WEBVTT 00:00:00.000 --> 00:00:04.940 It feels like every single week, there's a new 00:00:04.940 --> 00:00:09.300 viral headline about artificial intelligence 00:00:09.300 --> 00:00:11.599 just making something up. Oh, absolutely. It's 00:00:11.599 --> 00:00:13.800 literally everywhere right now. Right. Like a 00:00:13.800 --> 00:00:15.859 chat bot invents a massive corporate scandal 00:00:15.859 --> 00:00:19.760 out of nowhere, or it confidently gives you detailed 00:00:19.760 --> 00:00:21.640 driving directions to a restaurant that actually 00:00:21.640 --> 00:00:23.780 burned down 10 years ago. Yeah, and it almost 00:00:23.780 --> 00:00:26.359 always comes packaged with this very specific, 00:00:26.839 --> 00:00:30.780 almost medical -sounding label. The media developers, 00:00:30.920 --> 00:00:32.799 pretty much everyone says the AI is hallucinating. 00:00:32.859 --> 00:00:34.619 Okay, let's unpack this. Yeah. Because we are 00:00:34.619 --> 00:00:36.259 going straight to the source today to figure 00:00:36.259 --> 00:00:38.020 out what that actually means for you. I really 00:00:38.020 --> 00:00:40.159 need to, yeah. We've pulled together a comprehensive 00:00:40.159 --> 00:00:43.340 Wikipedia article simply titled, Hallucination 00:00:43.340 --> 00:00:46.420 in Artificial Intelligence. And our mission for 00:00:46.420 --> 00:00:48.979 this deep dive is to demystify justified the 00:00:48.979 --> 00:00:51.039 actual mechanics of what is happening inside 00:00:51.039 --> 00:00:54.159 these machines when they lie to us. Right, the 00:00:54.159 --> 00:00:56.420 actual gears turning in the background. Exactly. 00:00:56.520 --> 00:01:00.259 We are going to explore the hilarious and sometimes 00:01:00.259 --> 00:01:02.359 incredibly high -stakes consequences of these 00:01:02.359 --> 00:01:04.659 glitches in the real world. And perhaps most 00:01:04.659 --> 00:01:07.340 surprisingly, we're going to look at why this 00:01:07.340 --> 00:01:10.480 exact same glitch, the very thing causing so 00:01:10.480 --> 00:01:12.840 much panic in the tech world, might actually 00:01:12.840 --> 00:01:15.959 be the key to Nobel Prize -winning science. which 00:01:15.959 --> 00:01:20.780 is just wild to think about. But when I hear 00:01:20.780 --> 00:01:24.359 the word hallucination, I instantly picture like 00:01:24.359 --> 00:01:27.219 a psychedelic drug trip. Sure. Yeah. Or a severe 00:01:27.219 --> 00:01:29.739 fever dream. Right. And framing it that way just 00:01:29.739 --> 00:01:32.159 feels completely wrong for a computer program 00:01:32.159 --> 00:01:35.129 running on silicon chips. It doesn't have a brain. 00:01:35.489 --> 00:01:37.489 It really doesn't. If I had to use an analogy, 00:01:38.189 --> 00:01:41.090 AI hallucination feels a lot less like a psychological 00:01:41.090 --> 00:01:44.049 break from reality and a lot more like that one 00:01:44.049 --> 00:01:45.969 overly confident friend you have at trivia night. 00:01:46.430 --> 00:01:48.489 Oh, everybody has that friend. Do you know the 00:01:48.489 --> 00:01:50.879 one? The friend who, rather than just admitting 00:01:50.879 --> 00:01:53.159 they don't know the capital of some obscure country, 00:01:53.659 --> 00:01:56.079 will just invent an entire historically plausible 00:01:56.079 --> 00:01:57.719 war just to sound like they know what they're 00:01:57.719 --> 00:01:59.700 talking about. That is actually a remarkably 00:01:59.700 --> 00:02:02.340 accurate way to think about the behavior. Yeah. 00:02:02.359 --> 00:02:04.379 And what's fascinating here is that the word 00:02:04.379 --> 00:02:06.659 hallucination wasn't always used to describe 00:02:06.659 --> 00:02:09.800 a machine making a mistake. Wait, really? It 00:02:09.800 --> 00:02:12.620 wasn't a negative thing? No, not at all. To really 00:02:12.620 --> 00:02:15.000 understand why AI does this, we have to look 00:02:15.000 --> 00:02:16.960 at the evolution of the term itself because it 00:02:16.960 --> 00:02:19.620 used to mean the exact opposite. How so? Well, 00:02:19.620 --> 00:02:22.000 if you go back to the 1980s in the field of computer 00:02:22.000 --> 00:02:25.280 vision, hallucination actually had a highly positive 00:02:25.280 --> 00:02:27.659 connotation. I mean, I would imagine hallucinating 00:02:27.659 --> 00:02:30.740 in computer vision means the machine is seeing 00:02:30.740 --> 00:02:33.340 things that aren't there, like obstacles in the 00:02:33.340 --> 00:02:36.520 road. How could that be positive? So think about 00:02:36.520 --> 00:02:39.919 a really blurry, heavily pixelated, low resolution 00:02:39.919 --> 00:02:43.039 photo of a human face. Maybe it's from like an 00:02:43.039 --> 00:02:44.800 old security camera. OK, yeah, super grainy. 00:02:44.960 --> 00:02:47.560 Exactly. Researchers developed algorithms to 00:02:47.560 --> 00:02:49.780 automatically add high -resolution details to 00:02:49.780 --> 00:02:52.219 that image. It was filling in the gaps to make 00:02:52.219 --> 00:02:54.400 it look sharp and realistic for human analysts. 00:02:54.639 --> 00:02:57.500 Oh, wow. So, like, enhancing it. Right. And they 00:02:57.500 --> 00:03:00.439 literally called this process face hallucination. 00:03:01.120 --> 00:03:03.159 The first documented use of the term in this 00:03:03.159 --> 00:03:05.800 way was in a PhD thesis by Eric Majolsnes back 00:03:05.800 --> 00:03:09.379 in 1986. 1986. That is way earlier than I expected. 00:03:09.599 --> 00:03:13.219 Yeah, it goes way back. The AI was mathematically 00:03:13.219 --> 00:03:16.659 synthesizing details, pixels, really that weren't 00:03:16.659 --> 00:03:19.680 strictly in the original data, but it was doing 00:03:19.680 --> 00:03:22.259 it in a helpful, desired way to reconstruct an 00:03:22.259 --> 00:03:24.520 image. So it was essentially an enhancement tool. 00:03:24.659 --> 00:03:27.419 It was generating helpful filler, not deceiving 00:03:27.419 --> 00:03:30.199 the user. Exactly. It was a feature, not a bug. 00:03:30.740 --> 00:03:33.379 OK, so when did the tech industry start using 00:03:33.379 --> 00:03:37.180 it to describe this modern negative failure mode 00:03:37.180 --> 00:03:40.000 that we see today? The semantic shift really 00:03:40.000 --> 00:03:43.080 began in the 2010s. Researchers who were working 00:03:43.080 --> 00:03:45.719 on statistical machine translation, you know, 00:03:45.919 --> 00:03:47.900 getting computers to translate languages, started 00:03:47.900 --> 00:03:50.560 noticing a weird phenomenon. What kind of phenomenon? 00:03:50.740 --> 00:03:52.900 Well, sometimes the models would just spit out 00:03:52.900 --> 00:03:55.419 factually incorrect or totally misleading outputs 00:03:55.419 --> 00:03:57.340 that had absolutely nothing to do with the source 00:03:57.340 --> 00:03:59.280 text. Just completely making up translations. 00:03:59.439 --> 00:04:02.360 Right. The math was prioritizing linguistic fluency 00:04:02.360 --> 00:04:05.259 over factual accuracy. It wanted to sound smooth 00:04:05.259 --> 00:04:07.240 more than it wanted to be right. That sounds 00:04:07.240 --> 00:04:10.139 exactly like the trivia friend. Exactly. And 00:04:10.139 --> 00:04:14.699 in 2015, computer scientist Andrej Karpathy explicitly 00:04:14.699 --> 00:04:17.339 used the term hallucinated to describe his language 00:04:17.339 --> 00:04:20.360 model generating a completely fake, yet perfectly 00:04:20.360 --> 00:04:23.980 formatted citation link. A fake link. Wow. Yeah. 00:04:24.399 --> 00:04:27.220 And then the term just absolutely exploded in 00:04:27.220 --> 00:04:29.699 popularity during the recent AI boom, especially 00:04:29.699 --> 00:04:33.560 after OpenAI released ChatGPT in late 2022. Right, 00:04:33.560 --> 00:04:35.600 because suddenly millions of people were experiencing 00:04:35.600 --> 00:04:38.439 these incredibly confident falsehoods firsthand. 00:04:38.680 --> 00:04:40.279 It wasn't just researchers anymore. Exactly. 00:04:40.279 --> 00:04:42.240 It was right there on your phone or your laptop. 00:04:42.480 --> 00:04:44.439 Reading through the source material though, it's 00:04:44.439 --> 00:04:46.300 pretty obvious that not everyone is happy with 00:04:46.300 --> 00:04:48.620 the tech industry adopting that word. There is 00:04:48.620 --> 00:04:50.720 a massive amount of pushback from the scientific 00:04:50.720 --> 00:04:52.779 community. Oh, there is a lot of pushback. Researchers 00:04:52.779 --> 00:04:55.180 like Mary Shaw and statistician Gary and Smith 00:04:55.180 --> 00:04:57.740 argue that using the word hallucination is deeply 00:04:57.740 --> 00:05:01.439 dangerous. Dangerous? Because it heavily anthropomorphizes 00:05:01.439 --> 00:05:04.040 the software. It forces us to treat a machine 00:05:04.040 --> 00:05:07.029 like it has a biological human mind. Which it 00:05:07.029 --> 00:05:09.350 definitely does not. Right. Smith points out 00:05:09.350 --> 00:05:12.529 that these large language models, or LLMs, do 00:05:12.529 --> 00:05:15.470 not actually understand what words mean. They 00:05:15.470 --> 00:05:17.509 don't have sensory perception. They're just math. 00:05:17.689 --> 00:05:20.389 Just math. So they physically cannot experience 00:05:20.389 --> 00:05:23.209 a hallucination in the perceptual, psychological 00:05:23.209 --> 00:05:25.550 sense that a human does. Honestly, calling it 00:05:25.550 --> 00:05:28.810 a hallucination seems like it gives the AI entirely 00:05:28.810 --> 00:05:31.850 too much credit. It almost feels like a PR spin. 00:05:32.139 --> 00:05:35.079 That is exactly the concern Mary Shaw raised. 00:05:35.339 --> 00:05:37.779 Because it lets developers off the hook for bad 00:05:37.779 --> 00:05:40.980 programming by framing a critical software error 00:05:40.980 --> 00:05:45.050 as this quirky psychological human flaw. I mean, 00:05:45.129 --> 00:05:47.730 it sounds a lot better to say my AI is hallucinating 00:05:47.730 --> 00:05:50.250 rather than my software is broken and producing 00:05:50.250 --> 00:05:53.009 garbage. It absolutely softens the blow by spinning 00:05:53.009 --> 00:05:56.430 objective errors into idiosyncratic quirks. And 00:05:56.430 --> 00:05:58.370 because of this danger, researchers have proposed 00:05:58.370 --> 00:06:00.610 several alternative terms in the scientific literature. 00:06:00.649 --> 00:06:02.550 Like what? What should we be calling it? Well, 00:06:02.589 --> 00:06:04.509 you'll see words like confabulation, delusion, 00:06:04.629 --> 00:06:06.889 or fabrication. But my personal favorite comes 00:06:06.889 --> 00:06:09.069 from a paper published in the journal Ethics 00:06:09.069 --> 00:06:11.850 and Information Technology by Hicks, Humphries, 00:06:11.990 --> 00:06:13.949 and Slater. Oh wait, I remember this from the 00:06:13.949 --> 00:06:16.759 reading. They used philosopher Harry Frankfurt's 00:06:16.759 --> 00:06:18.879 definition of bullshit, didn't they? They did. 00:06:18.939 --> 00:06:22.000 They literally defined AI outputs as bullshit. 00:06:22.459 --> 00:06:24.779 And they aren't just swearing for shock value, 00:06:24.819 --> 00:06:28.160 either. They are applying Frankfurt's rigorous 00:06:28.160 --> 00:06:31.459 philosophical framework. OK, I love this. Explain 00:06:31.459 --> 00:06:34.019 the framework for us. So Frankfurt argued that 00:06:34.019 --> 00:06:36.779 a liar is someone who knows the truth and actively 00:06:36.779 --> 00:06:38.860 tries to hide it. Right. There's an intent to 00:06:38.860 --> 00:06:41.959 deceive. Exactly. But a bullshitter is someone 00:06:41.959 --> 00:06:44.610 who is completely indifferent. to the truth. 00:06:44.629 --> 00:06:47.350 Oh, that makes so much sense. Right. The AI model 00:06:47.350 --> 00:06:50.329 doesn't care what is real and what isn't. It 00:06:50.329 --> 00:06:53.269 is just mathematically generating text. If a 00:06:53.269 --> 00:06:55.329 statement happens to be true, it is accidentally 00:06:55.329 --> 00:06:58.069 true. If it is false, it is accidentally false. 00:06:58.089 --> 00:07:00.209 So it's just the ultimate bullshit engine. That's 00:07:00.209 --> 00:07:02.430 exactly what it is. So if it's not a psychological 00:07:02.430 --> 00:07:05.310 hallucination and it's just indifferent mathematical 00:07:05.310 --> 00:07:07.889 text generation, we really need to look at the 00:07:07.889 --> 00:07:10.149 gears turning in the background. We do. If you 00:07:10.149 --> 00:07:14.100 take an LLM like GPT at its core, It's essentially 00:07:14.100 --> 00:07:17.019 just an incredibly complex next word predictor, 00:07:17.019 --> 00:07:20.779 right? That is essentially it. These models ingest 00:07:20.779 --> 00:07:23.120 a massive amount of text during their training. 00:07:23.720 --> 00:07:25.819 When you ask them a question, they aren't looking 00:07:25.819 --> 00:07:28.319 up facts in a database like a traditional search 00:07:28.319 --> 00:07:30.459 engine. They're not Googling the answer. No, 00:07:30.480 --> 00:07:33.160 not at all. They are calculating the statistical 00:07:33.160 --> 00:07:35.779 probability of what the next word should be in 00:07:35.779 --> 00:07:38.699 a sequence over and over again. It sounds a lot 00:07:38.699 --> 00:07:40.819 like the predictive text feature on a smartphone 00:07:40.819 --> 00:07:43.500 keyboard. Yes, that's a great comparison. Like 00:07:43.500 --> 00:07:46.019 when you start typing a text message, your phone 00:07:46.019 --> 00:07:48.779 suggests the next word to keep the sentence going. 00:07:48.649 --> 00:07:51.709 based on your previous habits and LLM is doing 00:07:51.709 --> 00:07:54.509 that but on a massive highly articulate scale. 00:07:54.850 --> 00:07:56.889 Very articulate. It just wants to keep the conversation 00:07:56.889 --> 00:07:59.730 going and it is heavily incentivized to give 00:07:59.730 --> 00:08:02.470 a guess even when it totally lacks the underlying 00:08:02.470 --> 00:08:05.040 factual information. And that incentive creates 00:08:05.040 --> 00:08:07.800 a fascinating tension inside the model between 00:08:07.800 --> 00:08:10.959 two competing goals, novelty and usefulness. 00:08:11.180 --> 00:08:13.199 Novelty and usefulness. OK, break that down. 00:08:13.480 --> 00:08:15.620 Researchers actually draw a parallel to human 00:08:15.620 --> 00:08:18.720 creativity here. If an AI system is focused entirely 00:08:18.720 --> 00:08:22.199 on usefulness, it will just regurgitate and memorize 00:08:22.199 --> 00:08:24.600 safe, boring content. So it would just sound 00:08:24.600 --> 00:08:27.639 like a sterile robotic textbook. Exactly. Super 00:08:27.639 --> 00:08:32.220 boring. But if the system prioritizes novelty, 00:08:33.100 --> 00:08:36.600 you know, trying to generate diverse, unique, 00:08:36.860 --> 00:08:40.100 natural -sounding text, it drastically increases 00:08:40.100 --> 00:08:42.580 the chances of generating original but entirely 00:08:42.580 --> 00:08:44.679 inaccurate responses. It makes things have to 00:08:44.679 --> 00:08:46.500 sound interesting. Yeah. But how does that actually 00:08:46.500 --> 00:08:48.700 work in practice? Like, how do developers force 00:08:48.700 --> 00:08:51.559 a math equation to be novel? They use decoding 00:08:51.559 --> 00:08:53.639 strategies. And the most common one is something 00:08:53.639 --> 00:08:55.960 called top -case sampling. Top -case sampling. 00:08:56.000 --> 00:08:58.799 Yeah. Normally, an AI would just pick the single 00:08:58.799 --> 00:09:02.120 most mathematically probable next word. Always. 00:09:02.259 --> 00:09:04.600 Which would be the boring textbook output. Right. 00:09:04.679 --> 00:09:06.779 So to make it sound more conversational and human, 00:09:07.379 --> 00:09:09.600 developers forced the AI to randomly select from 00:09:09.600 --> 00:09:12.159 a pool of the top K most likely words. Oh, I 00:09:12.159 --> 00:09:13.960 see. So instead of the number one word, it picks 00:09:13.960 --> 00:09:16.279 from a group. Exactly. Say the top 40 words. 00:09:16.879 --> 00:09:19.120 By injecting that literal dice roll into the 00:09:19.120 --> 00:09:21.639 selection process, the AI sounds vastly more 00:09:21.639 --> 00:09:23.799 dynamic. But I'm guessing that dice roll is where 00:09:23.799 --> 00:09:26.610 the problem starts. You guessed it. The data 00:09:26.610 --> 00:09:29.570 shows that this exact injection of randomness 00:09:29.570 --> 00:09:32.970 is positively correlated with increased hallucinations. 00:09:33.629 --> 00:09:35.970 It branches off into fiction. Because it's trying 00:09:35.970 --> 00:09:38.549 so hard to sound interesting and conversational 00:09:38.549 --> 00:09:41.269 that it physically overrides its own accuracy. 00:09:41.490 --> 00:09:44.110 Exactly. But do we actually know what is happening 00:09:44.110 --> 00:09:46.950 on a microscopic level inside the neural network 00:09:46.950 --> 00:09:49.289 when it decides to make something up? Like, can 00:09:49.289 --> 00:09:51.149 we see it happening? We are actually getting 00:09:51.149 --> 00:09:53.070 much closer to understanding that microscopic 00:09:53.070 --> 00:09:56.980 level. In 2025, the AI company Anthropic published 00:09:56.980 --> 00:09:59.580 some groundbreaking interpretability research 00:09:59.580 --> 00:10:02.820 on their LLM clod. Interpretability research, 00:10:02.840 --> 00:10:04.940 meaning they are trying to interpret the black 00:10:04.940 --> 00:10:08.200 box. Yes. They essentially took a flashlight 00:10:08.200 --> 00:10:10.419 under the hood of the neural network. They weren't 00:10:10.419 --> 00:10:13.179 just guessing about probabilities anymore. They 00:10:13.179 --> 00:10:16.019 mapped millions of artificial neurons to find 00:10:16.019 --> 00:10:18.940 specific internal circuits. Wow. And what did 00:10:18.940 --> 00:10:21.629 they find? They found a specific circuit designed 00:10:21.629 --> 00:10:24.669 to inhibit the AI, to physically stop it from 00:10:24.669 --> 00:10:26.629 answering a question if it doesn't have the factual 00:10:26.629 --> 00:10:28.669 data. So there's an actual, I don't know, circuit. 00:10:29.070 --> 00:10:31.710 Basically, yeah. By default, the circuit is active, 00:10:31.870 --> 00:10:35.289 meaning the AI stays quiet. But when the AI determines 00:10:35.289 --> 00:10:38.269 it has enough information, that circuit is inhibited, 00:10:38.470 --> 00:10:41.110 and it speaks. Okay, so a hallucination must 00:10:41.110 --> 00:10:43.730 be a misfire of that exact inhibition circuit, 00:10:43.730 --> 00:10:46.889 right? Precisely. They found that if Claude recognizes 00:10:46.889 --> 00:10:50.330 a specific token, like say the name of a somewhat 00:10:50.330 --> 00:10:52.809 famous person, but it doesn't actually have sufficient 00:10:52.809 --> 00:10:55.570 data about that person's life, the circuit incorrectly 00:10:55.570 --> 00:10:58.669 inhibits itself anyway. Oh no, so it gets overconfident. 00:10:58.769 --> 00:11:01.809 Yeah. The AI essentially says, oh, I recognize 00:11:01.809 --> 00:11:04.309 that name. I can definitely answer this. It completely 00:11:04.309 --> 00:11:07.470 fails to check if the connected factual tokens 00:11:07.470 --> 00:11:10.549 are actually present and just proceeds to generate 00:11:10.549 --> 00:11:13.730 a highly plausible, completely untrue biography. 00:11:13.909 --> 00:11:16.210 Just totally wings it. Yeah. Knowing those mechanics 00:11:16.210 --> 00:11:18.669 makes the bizarre real world outputs we see make 00:11:18.669 --> 00:11:20.950 so much more sense. It really does. If the AI 00:11:20.950 --> 00:11:23.470 is constantly playing this tug of war between 00:11:23.470 --> 00:11:25.889 being mathematically safe and randomly selecting 00:11:25.889 --> 00:11:27.929 words to sound creative, you are going to get 00:11:27.929 --> 00:11:30.759 some wild swings. It truly is the Wild West of 00:11:30.759 --> 00:11:33.240 AI outputs right now. That tension produces some 00:11:33.240 --> 00:11:36.159 incredible artifacts. Some of the examples in 00:11:36.159 --> 00:11:38.879 the source material are purely hilarious. There 00:11:38.879 --> 00:11:42.240 was this data scientist named Teresa Kubaka who 00:11:42.240 --> 00:11:45.679 literally made up a gibberish phrase. Oh, the 00:11:45.679 --> 00:11:49.220 cycloidal inverted electromagnet. Yes, the cycloidal 00:11:49.220 --> 00:11:52.240 inverted electromagnet. She asked ChatGPT to 00:11:52.240 --> 00:11:54.559 explain it. Yeah. And instead of the inhibition 00:11:54.559 --> 00:11:56.840 circuit firing and the AI saying, hey, that's 00:11:56.840 --> 00:12:00.620 not real, the AI invented a deeply scientific, 00:12:01.039 --> 00:12:03.860 plausible sounding answer. Complete with fake 00:12:03.860 --> 00:12:07.250 academic citations, right? Yes. It sounded so 00:12:07.250 --> 00:12:09.269 authentic, she actually had to double check to 00:12:09.269 --> 00:12:11.110 make sure she hadn't accidentally typed the name 00:12:11.110 --> 00:12:13.629 of a real physics phenomenon. It's a perfect 00:12:13.629 --> 00:12:16.370 illustration of the model prioritizing conversational 00:12:16.370 --> 00:12:18.509 flow and pattern completion over fact checking. 00:12:19.029 --> 00:12:20.909 It saw a prompt that looked like a science question, 00:12:20.929 --> 00:12:22.950 so it generated a response that looked like a 00:12:22.950 --> 00:12:25.529 science answer. Exactly. There was also the time 00:12:25.529 --> 00:12:27.789 someone asked for proof that dinosaurs built 00:12:27.789 --> 00:12:30.450 a civilization. I remember that one. The AI just 00:12:30.450 --> 00:12:32.409 rolled with it. It claimed there were fossil 00:12:32.409 --> 00:12:35.210 remains of dinosaur tools and that they created 00:12:35.370 --> 00:12:39.649 engravings on stones. Which is amazing. But my 00:12:39.649 --> 00:12:41.850 absolute favorite example from the sources, and 00:12:41.850 --> 00:12:45.389 you're gonna love this, a prompt convinced Chet 00:12:45.389 --> 00:12:49.309 GPT that churros, like the fried dough pastries, 00:12:49.789 --> 00:12:53.309 are ideal cools for home surgery. Home surgery 00:12:53.309 --> 00:12:56.889 with a churro. Yes, the AI boldly claimed there 00:12:56.889 --> 00:12:59.049 was a study published in the prestigious journal 00:12:59.049 --> 00:13:01.649 Science proving that churro dough is pliable 00:13:01.649 --> 00:13:03.970 enough to form surgical instruments. Oh my god. 00:13:04.590 --> 00:13:06.389 You've been added that the cinnamon sugar flavor 00:13:06.389 --> 00:13:08.870 has a calming effect on patients. That is just 00:13:08.870 --> 00:13:11.250 incredible. The churro surgery is funny, sure, 00:13:11.450 --> 00:13:13.889 but it perfectly illustrates a really dangerous 00:13:13.889 --> 00:13:17.350 mechanical flaw called a cascaded error. A cascaded 00:13:17.350 --> 00:13:19.309 error. Let's dig into that because it generates 00:13:19.309 --> 00:13:21.629 text one word at a time, right? Right. So once 00:13:21.629 --> 00:13:23.889 it predicts that first false word about the churro 00:13:23.889 --> 00:13:26.110 being a medical tool, it essentially poisons 00:13:26.110 --> 00:13:28.929 its own well. Yeah. It has to use its own generated 00:13:28.929 --> 00:13:31.309 falsehood as the context for the next word. Yes, 00:13:31.309 --> 00:13:33.210 exactly. It locks itself into the narrative. 00:13:33.799 --> 00:13:36.360 Once it affirms the absurd premise of the prompt, 00:13:36.720 --> 00:13:39.120 it is mathematically forced to double down on 00:13:39.120 --> 00:13:42.740 the lie to maintain consistency. Ah, I see. That 00:13:42.740 --> 00:13:45.820 is exactly how it ends up inventing fake science 00:13:45.820 --> 00:13:48.799 journal articles. It needs evidence to support 00:13:48.799 --> 00:13:51.360 the initial random hallucination it just made. 00:13:51.539 --> 00:13:53.539 But here's why this actually matters to you, 00:13:53.539 --> 00:13:57.120 the listener. We can laugh about dinosaur civilizations 00:13:57.120 --> 00:13:59.820 and pastry surgery all day. Right. It's entertaining. 00:13:59.980 --> 00:14:02.980 But if these models are fundamentally guessing… 00:14:03.100 --> 00:14:05.779 And we know they are prone to these cascaded 00:14:05.779 --> 00:14:08.860 errors where they trap themselves in lies. Why 00:14:08.860 --> 00:14:11.740 are highly paid professionals blindly trusting 00:14:11.740 --> 00:14:14.960 them with critical high stakes tasks? That is 00:14:14.960 --> 00:14:17.370 the multi -million dollar question. And the real 00:14:17.370 --> 00:14:19.309 world consequences have already been severe. 00:14:19.389 --> 00:14:21.970 We are seeing major professional disasters across 00:14:21.970 --> 00:14:24.330 multiple industries. Yeah, the sources had some 00:14:24.330 --> 00:14:26.649 wild examples. Take the airline industry, for 00:14:26.649 --> 00:14:30.190 example. In February 2024, Air Canada was taken 00:14:30.190 --> 00:14:32.769 to a civil resolution tribunal. Because of their 00:14:32.769 --> 00:14:35.289 chatbot, right? Exactly. Their customer support 00:14:35.289 --> 00:14:38.009 chatbot had hallucinated a bereavement fare policy. 00:14:38.549 --> 00:14:41.029 It told a grieving customer they could retroactively 00:14:41.029 --> 00:14:43.490 request a bereavement discount within 90 days 00:14:43.490 --> 00:14:45.649 of ticketing. Which was just entirely false. 00:14:46.269 --> 00:14:49.009 And Air Canada actually tried to argue in court 00:14:49.009 --> 00:14:52.309 that the chatbot was a separate legal entity 00:14:52.309 --> 00:14:55.230 responsible for its own actions, which is just 00:14:55.230 --> 00:14:58.649 a wild legal defense. It really is. And the tribunal 00:14:58.649 --> 00:15:01.129 completely rejected that defense, ordering the 00:15:01.129 --> 00:15:03.629 airline to pay damages and honor the hallucinated 00:15:03.629 --> 00:15:07.169 policy. The precedent is set. The company is 00:15:07.169 --> 00:15:09.289 responsible for the text its tools generate. 00:15:10.090 --> 00:15:12.950 Then there's the legal world, the 2023 case of 00:15:12.950 --> 00:15:16.240 Mata v. Avianca. Oh, that was a huge deal. A 00:15:16.240 --> 00:15:18.659 lawyer named Steven Schwartz used ChatGPT to 00:15:18.659 --> 00:15:21.299 write a legal brief, and he submitted six fake 00:15:21.299 --> 00:15:23.779 case precedents that the AI completely fabricated. 00:15:23.960 --> 00:15:26.039 Sex of them. Yeah. Schwartz claimed he didn't 00:15:26.039 --> 00:15:28.860 know ChatGPT could make things up. He even noted 00:15:28.860 --> 00:15:31.159 that when he asked the AI if the cases were real, 00:15:31.580 --> 00:15:33.870 the AI assured him they were. Which is another 00:15:33.870 --> 00:15:36.049 textbook example of the cascaded error. You ask 00:15:36.049 --> 00:15:38.149 the bullshitter if it is lying, and it mathematically 00:15:38.149 --> 00:15:41.289 doubles down on the lie. Exactly. Judge P. Kevin 00:15:41.289 --> 00:15:44.309 Kestel ultimately dismissed the case, fined Schwartz 00:15:44.309 --> 00:15:47.629 $5 ,000, and described the A .I.'s legal opinions 00:15:47.629 --> 00:15:50.539 as gibberish that bordered on nonsensical. We 00:15:50.539 --> 00:15:53.019 have also seen massive financial waste in the 00:15:53.019 --> 00:15:55.840 consulting world. In late 2025, the Australian 00:15:55.840 --> 00:16:00.080 government received a $440 ,000 report from Deloitte 00:16:00.080 --> 00:16:02.259 that contained hallucinated academic sources 00:16:02.259 --> 00:16:05.860 and fake court quotes. Wow, $440 ,000 for fake 00:16:05.860 --> 00:16:09.000 quotes. Yeah. And just a month later, a $1 .6 00:16:09.000 --> 00:16:11.539 million Canadian health care report, also by 00:16:11.539 --> 00:16:14.039 Deloitte, was found to contain at least four 00:16:14.039 --> 00:16:16.659 false citations to non -existent research papers. 00:16:16.899 --> 00:16:18.759 The sources also highlight the friction these 00:16:18.759 --> 00:16:21.190 outputs cause. across other domains. There was 00:16:21.190 --> 00:16:23.690 a 2023 defamation lawsuit where a gun rights 00:16:23.690 --> 00:16:26.669 activist, Mark Walters, sued OpenAI after Chad 00:16:26.669 --> 00:16:28.929 GPT falsely claimed he was accused of embezzlement. 00:16:29.250 --> 00:16:31.610 Right. And for context, in 2025, a judge ruled 00:16:31.610 --> 00:16:34.330 in favor of OpenAI in that specific case. Yes. 00:16:34.710 --> 00:16:36.710 There was also an instance where Google's Gemini 00:16:36.710 --> 00:16:39.409 generated images of racially diverse 1940s German 00:16:39.409 --> 00:16:42.350 soldiers, which led Google to temporarily pause 00:16:42.350 --> 00:16:44.429 the tool's image generation of people entirely. 00:16:44.629 --> 00:16:46.970 Yeah. That made a lot of headlines. And we mentioned 00:16:46.970 --> 00:16:49.070 these strictly to show the broad spectrum of 00:16:48.840 --> 00:16:52.820 areas, from corporate consulting to law to historical 00:16:52.820 --> 00:16:56.679 image generation, where unpredictable probabilistic 00:16:56.679 --> 00:16:59.779 outputs clash with our fundamental need for factual 00:16:59.779 --> 00:17:02.399 accuracy. It's a clash that is happening everywhere. 00:17:02.600 --> 00:17:05.359 We have clearly seen the damage this confabulation 00:17:05.359 --> 00:17:08.259 causes when we need hard facts. But if we pivot 00:17:08.259 --> 00:17:11.359 and look at the broader landscape... what happens 00:17:11.359 --> 00:17:13.640 when we want the AI to make things up? Right, 00:17:13.660 --> 00:17:16.019 if we actually want it to invent. Exactly. If 00:17:16.019 --> 00:17:18.299 we connect this to the bigger picture, the line 00:17:18.299 --> 00:17:21.500 between a hallucination and a scientific discovery 00:17:21.500 --> 00:17:24.079 can actually be incredibly thin. It really can. 00:17:24.400 --> 00:17:27.039 This part blew my mind. The sources detail the 00:17:27.039 --> 00:17:30.319 2024 Nobel Prize in Chemistry. David Baker's 00:17:30.319 --> 00:17:32.759 lab at the University of Washington intentionally 00:17:32.759 --> 00:17:36.339 used AI hallucinations to design 10 million brand 00:17:36.339 --> 00:17:39.140 new proteins that do not exist anywhere in nature. 00:17:39.380 --> 00:17:41.920 10 million? That is just staggering. Nature only 00:17:41.920 --> 00:17:44.259 created a fraction of that over millions of years. 00:17:44.619 --> 00:17:46.839 This led to roughly 100 patents and the founding 00:17:46.839 --> 00:17:49.779 of over 20 biotech companies. And the Nobel committee 00:17:49.779 --> 00:17:52.380 specifically rewarded this work, though notably 00:17:52.380 --> 00:17:54.720 they avoided using the word hallucination. They 00:17:54.720 --> 00:17:57.420 did. They opted to call it imaginative protein 00:17:57.420 --> 00:18:00.279 creation. Which is brilliant, framing by the 00:18:00.279 --> 00:18:02.440 committee, honestly. And it is not just proteins 00:18:02.440 --> 00:18:05.019 either. At Caltech, researchers deliberately 00:18:05.019 --> 00:18:07.880 use AI hallucinations to design a novel catheter 00:18:07.880 --> 00:18:10.559 geometry. A catheter geometry. Okay, what did 00:18:10.559 --> 00:18:13.579 the AI come up with? The AI hallucinated a design 00:18:13.579 --> 00:18:16.180 featuring sawtooth -like spikes on the inner 00:18:16.180 --> 00:18:18.359 walls. And it turns out when they physically 00:18:18.359 --> 00:18:21.019 tested it, those microstopic spikes actually 00:18:21.019 --> 00:18:24.000 prevent bacteria from gaining traction. Oh, wow. 00:18:24.000 --> 00:18:26.940 So it stops infections. Yes. That hallucination 00:18:26.940 --> 00:18:29.440 has the potential to fight urinary tract infections 00:18:29.440 --> 00:18:32.519 globally. We see similar methods in meteorology, 00:18:32.740 --> 00:18:35.700 too, where scientists use AI to generate thousands 00:18:35.700 --> 00:18:38.660 of subtle forecast variations to discover unexpected 00:18:38.660 --> 00:18:40.799 weather factors. So it essentially boils down 00:18:40.799 --> 00:18:43.680 to the between brainstorming mode and fact -checking 00:18:43.680 --> 00:18:45.819 mode. Exactly. If I want to write a legal brief, 00:18:46.119 --> 00:18:48.859 an AI hallucination is a disaster. It's literally 00:18:48.859 --> 00:18:52.000 perjury. But if I am in a chemistry lab trying 00:18:52.000 --> 00:18:54.680 to invent a new medicine, an AI hallucination 00:18:54.680 --> 00:18:58.160 is just a really creative, highly advanced hypothesis. 00:18:58.559 --> 00:19:01.299 That is a crucial distinction. The scientific 00:19:01.299 --> 00:19:03.700 applications of hallucinations are fundamentally 00:19:03.700 --> 00:19:06.579 different from chatbot hallucinations. How so? 00:19:06.799 --> 00:19:10.160 Because of the data? Yes. Chatbots operate on 00:19:10.160 --> 00:19:12.819 messy, ambiguous human language pulled from the 00:19:12.819 --> 00:19:16.180 internet. But the AI models used by these scientists 00:19:16.180 --> 00:19:19.039 are, as Caltech professor Anima Anankumar points 00:19:19.039 --> 00:19:22.859 out, taught physics. Ah, okay. So they have real 00:19:22.859 --> 00:19:25.160 -world constraints. Right. Their hallucinations 00:19:25.160 --> 00:19:27.660 are strictly grounded in physical reality, and 00:19:27.660 --> 00:19:30.880 they are always validated through rigorous, real 00:19:30.880 --> 00:19:33.359 -world laboratory testing. Which brings us right 00:19:33.359 --> 00:19:36.349 back to the everyday user. Since hallucinations 00:19:36.349 --> 00:19:39.490 can be both revolutionary in a lab and incredibly 00:19:39.490 --> 00:19:41.670 dangerous in a courtroom or a business report, 00:19:42.049 --> 00:19:43.930 how do we rein them in for the rest of us? That 00:19:43.930 --> 00:19:46.069 is the big challenge right now. Especially in 00:19:46.069 --> 00:19:48.470 places like schools, where we need the AI to 00:19:48.470 --> 00:19:51.210 be a tool, not a cheat code, without breaking 00:19:51.210 --> 00:19:53.579 the system entirely. The classroom crisis is 00:19:53.579 --> 00:19:56.759 very, very real. A 2024 study at the University 00:19:56.759 --> 00:20:00.319 of Mississippi found that 47 % of AI generated 00:20:00.319 --> 00:20:03.079 citations submitted by students were completely 00:20:03.079 --> 00:20:05.859 or partially fabricated. It's 47%. That's almost 00:20:05.859 --> 00:20:08.559 half. Yeah. Featuring incorrect titles, wrong 00:20:08.559 --> 00:20:12.019 dates, or totally fake authors. It is causing 00:20:12.019 --> 00:20:14.940 massive academic integrity concerns globally. 00:20:15.079 --> 00:20:17.160 And we can't just rely on AI detectors to catch 00:20:17.160 --> 00:20:20.569 the fabricated text. Right. Definitely not. Because 00:20:20.569 --> 00:20:23.529 the Wikipedia article notes that tools like Turnitin 00:20:23.529 --> 00:20:26.609 frequently flag innocent papers written entirely 00:20:26.609 --> 00:20:30.089 by human students. The detection technology is 00:20:30.089 --> 00:20:33.630 so fundamentally flawed that OpenAI completely 00:20:33.630 --> 00:20:36.609 shut down its own AI detection software due to 00:20:36.609 --> 00:20:39.130 a lack of accuracy. They did. It just wasn't 00:20:39.130 --> 00:20:41.890 reliable enough. So how are developers actually 00:20:41.890 --> 00:20:44.829 trying to mitigate the hallucinations at the 00:20:44.829 --> 00:20:47.130 source? If we can't detect it, how do we stop 00:20:47.130 --> 00:20:49.609 it? There are several highly technical methods 00:20:49.609 --> 00:20:51.769 being tested right now. Data cleaning is the 00:20:51.769 --> 00:20:53.730 most obvious one. Just giving you better data. 00:20:53.890 --> 00:20:55.730 Right, ensuring the model isn't learning from 00:20:55.730 --> 00:20:57.609 bad information in the first place. But there 00:20:57.609 --> 00:21:00.250 are also some really fascinating structural approaches. 00:21:00.650 --> 00:21:02.509 Researchers are setting up systems where different 00:21:02.509 --> 00:21:04.529 chatbots debate one another. Wait, they have 00:21:04.529 --> 00:21:07.930 them argue? Yeah. Model A generates an answer. 00:21:08.140 --> 00:21:10.720 Model B critiques the logic and looks for flaws, 00:21:11.099 --> 00:21:13.500 and then Model A revises its answer based on 00:21:13.500 --> 00:21:16.039 the critique, repeating until they reach a consensus. 00:21:16.279 --> 00:21:18.880 That's actually brilliant. It is. Another method 00:21:18.880 --> 00:21:21.920 forces the AI to actively validate its own low 00:21:21.920 --> 00:21:24.519 -confidence answers by querying the Bing search 00:21:24.519 --> 00:21:27.819 API, essentially making the AI Google its own 00:21:27.819 --> 00:21:30.099 thoughts before responding to the user. But there 00:21:30.099 --> 00:21:32.599 has to be a massive computational catch to all 00:21:32.599 --> 00:21:34.680 those fixes. Oh, there is. Every time you ask 00:21:34.680 --> 00:21:37.539 a question, if the AI has to generate three different 00:21:37.420 --> 00:21:40.380 drafts, have two sub -models argue about the 00:21:40.380 --> 00:21:43.980 logic, and then do a live web search. You have 00:21:43.980 --> 00:21:46.519 just tripled or quadrupled the electricity and 00:21:46.519 --> 00:21:48.500 processing power required for a simple prompt. 00:21:48.640 --> 00:21:50.259 That is the ultimate trade -off in the industry 00:21:50.259 --> 00:21:52.720 right now. Evaluating multiple possible replies 00:21:52.720 --> 00:21:55.019 or checking search engines drastically increases 00:21:55.019 --> 00:21:57.579 the computational cost and the financial cost 00:21:57.579 --> 00:21:59.819 of every single query. And is that worth it? 00:22:00.099 --> 00:22:02.299 In high -stakes fields like medical diagnostics 00:22:02.299 --> 00:22:04.480 or chip design, yes, that extra time and money 00:22:04.480 --> 00:22:07.279 is definitely worth it. But for a general consumer 00:22:07.279 --> 00:22:10.259 chatbot, the data shows an uncomfortable truth. 00:22:10.599 --> 00:22:13.920 What's the truth? Customers actually prefer rapid, 00:22:14.279 --> 00:22:18.440 overconfident answers over slow, cautious, uncertainty 00:22:18.440 --> 00:22:21.420 -aware ones. Oh, wow. So are the tech companies 00:22:21.420 --> 00:22:24.839 failing to fix the problem or are we, the users, 00:22:25.019 --> 00:22:27.779 the actual problem? Because our behavior rewards 00:22:27.779 --> 00:22:31.319 the AI for being a fast, confident liar rather 00:22:31.319 --> 00:22:34.369 than a slow, cautious truth -teller. This raises 00:22:34.369 --> 00:22:36.509 an important question about the nature of the 00:22:36.509 --> 00:22:39.349 technology itself. Many top researchers propose 00:22:39.349 --> 00:22:41.910 that hallucinations aren't just a bug we can 00:22:41.910 --> 00:22:44.269 eventually patch with a software update. They 00:22:44.269 --> 00:22:46.809 think it's permanent. They believe hallucination 00:22:46.809 --> 00:22:49.630 is an innate, unavoidable limitation of large 00:22:49.630 --> 00:22:52.130 language models. Because of how they are built 00:22:52.130 --> 00:22:54.849 as probabilistic prediction engines, they will 00:22:54.849 --> 00:22:57.029 always have the mathematical capacity to make 00:22:57.029 --> 00:22:59.529 things up. So as you continue to use these generative 00:22:59.529 --> 00:23:01.529 tools in your own work, your studies, or just 00:23:01.529 --> 00:23:04.359 your daily life, This approach is to completely 00:23:04.359 --> 00:23:07.700 stop treating the AI like an omission, all -knowing 00:23:07.700 --> 00:23:09.799 oracle. Absolutely. It is not a search engine. 00:23:10.240 --> 00:23:13.359 Instead, treat it like a brilliant, incredibly 00:23:13.359 --> 00:23:17.700 fast, but over -eager intern. An intern who really 00:23:17.700 --> 00:23:20.279 wants to please you, who can brainstorm amazing 00:23:20.279 --> 00:23:23.559 ideas, but who requires heavy, constant supervision 00:23:23.559 --> 00:23:26.140 and rigorous fact checking before you publish 00:23:26.140 --> 00:23:28.190 their work. I think that is the perfect mindset 00:23:28.190 --> 00:23:30.769 to have and it leaves us with one final deeply 00:23:30.769 --> 00:23:33.329 philosophical thought to mull over. Let's hear 00:23:33.329 --> 00:23:37.460 it. If AI hallucination is, at its core, just 00:23:37.460 --> 00:23:40.339 the mathematical equivalent of imagination, is 00:23:40.339 --> 00:23:43.319 it even possible to cure AI of hallucinating 00:23:43.319 --> 00:23:45.400 without completely destroying its ability to 00:23:45.400 --> 00:23:47.619 be creative? Oh, that's a wild thought. Think 00:23:47.619 --> 00:23:49.660 back to your overly confident trivia friend. 00:23:49.980 --> 00:23:52.119 If we somehow build a machine that is physically 00:23:52.119 --> 00:23:55.200 incapable of ever telling a lie, do we inadvertently 00:23:55.200 --> 00:23:57.759 build a machine that is incapable of ever inventing 00:23:57.759 --> 00:23:59.920 anything new? Definitely something to think about 00:23:59.920 --> 00:24:01.700 the next time your chatbot tries to convince 00:24:01.700 --> 00:24:04.099 you to do surgery with a churro. Thanks for joining 00:24:04.099 --> 00:24:05.940 us on this deep dive. We'll see you next time.