even higher.
• 21:07
  They're really sending a message, responsible AI development. This isn't optional anymore. This is the law. [lips smack] Wow. This deep dive has been quite the journey.
• 21:16
  We went from, like, what even is an AI system to the nitty-gritty of high-risk applications, regulatory sandboxes, systemic risk enforcement. My brain's a little fried.
• 21:28
  It's a lot to take in, for sure, but incredibly important. I mean, the EU's AI Act is groundbreaking, and it really makes you think about the future of AI, not just in Europe, but everywhere. It really does.
• 21:38
  And speaking of things to think about, we always like to leave our listeners with a final thought, something to ponder as they, you know, continue to navigate this crazy world that's increasingly powered by AI.
• 21:49
  What do you think? Well, if there's one thing to take away from all of this, it's that the EU keeps saying this isn't about trying to stop AI innovation. It's about making sure it's done responsibly. Right.
• 22:00
  So I would ask your listeners this. As AI gets more and more integrated into our lives, how do we make sure it reflects our values, not just the values of the people who happen to be building it?
• 22:11
  That's such a good question, and not an easy one to answer, but it's something we should all be thinking about. Hmm. The future of AI isn't something that just happens to us. It's something we're all building together.
• 22:22
  Well said. And on that note, we've reached the end of our deep dive into the EU's AI Act. Hopefully, you all found this insightful, informative, maybe even a little inspiring.
• 22:31
  Keep those questions coming, keep exploring, and keep diving deep. Until next time.