Learn how to detect AI generated content. This will help you avoid a very common vector of scams.
Files related to this podcast:
https://www.icloud.com/iclouddrive/0cbZpqjp4ycup-C93ZXurW5zA#Detect_AI_-_Links_Document
https://pmug-nj.org/manage/content/f/id/14
Video version of this podcast:
https://pmug-nj.org/manage/video/p/id/75
Transcript:
Aric • 00:01
This is the May 2026 PMUG podcast. Today we’ll be discussing detecting and dealing with AI. Let’s get started. Okay, welcome everyone to tonight’s PMUG meeting. Michael, would you like to give your spiel tonight?
Michael • 00:27
No, I gave in the past and it hasn’t helped.
Aric • 00:31
Fair enough. Well, tonight we’re going to be talking about AI. It’s very many uses, both good and bad, and ways in which you can detect AI use so that you don’t fall for scams. So let me go over here and share my screen. Allow. Okay. Share that one and share sound and then share. Okay. So now hopefully you should be able to see detecting AI on the screen there.
Michael • 01:09
We do.
Aric • 01:10
Okay, excellent. I am going to just switch over to my notes. And then I will post in the chat. There’s a link to a PDF document. I recommend that you all download it and take a look because there’s some – where is the chat? All right. I’m going to stop sharing my screen for a moment. Oh, there we go. Okay. They move stuff around in Zoom and I always forget where the stuff is. Okay, so now I am going to go to the chat room and paste in some links. Okay, the first one is the PDF document. It contains all the links that we’re gonna talk about tonight so that you don’t have to write anything down. It also contains the three samples of writing that we’re going to talk about later. So I recommend having it on hand so you can refer to it later. that is a list, a link to a compressed version of the media files that I’m going to be using tonight. So if you want to review them yourself, you can do so. You just have to be logged in as a member in order to be able to download that stuff. The first one’s available to anybody. Okay, there we go. I’ll put this over here. Okay, so tonight’s presentation will contain links and examples of AI use. It’s important that you see how it’s being used. Because AI use and misuse meets people where they are, some of these links may make reference to a variety of potentially hot button topics, though I’ve done my best to pick sources and subjects that just stick to a discussion of the facts around AI use, but just a note there. So what is AI? Well, what we refer to as AI today isn’t what we would have called AI, artificial intelligence, even 10 years ago. What we call AI now is essentially just a predictive algorithm or multiple algorithms that are capable of some surprising stuff when you put them to work. none of them are capable of independent thought regardless of what people say so true ai that is capable of independent cognition is now referred to as agi artificial general intelligence there’s no agi right now but there are people working on it the tools that we have now that we call ai can’t lead to agi new tools and methods are going to be required so So you don’t have to worry about chat GBT becoming sentient. Even though AI isn’t capable of real thought, it can be used by humans to do some incredible things. But that being said, AI, although everything gets branded with the letters AI these days, it’s not one monolithic thing. Every use requires specially designed algorithms and training. So while you may think of AI as a group of things like ChatGPT, Sunno, or Apple Intelligence, these are all different products designed for different use cases. So why are we talking about AI? Well, even if you’ve never personally used AI directly, you’ve been affected by it in some way without even realizing it. Companies have been cramming AI into everything these days. It is used in advertising, software creation, music, media generation, both text, pictures, and videos. It’s used in writing, customer service, web design, journalism, dating apps, even job applications. If you’ve never been interviewed by AI, I’ve got to tell you, it is just as horrific as it sounds. It’s used in games and so much more. AI is a powerful tool when used properly, but it can really ruin things when used poorly. You may have heard of the term AI slop. This refers to the nearly ubiquitous, very lazy overuse of AI. So, for example, now we get ads for fake products. We get websites that pretend to contain real valuable content that actually isn’t. It’s lazy and just straight up wrong in most cases. AI bots listening to AI-generated music on streaming services to generate money for scammers. That has been a big thing, especially on the biggest music streaming sites. Apple hasn’t said specifically how much of a problem they have there, but I’m sure they do. Companies using AI also use it to replace people who write software. They replace marketers and customer service reps. If you’ve tried to hit up a company via chat, you probably ended up using some sort of simplified AI chatbot, at least initially. It’s yet another level that you have to go through before you can get to a human, if you can. They also replace journalists with AI, unfortunately, as if we don’t have enough problems. They’re poisoning social media and news sites with breaking news that looks real but isn’t. Chatbots that very confidently hallucinate answers that people never check. AI is used to manipulate you and drive propaganda. All of this while using a lot of electricity and driving RAM prices through the roof. You may have noticed that or may have read an article noticing that things like the Mac Mini and the Mac Studio, a lot of them are just straight up not available now because RAM is not available. So it is affecting absolutely everyone. It’s honestly easy to look at all of this misuse and say AI is bad. But the truth is, used properly, AI can be really helpful. It gives people without skill, that don’t have the skill to make things, the opportunity to create what they wouldn’t otherwise be able to. So, for example, I used AI to create the opening picture here. The actual picture itself is AI. And you’ll notice I put in some pointers here. These are all indications that this is AI. In those spots, there are things that you can look at to detect that this is not a real picture. So I wouldn’t have been able to create this myself. I am not a graphic artist. And honestly, I wouldn’t have had time or the money to be able to create this for tonight. So I’m not taking anybody’s job. But it does allow me to improve things for you guys. But other than that, I haven’t used AI for anything else tonight, other than the examples of actual AI that I will be giving you, which I, for the most part, didn’t create. Okay. So not only does it help me do things like this, but it’s also used for health and medicine. Specially designed AI is being used to detect various types of cancer with much higher rates of accuracy than humans. So, for example, here’s a new AI tool can diagnose cancer, guide treatment and protect patient survival. AI finds signs of pancreatic cancer before tumors develop. And pancreatic cancer is particularly deadly. So early detection is critical. So that’s great news. Pancreatic cancer mRNA vaccine shows lasting results in an early trial. They actually used AI to help design some of the RNA structures that are used in the vaccine to make it more effective. So it’s being used all over the place. Okay. AI has also helped us to determine methods of protein folding. So a lot of diseases and whatnot, they occur because proteins in our body get misfolded in ways that they shouldn’t be. And so scientists have been working for many, many years trying to figure out how proteins can get misfolded so they know what to look for, for diseases, etc. and to understand how the body works. So AI has been really, really good with this. In the past, I think it’s even still a thing. There’s a folding at home project where you can use your computer to help model these misfolded proteins. But AI has been a game changer because you can go through stuff a lot faster than humans can. So that’s pretty good. So on top of all of that, AI tools can be used for accessibility. So for example, Apple’s personal voice feature for those who are losing their voice. The AI-powered voices that are included with your Apple devices allow the blind to better interact with what they’re using. AI is powering object detection to allow those with poor sight to better interact with the world around them. So now you can point your phone at stuff and essentially your phone can tell you what it is that the phone sees. So, you know, it’ll tell you what the sign says. It’ll tell you where it is, how far away it is, that sort of thing. So it’s very useful for accessibility, which truthfully, you know, even if you’re not blind or deaf, et cetera, The accessibility features are very, very useful for just everyday use. If you haven’t already, I do recommend that you go into all your devices and check out the accessibility settings. And science in general. AI helps to speed up scientific efforts using custom-designed machine learning tools. So, you know, here’s an article talking about that very thing. Okay. So why do you need to try to detect the use of AI? Even if you don’t spend much time in social media, you’re still exposed to AI nearly every day. AI is used to write at least some news stories, even at some large news outlets. So, for example, the following examples are outlets that have said they are using AI to one degree or another to actually write news stories. The Washington Post is one of those. They have doubled down on AI, unfortunately, and got rid of a lot of their journalists. The Washington Times is another one. And then Gannett-owned newspapers. So USA Today and tons of local newspapers that Gannett owns are using AI to one degree or another. Sports Illustrated, Men’s Journal, The Street, Parade, etc. All of them are using AI to generate actual content. Newsweek is another example. They’ve been doing it for a while now. and you can tell the difference. There are some news outlets that don’t use AI to write actual stories, but will use it to write headlines and article summaries. So examples in this category are the New York Times, the AP News, how they’re leading with AI, and also Fox News. So honestly, no valid news sources should use AI unchecked. And even when it is checked by a human, the current generation of AI should be used sparingly as it’s way too easy to make mistakes. But it’s kind of funny because based on recent studies, the people who tend to use AI the most, which is Gen Z and Gen Alpha, the more they use it, the more they absolutely despise AI use. So a lot of them are actively avoiding AI of any sort because they realize how generally toxic it can be if you use it too much as a crutch. But unfortunately, we can’t put the genie back in the bottle, so we have to find some sort of middle ground. Unfortunately, AI slop is just about everywhere these days and is often used to try to trick you. Entire websites with extensive fake content can be created with very little effort and are used to scam money out of both advertisers and you. So, for example, here’s a recent article that Axios did where a particular fraud group created more than 200 AI slop websites, basically to scam people and to earn ad revenue for content that is non-existent, basically. AI generated video even fools major news outlets. So if you’re thinking that it can be hard to detect AI, well, you’re not the only one. So there are people whose job it is to make sure that they don’t get fooled that have occasionally gotten caught up in AI content, unfortunately. AI has been used to flood Amazon and other print-on-demand sources and has made it harder for genuine people who are trying to publish content for their stuff to be found. But here we have it on Apple News, because Rolling Stone is behind a paywall and you can’t read anything more than the headline. So what I linked to was the Apple News article. But it really is pretty amazing how much AI slop is out there. Like travel books, etc. There’s a lot of just straight up garbage. And there’s a lot of stuff that is trying to pretend to be relatively famous books by famous authors. That’s just AI slop. They’re trying to make money off people who don’t know the difference. So you do need to be careful there. because it can be used to basically take your money. So how do you detect AI? There are a lot of paid services out there that claim to be able to detect AI content with around 90% accuracy. However, when put to the test, they all seem to fall short, especially with AI-generated text. So the average text correct detection rate for AI text is around 60% for most of these paid services, which honestly isn’t much better than random chance, which would generally come up with the right answer about 50% of the time. Is it AI or not? Yes or no? And so, where are we here? Okay. So if you cannot trust these services to detect AI, how can it be detected? Well, every AI model has different strengths and weaknesses, though they almost always have some tells or some issue that crops up often enough to suggest that what you’re looking at or listening to wasn’t made by a human. So I’ll show you how to look and listen critically to what you’re interacting with to determine if AI might be involved. This will help you get your detection rate up over 60%, hopefully. But keep in mind that everyone can be fooled. Some people, they’ve done testing. Some people are better at just natively detecting AI, and they’re not really sure why that is. But everyone can be trained to detect it more frequently. So in addition, new models are being generated all the time. So what works right now to detect AI may become less effective a year from now. Though many of these tips should still give you a skeptical eye that makes it less likely that you’re going to fall for AI content. AI fake detection differs depending on what the AI model is generating. So we’ll break down our detection efforts into text, pictures, video, voice, and music. Also keep in mind that current scientific testing shows that some people are just better at it than others. So maybe you’re one of those lucky people who just inherently can look at something and go like, yep, that was made by AI and generally be right. General approaches to detecting AI. These things apply to detecting any sort of AI, no matter what you’re looking at. So always approach anything important that you’re looking at, listening to, or reading with a general air of skepticism. Just because you see it somewhere doesn’t necessarily mean that what you’re seeing is true or completely accurate. So generally approaching things with a bit of a skeptical attitude is a good thing. And that doesn’t mean you can never trust anything ever again. It just means that if you start from a place of, you know, I’m going to ask questions and make sure that what I’m seeing is right, is a generally good approach to take. Look for multiple reputable sources to confirm what you’re reading, listening to, or viewing. So, you know, if you run across a website you’ve never heard of before, making some incredible claims about something that’s happening in the world. And what I would do is go check, you know, sources, media content, you know, places that you trust that have track records of posting, you know, quality information. Do you see that same information getting reported at those places? If not, then you should assume that what you’re seeing may not be real. Until reputable sources start reporting on it, you might want to hold off. You can also check reputable fact-checking sites. There’s lots of them out there. PolitiFact, Snopes, et cetera, are probably two of the more famous ones. But there’s lots of them out there. And a lot of media sites now have their own fact-checking departments that will report on issues like this. So, you know, take a look before you share that post on social media. Also, watch for leading text or media. If someone tells you that you have a very limited time to get something, always be skeptical. Especially if there’s a countdown clock on the website. I guarantee you if you see a countdown clock, yeah, this is somebody who is trying to get you to not think about your purchase, your potential purchase, and getting you to just hurry up and buy before you have a chance to think about things. If someone tells you that the government or some nebulous group or they, with a capital T, have been suppressing a miracle product or a cure or some critical news and no one else is reporting on it, you should not only be skeptical, you should probably leave that site immediately. Because I almost guarantee you that what they’re about to tell you is just garbage. So while distrust of big pharma or big tech is fairly pervasive right now, one thing we can all agree on is that if there is money to be made, the big players will get themselves involved somehow. They would not hide a cure or an effective product. They’d monetize it because that’s what capitalism is. So, yeah, they don’t try to hide stuff that could make money. We are unfortunately very much back in the age of snake oil salesmen. I remember when I was a kid, I always, you know, I remember reading about like, you know, this, you know, 18 and 19, 1800s with snake oil salesmen selling all kinds of elixirs and random bottles full of, you know, god awful, terrible, basically poison. And I was always thankful that like we don’t live in an era like that anymore. Unfortunately, that’s not the case now. There are a lot of people out there selling the equivalent of snake oil. And yes, some of it is very dangerous. In fact, you know, there’s a lot of stuff that you can just go to a local grocery store or a gas station and pick up stuff that could potentially harm you or even kill you. So it’s unfortunate, but lots of products being sold as providing a cure for a common problem, they haven’t been independently tested or verified. So you have no way of knowing if the stuff that they claim is in that bottle is actually there. For example, supplements are big business right now. So if you see the warning, these statements have not been evaluated by the FDA. This product is not meant to diagnose, treat, or cure, or prevent any disease. You should immediately be skeptical of the product. That doesn’t mean that it can’t be useful, but that it has not been independently or scientifically tested to prove the words that are on the bottle or on the package. Groups like Consumer Reports and Consumer Labs often do unprompted testing of supplements and very often, especially supplements from no-name brands, do not contain the amount of ingredients that are listed in the package. Sometimes they have way more than is listed and sometimes they have nothing at all or they have other stuff like heavy metals in concerning quantities, etc. If you have to take supplements like vitamins, make sure that you obtain them from a reputable source that is willing to give you a recent third-party testing report on the product you’re going to take. If it’s a well-run company that’s not trying to take your money, they do testing, and they will provide, if you contact them, they should be able to provide you, here’s our most recent testing on the products that you’re looking to buy from us. Recently, I started taking a fiber supplement, and the first thing I did was contact the company for the third-party testing. And yes, in fact, they did third-party testing. They tested for heavy metals, et cetera. It’s good to go. And, of course, if you’re going to take supplements, only take as much as you need. Avoid products with lots of different vitamins in one pill. or even worse, anything that says it’s over the counter, that says they have a proprietary blend of anything at all. Because those proprietary blends, while the individual items in it may not be particularly concerning, although a lot of that stuff can interact with medications you may be taking, but, you know, taken together at the levels that may be in the product could potentially be concerning. Also, when you see ads claiming to have a doctor who supports the product, your default should be to assume that the person you’re seeing is just a paid actor, if they exist at all, because this is a very common scam. Oh, hi, I’m Dr. So-and-so, and there’s no such doctor. But, you know, it gives the product an air of, you know, competence. So those are the general things that you should do. Now let’s talk specifically about AI text. From helping you write or rewrite an email, to help a student cheat while writing an essay, to creating ad copy or even an entire website, AI text generation is very common. Here are some writing samples. They’re included in the PDF that I linked to earlier. so you can see them in their glory there. I’m not going to read them in their entirety. I’m just going to read a little bit of it here to give you a sense of each one of these.
Jim • 27:23
By the way, Aric, you know you’re still on the Business Insider page on what you’re displaying to us?
Aric • 27:30
Yes, yes. I’m not going to actually display the text. I’m just going to read it.
Jim • 27:35
We had somebody giving a presentation that thought that they had moved ahead five or six pages and the display was not moving.
Aric • 27:42
Yeah. In fact, I can probably stop the share for now. But then I need to go here. Before we talk about ways to detect AI, I’m just going to read a little bit of samples. See what you think. See if you think that the following samples are AI or not. Some are, some may not be. The first writing sample is called the keeper of the second sky. Once upon a time, in a village nestled between two hills, each rising no higher than 300 feet, there lived a girl named Ellera, who collected broken things. One autumn morning, around nine o’clock, she found a cracked clay bird in the tall grass. When she pressed her thumb to its beak, it sang, not with sound, but with color. a single thread of blue light unwound from its chest and drifted upward, tangling the branches of an oak tree. The villagers all had long forgotten that the sky above them was only the second sky. The first had been woven ages ago by the warden of small truths, who stitched stars from forgotten lullabies and mourning apologies. But the loom had broken, the first sky had frayed, and the dark slipped through. People built roofs and stopped looking up. Elra followed the blue thread deep into the woods. It ended at a cave where an old woman sat sorting piles of snapped thread. The loom needs mending, said the woman, but I cannot see the colors anymore. Elra offered her the bird. The woman smiled. Not the bird child, the listening. And we’ll stop there. Now here’s writing sample number two. This is called How the Whale Got His Throat. In the sea, once upon a time, oh, my best beloved, there was a whale, and he ate fishes. He ate the starfish, the garfish, and the crab, and the dab, and the place, and the dace, and the skate, and his mate, the mackerel, and the pickerel, and the really, truly, twirly, whirly eel. All the fishes he could find in the sea, he ate with his mouth, just so. Until at last there was only one small fish left in the sea, and he was a small stute fish. And he swam a little behind the whale’s right ear so as to be out of harm’s way. And then the whale stood up on his tail and said, I’m hungry. And the small stute fish said in a small stute voice, noble and generous citation, have you ever tasted man? And that’s where we’re going to stop there. If you read the writing sample, there’s more of it. in the PDF. Writing sample number three. This doesn’t have a title. It just says, Consider the common harbormaster who, whilst ye sleeping landsmen do smugly a bed lie and dream of your green meadows and stillborn ambitions, stands sentinel upon the weather-beaten pier as the first gray fingers of dawn do but feebly grope that endless and indifferent waste of waters. He is a man, yet he is more than a man. He is the very hinge upon which the vast door of commerce and catastrophe doth swing. Having looked upon the faces of 10,000 mariners, each bearing in his watery eye the same unspoken interrogatory, what doom, what distant and unattainable mercy lieth wrapped in the heart of that same deep which rolleth now so docilely at his boots? And we’re going to stop there. Okay, so now we’re going to talk about what you should look for in AI text. First, remember that different AI models will return different results. So the items I discussed below are what are most frequently seen with the current models, but there may be differences depending on what models get used. There are a lot of them out there. The first and best way to detect AI is to notice a difference in the writing style for the person who supposedly wrote the content. Of course, this requires you to know what the person’s writing style is like. So this isn’t going to work if you don’t know the person who wrote the content or if you don’t know them well enough to know their writing style. However, if your friend usually writes simply using straightforward language and you suddenly get something from them that uses very flowery language and punctuation they don’t normally use, well, that’s a dead giveaway. Most writing you see won’t be from somebody that you know personally well enough to judge their writing style. But thankfully, there are other tells. Look for regular use of uncommon punctuation, like n or m dashes, or liberal use of semicolons. Most people don’t use these in their typical writing, mostly because people don’t generally know for certain how they should be using them. So, for example, hyphens are regularly sized, and you see these most often in text. They’re used to link words together and compound words like the pre-war period or woman-owned business or to hyphenate words at the end of lines. N dashes are slightly bigger than a hyphen, about the width of a capital letter N, and whatever font is being used. That’s why they’re called N dashes. They are most often used between dates or ranges of numbers to indicate inclusivity, such as July 15th to July 20th, where the two would be an end dash. M dashes are the biggest of all. They’re the width of a capital letter, M, in the current font, and they’re used in places of parenthesis to denote or to break a sudden shift in the direction of a sentence. So let’s see if we can go here. Yep. Okay. I’m going to switch back now to share my screen to share sound. That one. Okay. Now we’re back on the view of how to use em dashes. Semi colons are sort of like super commas. They’re most frequently used to separate items in a list, especially in lists where commas are used as part of the list item. You can learn more about that at Merriam-Webster as well, right there. In longer passages, check that what you are reading actually makes sense and doesn’t wander from paragraph to paragraph. AI can sometimes get off-topic or assert something that is absolutely ridiculous. But you may not notice that because at first it sounds reasonable, and then it just gets weirder as time goes on. So it may take some time for you to read a source before you know whether it’s AI or not. If the texts happen to have footnotes or references, check them to make sure they’re real. Because a lot of times when AI does that sort of thing, they will link to non-existent sources. And that’s actually gotten a number of lawyers in trouble because they’ve tried to use AI to write briefs for the court. and then they reference court cases that never happened. But it sounds good if you don’t actually check the sources. If you can’t tell for sure and nothing else works, and you end up needing to use an AI text detection tool, and there are literally hundreds of them out there these days, make sure that you use several of them on the same passage, and that they all agree that what you’re looking at is is or is not AI. Because that’s the only way to be hopefully more certain that what you’re saying is or is not AI, if you’re going to use a tool like that, especially since they’re not super reliable. Okay, so…
Frank • 35:48
Aric, when you have them write something and you give it certain parameters, If you do the same thing several times, do you get the same results? Or is it different results every time you put it in?
Aric • 35:59
It’ll be a different result every time you put it in, except for like, so if you want the AI that you’ve been talking to, to like continue writing in the same vein of what has been written, or if you want it to rewrite what it already wrote, then you have to tell it specifically, hey, that thing you just wrote, please rewrite it to make it sound nicer or use the word pink or whatever. And it will take the existing text and rewrite it. Otherwise, it’s just going to go off in a different direction.
Frank • 36:37
But what I’m saying, if you put exactly the same request in, you turn the machine off or you do it again, do you get the same results again or do you get it different?
Aric • 36:46
Nope. It’s different pretty much every time. You may notice certain commonalities depending on what you’re asking, but for the most part, you’re going to get something different every time. Okay, so take a look at the samples that I read earlier. Let’s figure out which ones are AI. Well, sample number one. Right. The sample number one was AI. and this is a fake fable that AI generated on the spot. It took about 30 seconds to come up with the entire story. If you read the entire thing, it’s a little bit all over the place by the end, and if you look at the actual text, you’ll notice that there are literal uses of em dashes, and as I said, the fable doesn’t really go anywhere if you read the entire passage. Sample number two, Oh Best Beloved. Was this AI or not? It was not AI. It was actually a passage from Rudyard Kipling’s Just So Stories. If you’ve ever read Just So Stories, you may have recognized the phrase, Oh Best Beloved, because he uses that particular phrase quite a bit.
Jim • 38:02
I was confused because, oh, my best beloved is in the middle of the first line. took me a while to realize, to get congruous with you. But it did sound to me like what I might hear in a children’s tale.
Aric • 38:16
Yes, and in fact, that’s basically what he was writing, is sort of fake parables to explain how things became. So like how the camel got its hump, or how, you know, this thing evolved, etc. So they’re fanciful stories that aren’t meant to be taken seriously, but they’re just sort of fun ways of sort of explaining how the world became the way it is. It was written a long, long time ago. And sample number three was AI. It was written vaguely like Herman Melville, if he was an angsty 16 year old boy. They got, again, liberal uses of em dashes, semicolons. Oh, my. And frankly, it’s, you know, it reads a little bit like somebody who is like trying to pretend they’re a pirate, particularly as you, if you read further into the sample that I provided.
Jim • 39:11
Okay. So what about people like me that like to use semicolon, that have enjoyed using semicolon and the various styles of hyphens and dashes? I’ve realized that, and I’ve heard those kinds of stories in the last month or two, is that they’re basically saying, don’t trust Jim Thursby because obviously he’s an AI because he things and he wants to end the fact that i typically use them correctly is of no accord because most people don’t know how to use them correctly and can’t recognize it so i’m like getting mad that ai is basically saying don’t trust we’re ruining perfectly good people yeah
Aric • 39:53
yeah uh it’s funny because honestly um the the current generations uh punctuation tends to scare generations. So it’s used less and less, because even things like a period can be considered quite hostile.
Jim • 40:12
Yeah, I’ve seen things where I don’t post to social media, but they’re basically saying if anybody posts to social media and they use sentences, they’re not trustworthy people.
Aric • 40:25
Right. You know, and as a, you know, I used to be an English teacher for many, many years. I taught English second language to adults. So, you know, I, too, also love using Em and En dashes as appropriate. But I can understand where, you know, it’s not surprising that people haven’t figured out how or even what an en or em dash is. Because it’s not part of your, if you look down at your keyboard, you’re not going to see that punctuation there. You have to know how to generate it and then why to generate it.
Jim • 41:04
I actually use the original that I was taught where we didn’t have word processors. We had typewriters and we learned the dash dash. There was the one hyphen and double hyphen. Right, yeah, because you couldn’t generate. It represented the am, la, and dash.
Aric • 41:20
Right, exactly.
Jim • 41:20
And I found it hard to read, so I typically add a space double hyphen and then space. Right. So I don’t use standard typography in terms of spacing between the things. I try to be readable, just like I try to follow the old standard of putting two spaces past the period on sentences to help separate them. That wasn’t pica type. What was the other type of typewriter that was evenly spaced?
Aric • 41:54
There was the Selectric, the IBM Selectric, the one with the ball.
Jim • 41:58
Before you had Electric?
Aric • 42:01
Before Electric. Yes.
Jim • 42:04
There was like PICA and some other type of type. And PICA, I think, was the one that was somewhat proportionally spaced. And the other was equally spaced.
Aric • 42:14
Yeah. And most typewriters ended up on, you know, equal spacing just because it was a simpler thing to accomplish.
Jim • 42:23
Mechanically.
Aric • 42:24
Yeah, right.
Michael • 42:26
I’d also like to point out that back in the 1800s, writers used end dashes and end dashes a whole lot more than they do today. So if you read the works of, say, like Mark Twain or Emily Dickinson, if you’re a fan of either of those, you’ll use end dashes and end dashes.
Aric • 42:47
Yeah. And the reason for that is because a lot of that stuff was written by hand. So it’s much easier to generate, you know, the appropriate size when you’re doing it by hand. But so few people end up writing by hand now. You know, that’s why it’s sort of falling out of favor. So I can’t say that it isn’t sad. And the truth is, eventually, AI models will get better. Because honestly, not only do they use en dashes and em dashes a lot, they often use them wrong. So, you know, you need to be careful with that. But if you see a lot, particularly from somebody that you know, then that’s a good sign that, you know, somebody helped them ChatGPT or otherwise. Okay, so let’s move on to detecting AI pictures. Okay, let me go over here. Close that off. And then I’m going to show you some pictures. There’s one. Just decide whether these are real or AI. There’s another. There’s a lovely flamingo. And the cutest baby bear with a bear costume.
Frank • 43:55
Lion.
Aric • 43:56
Yeah, lion, sorry. With the… The reason why I said baby bear is because the file that I downloaded says baby bear. Okay. So, yes, it’s an old-timey picture.
Jim • 44:12
Can I interject for a moment? Go ahead. What’s the importance of being able to detect a fake as AI or a fake as traditional methods?
Aric • 44:22
The truth is, it doesn’t matter. A lot of these tools were… The thing is, right, like, Detecting fakes no matter where they exist are important, but when somebody puts real effort into a fake, it’s much harder to detect. Think counterfeiting money or artwork, etc. These are not instantly generated on the fly. Somebody spent real time to do that. So they can be harder to detect. But AI makes it very, very easy to generate hundreds and thousands of fakes within just a few moments. So they can just spray all kinds of fake stuff with very little effort. And that’s why it’s in this day and age important to recognize when AI is making the thing that you’re looking at. so that you know, you have a sense of whether you can trust it or not. Okay, there’s a picture of a car. Pretty cool looking old car.
Frank • 45:28
How can you tell if it’s fake or just Photoshopped?
Aric • 45:31
Well, I will tell you that in a few minutes. Oh. So here is a lovely looking dessert. By the way,
Jim • 45:40
Frank, when I ask about traditional fakes or AI, I was including Photoshop in traditional fakes. Yeah.
Aric • 45:49
Yeah, yep. And it’s just, again, like I said, back in the day, to make a fake, a convincing fake, required real skill. There is a YouTuber called Captain Disillusion that basically deals with a lot of early fakes before the time of AI. and the amount of effort that people needed to put in, even with tools like Photoshop, et cetera, a lot of work to do it properly. And now AI makes it just so much faster to be able to do these things. So there’s another set of flamingos. Here’s a man hanging out. There’s another guy. There’s a woman. Oh, and here we are in OpenAI’s headquarters. As it says in the file name, somebody’s writing up some details about working with various modalities and transformer models. Here’s a lovely looking open floor plan of a house. And then here’s an athletic woman wearing both ballet slippers and a bathing suit. Now, this was obviously created, but was it created by a human or was it created by AI? There’s another woman. Okay. So keep in mind what you think are AI and which are not. Okay. So how do we detect pictures that may have been generated by AI? Well, one very obvious thing is that because AI gets trained on a lot of the same material on the Internet and, you know, Pixar, etc., has been very popular. So there’s a lot of content out there from Pixar. It’s funny how much Pixar content actually got absorbed into the various picture creation models that are used to create pictures. So if you see people that have glowing skin or where it looks slightly plastic, that’s usually a good sign that AI may have been involved. There may be obvious or even subtle mistakes in the picture. Small details matter when you’re trying to figure out whether it’s fake or not. So look in the background. See what’s going on. Because just because something is blurred doesn’t mean you can’t find interesting or important information that will help you determine whether the thing is real or not. While most newer models don’t have issues with hands or feet anymore, They used to not know how to generate the correct number of toes or fingers, or they would generate too many hands. That was very common in early models. And so a very, very easy tell. There are still some models that struggle with that, but they’re becoming less and less these days. But if you see somebody holding their hands at an odd angle or, you know, an extra hand sprouting out of a spot, then that’s a good sign that it was AI. Don’t forget to check, you know, not only the background, but things like the hair in a picture. Because small stuff like little wisps of hair, they may actually break in odd ways or, you know, grow at weird angles out of the skin. So it’s always worth looking at that too. AI picture generating models are getting better with text. So, you know, creating, you know, artwork on your shirt, et cetera. But they do still make mistakes. So the text looks odd or nonsensical. It’s possible it could be AI. So when you’re checking text, not only should you look for like, is this a real word? Because again, it could be a foreign language. There are other ways that the text can look odd. If it looks smeared or broken in odd ways, or if it just clearly isn’t a proper word in any language, it’s almost certainly AI. Another thing to look for is super tight bokeh, which is the name for that blurring effect that you get on higher end real photography and also faked in things like the iPhone in the cinematic mode, etc. It generates a sort of fake blur in the background instead of everything being in focus. So AI tends to make, when it does bokeh, it tends to make it a little over tight. So things that are at the same level, like in the same, like right next to the main subject, may be blurred where like clearly it shouldn’t be if this was a real picture. So always worth looking at what’s being blurred in the picture. Another thing to check for is the metadata in pictures. and go over here here’s a tool that is free and is available in the Mac App Store called exit Fox you basically just drag and drop a photo into exit Fox and it will show you the metadata that isn’t contained within that particular file so here I’ve brought up two of the of the images that I showed earlier one is the picture of a room and this is in Exifox and this is all the metadata that is available and if we scroll down a bit we go through all of this metadata about the picture and how it was taken and ways that it’s been you know the color has been added to the file etc if we keep going we see right here it’s a source mid journey. Mid journey is a picture generation tool. And so this picture was generated by mid journey. So even if you look at the picture and like, I don’t know, sometimes the metadata can give the picture away. If you think to look there again, you’re not going to do that every picture, but if it’s important to you, then you, it might be well worth looking at.
Jim • 52:26
So you can get the metadata from a picture that’s on a web page from the Washington Post, for example?
Aric • 52:33
Yeah, yeah. You can just right-click on almost any picture in any website and just save it to your computer. And then just drag and drop that file into Exit Fox here. And then you can see the metadata for the picture. Now, here is an obvious one. This was manually generated by a local AI. And one of the ways you can tell is because the prompt used to make it is actually embedded in the metadata. So it actually says ultra realistic, high quality man’s face smiling at the camera. So this was a model that’s even it actually lists the name of the model that was used and the other settings so that if you wanted to, you can recreate this exact picture using this data. So it is possible if you know all of the data that went into creating the picture and you use the exact same model and the exact same settings with the exact same seed, you’ll get the same result. But that only works for pictures. Okay. And obviously some generation tools will add a watermark to the picture. Sometimes the watermark is visible in one of the corners, which obviously, like if somebody’s trying to scam you, it’s not going to help because they’ll probably just either blur it or crop it out, etc., or otherwise try to hide it. But for a lot of tools, what they’ve taken to doing is there’s an invisible watermark that you can’t see that’s embedded in the file. And some AI detection tools can detect the watermark. And if you have a sense, if you likely know what service might have generated the content that you’re looking at, you can very often go back to that service, whether it’s like Google’s DeepMind, et cetera, and you can provide the picture and say, hey, did you generate this? and it will be able to tell you, even if the item has been cropped, if it was at least involved. Maybe it didn’t do the entire picture, but it did some part of the picture.
Frank • 54:53
If you just copy a picture and drop it into ChatGPT or something like that and ask it to verify, will it do it?
Aric • 55:01
Well, it depends on the model that you’re using. It does make an attempt to detect, but it’s not, it will, as is so often the case with chat bots, it can very confidently tell you the wrong answer. So it’ll tell you, yes, it’s definitely a fake. So maybe it’s true, maybe it isn’t. Generally speaking, the best way to detect is there are dedicated tools and services that can look for those embedded watermarks that are not using AI to do it. And they tend to be a little bit more reliable if you’re concerned. But keep in mind that not every picture that’s generated by AI is going to have a watermark, visible or otherwise. But it’s certainly another area that you can go look for stuff. So let’s take a look at the pictures I showed you earlier and I will explain how those things are or are not AI. Well, this one, it’s obviously AI. It’s not real. Those don’t even look like convincing aliens. They look like plastic mannequins made in the shape of aliens. But, you know, obviously somebody made this as a joke. You know what I’m saying? Like he’s on a, they’re on some sort of talk show. But if you look in the background, a lot of these people don’t look like they actually belong in the group. And some of them have their collars are messed up or their hair is out of place. So there are a number of ways to look at this and see that it’s very clearly not real. Also, let’s face it, there’s no… When’s the last time you saw somebody with some sort of talk show wearing, you know, basically a tuxedo while talking to people? So, definitely AI. Here’s another. Now, is this AI? Yes, this is AI. Either one of these planes are real, and they are not actually flying in the sky near one another.
Frank • 57:20
Not like that. No way.
Aric • 57:22
Yeah. And one of the ways you can tell is because they were kind of sloppy. If you look at the American Airlines plane here, it’s got like a gray haze around pretty much the entirety of the plane. And that is not normal. And furthermore, I went looking for like a Delta plane that has this exact kind of marking on it. And I couldn’t find anything that Delta ever had a paint job that looked like this. And not only that, but, you know, if you look closely at parts of this, it looks like like back here in the American plane, like right around here. It looks like the planes actually dented, like something like messed up the back. and they just were like, oh, we’ll fly anyway, which is, you know, if they’re going to do that kind of damage to an airplane, they’re not going to be flying it anytime soon. So, I mean, you might occasionally see speed tape on the plane, but that’s just a cosmetic thing. It’s not, they don’t, they don’t fly when structural things are wrong. So, okay. So this flamingo, is it real or it fake? It’s real. This is a real flamingo. I got it directly from a nature website. It’s a Creative Commons picture. That is a real flamingo. This one that I got that says baby bear, which is a baby lion, is that a real old-timey picture? Notice the sort of burn-in down here, which we’re all old enough to remember times when cameras did this automatically sort of burned in a little area telling you when the picture was taken but you’ll notice there’s the date is not quite right so it is it looks closer to 1929
Jim • 59:26
than to 1989.
Aric • 59:27
Yeah, even though 1989 would be a better example. In addition, so like there are some of these wisps do some weird things in places. This telephone, if you look closely at the dial, it’s not quite right. It’s close. And, you know, there is the sort of area that where you would, you know, that would stop your finger when you’re dialing that, but it’s not in the right place for a typical rotary dial phone. In addition, some of the items in the background here, some of the flowers, they just stop or get cut off in places. So there’s a number of different tells here that this is in fact AI. Now this car, if you look at it quickly, you might think, Like if you don’t know cars particularly, you might look at this and go like, oh, yeah, this looks pretty real. It is not. This is AI. And, well, that license plate is definitely not a thing. No, but maybe, hey, maybe somebody, you know, made a change because they don’t want to show somebody’s real license plate example. But over here, the window is kind of weird right in this section over here. And some of the chrome here doesn’t flow properly. And there’s these little lights here. If you look closely at them, they’re smudged and they look weird. Like somebody took a Gaussian blur and accidentally painted over the spot. So this is, in fact, a fake. Now, also, the sort of fourth wheel back here is hard to find. So, you know, another sort of situation where, yeah. And this is no car like I’ve ever seen. Now, I am no car expert. Maybe you are, and maybe there are additional tells that this is not a real vehicle. But even if you don’t know the car, there are still plenty of obvious ways to detect that it’s AI. Now, this dessert looks kind of real. It is not. This is purely AI generated by a model. And one of the things that you can tell is the bokeh. So you’ve got the blurring back here behind the plate, which, again, is what you would expect to see for your average bokeh. And you would want all of this dessert more or less to be in focus. But there are places like right around here and right around here that there’s little spots of bokeh and spots of clarity that give it away. Plus this frontmost blueberry is a little bit oddly shaped. Now, it could just be like, hey, you know, it got damaged or something when they were making it. But yeah, the bokeh is very clear because there are things that are blurred here that shouldn’t be in a normal picture. So that’s the tell for that one. Now here’s more flamingos. Is this real or fake? This is fake. And how do you know? Well, for one, there’s a little white dot here that is not like a blown out section of the picture. It is literally, there’s just missing data. It’s just plain, pure white that’s in that one spot right there. If you look closer at the lead flamingo’s eyes, that eye is definitely messed up. It’s not real. It’s sort of like somebody took like an eye from a flamingo and then just did another one of those like Gaussian blurs and sort of did a swirly pattern to make it more look like an eye. So, so yeah, so even ones that look fairly convincing still have tells. Now this one is also a fake. This one’s a little harder to tell. And down here, I specifically hid this spot myself because it gives away the results. So let me show you the original. Here’s the original from 2024. It was specifically a design winner where the goal was to take an AI model and make as realistic looking a picture as possible. And they did a really pretty good job. There are some small things, like if you look back at the way that the light bends, some of the light coming from the sides here, there are a couple of places that look a little unusual. But other than that, it would be pretty hard to tell. So if this wasn’t here, you could easily get fooled by this particular picture, Because it’s hard to see that there’s any difference.
Jim • 01:04:58
Yeah. I was attracted to the first idea on it, which is, well, if it were real, it would be a composited picture where the person wasn’t in the background and the person were combined together.
Aric • 01:05:10
Right, right.
Jim • 01:05:11
The artifacts of the person, I don’t think you could get lighting and other things to create.
Aric • 01:05:17
Yeah, yeah. And sometimes between them. Right. And sometimes, you know, you get that. You’ll see this even in high-end movies, because anytime that they’re doing CGI in the background, that CGI doesn’t reflect off of things in the foreground like you would expect. And it’s hard to fix, even in big-budget movies. So that can be a way to tell. But in this one, you know, you could make a convincing argument that maybe this isn’t AI, that like somebody had just used Photoshop to edit the picture a little bit, but it is in fact AI. Now, this one we already know is AI because I showed you the metadata, but honestly, look at his skin. Look how plastic it looks, how shiny it is. And that is definitely a tell. Now, this one is also AI. one, take a look at that background. Where is he? It looks kind of like he’s in like NORAD, like a bunker of some sort with all kinds of like technical data. But yet it looks like he’s sitting at a coffee shop, just enjoying a coffee, looking away from the end of the world. So he is in an odd spot. His skin is very plastic and shiny looking. He’s got the right number of fingers. I mean, you can’t see the thumb, but, you know, at least the fingers are there. But yeah, you wouldn’t expect to see a person sitting like this with that particular background. and add that to his face and you’ve got a good tell now this this is also AI this is a little harder to tell because she’s wearing such a plain outfit but basically her skin look at all of her skin it’s all very shiny and all equally shiny and slightly plastic looking not too bad you know there is some diffusion etc but it’s a little too perfect even for a model you know where you have somebody doing your makeup etc they’re not going to cover your entire body and makeup
Jim • 01:07:44
now a few a decade or so ago they had a few things that they went into how photoshop was being used and especially the idea of presenting a perfect woman to people as real and they went into like 14 levels of adjustments on a photo that was being used for an ad beyond just the makeup and the other thing. So that kind of presents two things is where I ask on the words, the same thing on the pictures. If somebody is trying to create a picture rather than take a photograph, does it really matter if it’s AI or not?
Aric • 01:08:22
It doesn’t matter. But like I said, the reason why we’re concerned about AI now is because it’s so easy to use AI to create scam pictures where somebody needs real talent to do it with Photoshop. And so most people aren’t going to go through that kind of effort.
Jim • 01:08:40
That was the other thing I’d seen where I think I wandered onto a posting on social media where somebody was claiming that they didn’t use AI to create a picture. And they were just that good of an artist and people were like, no, no, no, no, no.
Aric • 01:08:58
Yeah.
Jim • 01:08:59
Yep. But in terms of advertising, typically they want to get something that gets your attention. And so fake might be even better for them than reality in most cases.
Aric • 01:09:11
Yeah. Cause it’ll stand out visually.
Jim • 01:09:13
Yeah. And I think they’ve said for years that there aren’t any ads on TV that aren’t Photoshopped or created, modified or whatever. So I guess it’s like, don’t get too hung up in advertising photos because, you know, unless there’s a purpose associated with them, and certainly don’t, essentially don’t believe anything you see in the advertising. It’s very detailed, correct?
Aric • 01:09:39
Right, right, right, right. Now here’s one that is kind of interesting. So I specifically got this on a site where they said they took a picture in OpenAI’s headquarters in California of the, you know, this woman working at the whiteboard, you know, on some stuff for OpenAI. And now you’ll notice the reflection off the whiteboard is a bridge. Well, even if you’re not sure whether this is AI, you can certainly use the picture there to figure out if they are in fact in OpenAI’s headquarters. And so I went and looked at the two different places where OpenAI has their headquarters. And this is where supposedly at the time that this picture was taken, this is where they were. Now, if you turn around from this building, there are more buildings that are even taller than this one. So there are no windows on this building facing out towards the ocean that would be able to see any kind of ridge. So I don’t care which window this is from. We’re sort of looking at it from across the street where there’s much higher buildings that block the view of the ocean. Now, there is an ocean out there. There is a bridge that’s not terribly far away, but it’s far enough away that it would not look like that picture. And so sure enough, this was generated by AI, one of their internal models. And so it looks pretty convincing, that reflection gives it away.
Michael • 01:11:37
I have to admit that I recognized this as being AI when I first saw it, because I remember Ars Technica had a story about a new model of ChatGPT, maybe it was ChatGPT 3, that had advanced graphic capabilities. And this was one of the examples that they used.
Aric • 01:12:01
Yep. Yep. Because, you know, again, a lot of it looks very realistic, but yep, it’s not perfect. So, and this one we already know is AI because I showed you the metadata. This is AI because, okay, maybe she’s a ballerina who also enjoys swimming, but putting those two things together. Also, if you look closely, there’s a brand name on the swimsuit, which is not a thing. So yes, this is very much AI. Now, this is also AI. The background looks fairly realistic, but she looks like Pixar may have generated her. But if you’re, you know, not sure whether AI was involved, take a look at the background. Whoops, I clicked the right button right over here. We’ve got some text that is kind of legible and is not any language that I have been able to find. It looks vaguely like it might be European in nature, but this is not a thing that I have been able to discover. So if you’re curious, that’s how I figured it out. Now, here’s the woman in the church. And she looks fairly realistic until you look at this braid right here, where you have got an extra braid just sort of jumps right out of the other braid. And there is not enough hair up here to justify there being another braid in that spot. So that is a dead giveaway that this is not real. Okay, this one’s a little harder, but this one is also AI. Her face isn’t quite as plastic, but it is fairly shiny, as is her skin. It looks better than some of the other samples we’ve seen. But if you look at this little bracelet that she’s wearing, The thing, which I assume is supposed to be maybe a clasp, is very clearly just like a morphed version of these little, you know, pearl, rounded pearl things that are on the. Yeah. So it’s just sort of like clumped together. So it’s just like an error that the thing generated when it was making this. So but otherwise it might might be kind of hard to detect. That one. Is that Pixar or is it AI? Well, it’s AI, but it looks like something Pixar might do. Super shiny, very much looks like Pixar, but there are no Pixar movies with characters like this. Okay, is this woman real or not? She’s real. That is a real person. That is a real photo. There’s no AI that has been used in it whatsoever. This is an AI photo. The reflections are not perfect off of her skin and off of her clothes. Plus, if you look at her jacket, like around here, that what we assume might be some sort of hood or collar of some kind, just looks very odd. It doesn’t quite look real.
Jim • 01:15:25
My reaction to this was, why would you create this with anything other than AI? It doesn’t make sense. that it would exist other than as an AI.
Aric • 01:15:35
Yeah, I mean, I could see it being used for maybe some sort of ad where you’re trying to get somebody to focus on the model’s face or something. But yeah.
Jim • 01:15:46
It’s highlighted correctly for anything like that.
Aric • 01:15:49
Yeah.
Jim • 01:15:49
So that was my reaction to something like this. If it’s designed to look like AI, it’s almost certainly, why not use AI to create it?
Aric • 01:15:58
Yeah. And this picture, speaking of women, This is also AI. All of her skin is very sort of glowy. It’s consistent across her basically entire body. And there are wisps of hair in spots that don’t quite behave normally. So if you look around, you can see evidence of that. Now, this one is more obvious because this one has some text on it that is not a thing. This is not any language. This is not any brand. It’s just random letters placed on the front of this piece of clothing.
Jim • 01:16:42
So when I saw this, it looked like it was the same script that was used on your opening picture.
Aric • 01:16:49
Yeah, it’s similar. Not the same model, however.
Jim • 01:16:52
I actually got this from somebody. But I’m wondering, is that something where the AIs, it might have been made from the same AI and it came from the same place? of random text.
Aric • 01:17:03
Well, I mean, yeah. I mean, they might have used a lot of the similar training data. So that might explain the reason. Because it’s not uncommon when they use similar training data that you get similar results.
Jim • 01:17:15
Okay.
Aric • 01:17:16
Now this one, if you checked everywhere but here, might be hard to tell. But this is in fact very much AI because her wrist is at a very unnatural angle here. So this is not real. And there we go. That’s the pictures. Okay. All right. Let’s go on to AI video. So I’m going to show you just a quick little snippet of some video here. Is that AI or not? Okay. Here’s another one. This is fun. This is actually from TikTok. Ah, fascinating. And then we have this. looks like somebody caught some very unusual situations and by with their security camera there okay and then there is one more let me go here okay already okay let’s move on um so how do detect AI video? Well, when you see a video of seemingly impossible stuff, like a bear solo riding on a snowmobile like a human, or a human flying around on their own power without a glider or an airplane, etc., especially if it’s on social media and it’s not in a movie, you’re almost certainly looking at AI footage, even if the found footage looks older. The most common AI video that attempts to fool people is footage that looks like it was taken by a home security camera because we’re used to seeing that footage as being a bit grainy or slightly blurry. So this lets scammers hide the typical AI video imperfections because even if you see something like that, you might think, oh, well, it’s just a compression artifact. So, you know, that’s to be expected. Generating AI video is very expensive. So generally, AI video slop is going to be short, a few minutes or less, typically, often just a few seconds. It’s certainly not feature length, though AI slop may be inserted into longer real footage to try to trick people into a false sense of reality. As with other AI content, always be on the lookout for something odd in the background, a sidewalk that ends abruptly, a door that doesn’t seem to go anywhere, a bunch of trees that are too close together, shadows that don’t appear where they should or they don’t behave the way they should, etc. Since video may also have AI voice or music, you can use the tips that we’re about to discuss for those things here as well. So let’s take another look at those videos that we saw. Okay. So this one with the funny-looking bunny. Is this real or is this AI? Like, obviously, it’s not a real bunny, but did a human make it or did AI make it? And the answer for that one is AI did not make this. It’s Big Buck Bunny, which is a Creative Commons-released video file that has been around since 2008. It has been updated and made into 4K and improved in various ways over the years. But it is made by hand by real people who use Blender to make all of the 3D assets, etc. So it is, in fact, a real human-designed item. Now, this is the horse thing. While it looks interesting, it’s a complete fake. And it is designed by a company that will sell you a horse that they claim is a recliner. But what it is, is a tiny little, almost like rocking horse, with a sort of plush top on it that maybe a small child might be able to recline in. But it does not move. It does not walk. And it does not look like anything you see in the ad. So it is, in fact, a scam. And you can actually see this because they sell something that claims to be this on Amazon as well, like third-party sellers, not Amazon themselves. but yeah so it’s getting harder and harder you know because like certainly you could see where people might be interested in a product like this just because it’s so interesting and different I mean it’s not very you know useful but but yeah it looks cool so and then we have this one right here. And this is also fake. Now, the raccoon riding on a bear, that seems like it could be fake, but believe it or not, raccoons do sometimes actually hitch a ride on bears. Bears don’t, as long as the raccoon doesn’t try to, you know, hurt the bear, bears generally don’t see raccoons as a threat. So sometimes they’ll let the raccoon just sort of like hang out and take a ride. for a while. So that is not in and of itself an indicator of AI. But what you notice is that, first off, this is found footage, it’s grainy. So there are places where, you know, as it takes off here, it’s not very far from the street. And it then, you know, it hits the truck right here. But take a look at that truck. That truck was, before it ever got anywhere near that supposed short little driveway there, it was out in the street. It’s literally blocking the road. So that seems odd. Also, the bear sort of glances off the thing and seems to knock itself out. And there’s not enough room for the bear to have gotten up enough speed to hit hard enough to have knocked itself out because it didn’t hit its head it hit like its shoulder uh and it might fall over but it would uh it would keep going so this is in fact uh very fake okay all right so um detecting ai voice so i’ve played with ai voice before to show how it can be used for, you know, good purposes. And so I will play a couple of samples here for you before we get started. We’ve got, yeah, all right, we’re going to do that one first. So just listen.
Voice Sample 1 • 01:24:41
Chapter one, down the rabbit hole. Alice was beginning to get very tired of sitting by her sister on the bank and of having nothing to do once or twice.
Aric • 01:24:52
All right, I’m going I’m going to jump ahead a little bit here because I don’t want to play the entire clip, but I do want you to hear a bit more of it.
Voice Sample 1 • 01:24:58
…feel very sleepy and stupid, whether the pleasure of making a daisy chain would be worth the trouble of getting up and picking the daisies, when suddenly a white rabbit with pink eyes ran close by her. There was nothing so very remarkable in that, nor did Alice think it so very much out of the way to hear the rabbit say to itself, Oh dear, oh dear, I shall be late. When she thought it over afterwards, it occurred…
Aric • 01:25:22
Okay, that’s enough for that one. Now here’s the same book.
Stacy Arthur • 01:25:27
Chapter one, Down the Rabbit. Alice was beginning to get very tired of sitting her sister in the bank and of having nothing to do. Once or twice, she had peeped into the book her sister was reading.
Aric • 01:25:37
Okay, now I’m going to jump ahead a little bit. Making a daisy would be worth the
Stacy Arthur • 01:25:40
trouble of getting up and picking the daisy when suddenly a white rabbit, pink eyes ran close by her. There was nothing so very remarkable in that, nor did Alice think it so very much out of the way to hear the rabbit say to itself, Oh dear, oh dear, I shall be late.
Aric • 01:25:56
Okay. And then here’s another sample. This is not a book.
Voice Sample 2 • 01:26:00
Like just when you think you’ve got a handle on the world, bam, everything’s different. Right. It’s like the pace of change is, well, kind of freaking people out. Totally. So today we’re tackling that head on. We’re doing a deep dive into the potential downsides of all this crazy fast technological advancement. And we’ve got a heck of a guide for this one. We do. We’re looking at this thought-provoking essay by Bill Joy called Why the Future Doesn’t Need Us. Which, let’s be honest, is a title that makes you stop and go, OK, what’s he worried about? Seriously. But here’s the thing that makes.
Aric • 01:26:34
OK, that’s enough of that. Now let’s talk about how to detect most AI voice samples. Older AI models have obviously generated sounding voices. There will be a buzzing or a barble pretty much in the entire voice. while it may sound reasonably realistic, you’re not going to mistake it for a real person. But the newest AI models sound much, much better. In fact, they’re even programmed to add in breathing sounds and the occasional filler word, which makes it a little harder to detect whether the voice is real or not. However, in most samples of more than a few seconds, there’s likely to be a slight buzz or warble in the voice at certain spots, even if it otherwise sounds normal. And also, when AI is generating a conversational tone, and it’s not following a defined script, you’re likely to hear the AI express things in ways that typical native speakers probably wouldn’t. For example, if one AI voice says, so my friend passed away the other day, without the sensitive way a normal human might otherwise say that. And then the other AI voice might say something like, oh yeah, instead of expressing sympathy, as you might expect. AI voices are really often used for scammy ads, stuff that you find on YouTube and TikTok for products of very dubious quality. So generally speaking, if I see an ad that has nothing but an AI voiceover, I usually run screaming in the other direction because that means that I’m probably not seeing what I think I’m seeing. Those ads often will follow a script as well, which sometimes the person generating the ad isn’t even a native English speaker and it’s obvious. So you might hear it say things that a normal native speaker would not say. So if you hired someone to read a script, there are certain places where there would be ad-libbing by a professional where you wouldn’t get it from AI. So things like missing contractions in places, because most people, when they’re using normal speech, will use contractions for the most part. They use odd phrases or words spoken in ways that a native speaker just wouldn’t. Okay, let’s see if we can go here. Okay, let’s take a look at this. All right. Okay, let me play this for you. We will play it in VLC. Okay. All right. Let me show you. That will go right here. Okay. So this I’m going to show you is a video I took off of YouTube. It claims to show you how to use AI to generate a video script for products that you are interested in.
Voice Sample 5 • 01:29:53
You need these basic things. A script, the visuals, a voiceover, and some music. I’m going to use ChatGPT for the script, MidJourney and Runway for the visuals. I know they are paid, but you can use Leonardo and Pica Labs if you are looking for free alternatives. For the voiceover, I am using Fleeky. You can use any text to audio of your choice like Eleven Labs, Murph AI etc. And for the music I will be using free open source model called MusicGen. Now first we need a script. We will ask ChatGPT to write us the script. And just like that our script is ready. Now let’s ask ChatGPT to give us some image ideas for our video. Write a prompt like this, and it will give you detailed video shot descriptions. Now just copy and paste these image descriptions into any text-to-image generators. For example I’m using Midjourney. Just type in the image description, and voila! You get images like these. Download the image you like, and repeat the process for all other shots. If you don’t have Midjourney, you can do the same in Leonardo AI, and you will get the desired images. Now, let’s add some life in these images. For that, I’m gonna use Runway. They give 125 credits in the free trial. Try for free. Select Image to Video. Click on Upload a File. Select any of the images that we created, and click on Generate. Give it a few seconds and it will generate a video like this. You can use Pika Labs if you don’t want to use Runway, and it will create similar results. Upload your image and give it a prompt, and your video will be ready. Repeat the same steps for all the images and download the videos. Now for the voiceover I’m using Flicky AI. The website link is in the video description so don’t worry. Just sign up with your email. Click on new file. Select the language. Give it a file name. Select audio only and submit. Now paste your script here. Select the narrator’s voice. For example I liked Davis’s voice. Flicky helps you create audio and video content. Let’s apply this voice to all the voiceovers. Change the rate or pitch of the voice. You can also add some pauses if you want. Let’s see how it turned out.
Voice Sample 3 • 01:31:55
The symphony of flavors, a masterpiece of taste, a burger beyond imagination. Supreme sizzle, where every bite tells a story.
Aric • 01:32:02
Awesome. Let’s download this. Now for the background music, I’m using a free music generator called MusicGen. Just describe your music, something like upbeat, energetic music, and click on generate. Let’s see how it turned out. Click on download and we have our background music. Now let’s bring it all together in a video editor. I’m using CapCut as it is simple and has the sound effects I’m looking for. Let’s import all our assets. Arrange the videos on the timeline according to the script. And in the audio section, go to sound effects and you will see the food section. Here select the sound effect for your scene like chopping, sizzling meat, etc. Add these sounds to the timeline. Add some transitions, the background music, and the voiceover that we created with Flicky. Let’s see how it turned out.
Voice Sample 3 • 01:33:00
Experience a symphony of flavors, a masterpiece of taste, a burger beyond imagination. Share the joy, share the taste. Supreme sizzle where everybody tells a story. Okay.
Aric • 01:33:12
Now that is disgusting. Now, the interesting thing about that particular video is none of it’s real. The entire thing is made up by AI from start to finish. There’s no human involved. There’s nobody, there’s no human telling you how this works. I mean, some of the services they show you are real and that’s what they’re really trying to get you to do is to sign up for various services. So, and that ad is terrible. I mean, you saw the supposed chopping and the onions were just sliding out of the way. It was very weird. So probably the most frequent use of AI these days, thanks to Amazon, which has really started pushing it for self-published offers, is AI narrated books. And they tend to lack real feeling in the voice and energy, if any, doesn’t ebb and flow like a natural human voice would, particularly if you’ve got someone who’s really skilled at doing that work the way they would read. AI voices also don’t tend to get timing right. Sometimes they will pause too long or they won’t pause long enough. They’ll brush through a section they shouldn’t have. you also have to beware famous voices it’s very easy and tempting for scam artists to make AI voices that sound like world leaders or celebrities here it is I think this is it here so I’m going to play you a sample that I just created
Jim • 01:34:52
hi I am Barack Obama and I love attending Princeton Macintosh user group meetings They are so informative.
Aric • 01:34:58
So that was literally just from a free tool. And I just typed in the text and hit create. And it generated that for me. So, you know, is it perfect? No. You know, but it does a reasonably good job of sounding like Barack Obama. And of course, famous people are very common. And so because it’s easy to, you know, oh, there’s somebody I like or dislike or whatever the case may be, but it gets you engaged with the content in a way that like, you know, a voice from somebody you don’t know might not. So let’s go back to the original samples. Where here? Roll back up here. OK, here and back to the voice. Okay. And so this first chapter one, down the rabbit hole,
Voice Sample 1 • 01:35:54
Alice was big.
Aric • 01:35:55
This was a an AI fake that I generated a long time ago when we were first talking about AI voices. And this is from 11 Labs. It sounds reasonably good, but not perfect. You know, the text really just doesn’t come to life with an AI voice because it just doesn’t have the right inflection.
Frank • 01:36:19
Can you tell it to add inflections?
Aric • 01:36:21
You can add it to a point, but it’s very, very difficult to get it to read. Because again, if you’re giving it a script, you have to basically program in, okay, you know, use a filler word here or whatever. So it can be very, very difficult to make it read in a way that sounds believable. And for the most part, when you’re using AI voices for voiceovers, it’s the path of least resistance. You’re not doing this because you want to spend real time and effort editing a product. So you might spend a little bit of time doing that, but it can be a lot of work to get it. And even so, there are some things you just can’t get an AI voice to do if it hasn’t been programmed properly.
Jim • 01:37:13
The breathing was so consistent that it was very odd to me on that one.
Aric • 01:37:19
Yeah.
Jim • 01:37:19
A little breath at the end.
Aric • 01:37:21
For sure. Now, the second one, which is also a book.
Stacy Arthur • 01:37:25
Chapter one, Down the Rabbit. Alice was beginning to get very tired of her sister in the bank. However, the pleasure of making a daisy would be worth the trouble of getting up and picking the daisy.
Aric • 01:37:35
This one is real. I actually got this from Apple Books. They were giving it away for free some time ago. It’s not my favorite narrator, but it is a real human who read the entire story. And you can tell she doesn’t have that bright and realistic way of speaking when somebody is speaking that gives it more of a natural feel. But that’s just because she, I think, lacks a bit of skill in that. But she is, you know, reading competently. And there’s no buzz or hum or anything weird. They actually did record that in a sound studio.
Jim • 01:38:25
I noticed a couple of the jumps between phrases seemed off. Would you attribute that to editing then?
Aric • 01:38:31
Yeah, the thing is, when you’re talking about audiobooks, just the occasional stop and start again, that’s very common. Because, you know, if you flub a line in when you’re recording an audiobook, you have to stop, go back and rerecord. Most professional narrators say that, like, however long a book would normally take you to read out loud, if you’re, you know, recording it, you know, say for Audible or something like that, you can expect it to take at least double that amount of time because invariably you’re going to flub a line. There’s going to be need to be breaks. You know, you need to go to the bathroom, that sort of thing. You know, it’s not all done in a single take, especially when it’s a longer book. So that’s not uncommon. But but you get when the voice sounds dead, like there’s very little inflection. Right. Then it’s a good sign that like you’re probably using AI. Now, this last one, Is this real or AI?
Voice Sample 2 • 01:39:32
Like just when you think you’ve got a handle on the world, bam, everything’s different. Right. It’s like the pace of change is, well, kind of freaking people out.
Aric • 01:39:41
Totally.
Voice Sample 2 • 01:39:42
So today we’re tackling that head on. We’re doing a deep dive into the potential.
Aric • 01:39:47
Okay. Is that real or AI?
Jim • 01:39:50
Highly scripted was my reaction if it was real and likely to be AI because of that.
Aric • 01:39:58
It is, in fact, entirely AI. It was made by the Notebook LLM from Google. And there, the Notebook’s claim to fame is sort of like, you can feed it notes, and it will generate like a podcast with multiple hosts talking about the thing that you fed it. So in this case, they fed it this piece of, you know, text. And now the AI has generated a podcast to talk about the topic. But if you go back a little bit and listen.
Voice Sample 2 • 01:40:38
Is, well, kind of freaking people out.
Aric • 01:40:40
Totally.
Voice Sample 2 • 01:40:41
So today we’re tackling that head on. We’re doing a deep dive into the.
Aric • 01:40:45
Okay. Did you hear the word dive? Yeah. Go back and play it again.
Voice Sample 2 • 01:40:50
Haste of change is, well, kind of freaking people out.
Aric • 01:40:53
Totally.
Voice Sample 2 • 01:40:54
So today we’re tackling that head on. We’re doing a deep dive into the…
Aric • 01:40:59
You hear how it warbled there? Yeah. That’s a clear sign of AI. Either that or something went seriously wrong with their recording equipment. But no, it’s…
Jim • 01:41:10
Or somebody’s tried to do a takeoff of Max Edrum.
Aric • 01:41:14
Yeah, right. Exactly, from the 80s. So yeah, this is in fact fake. They do a pretty good job of, you know, sounding like real people. much better than most models uh but again it’s not perfect um and honestly if you listen uh to the uh the entire file which you know i provided a link to that earlier um it gets it gets worse as time goes on like the way they’re talking about uh the topic also for example you know they’re like when they when she talks about the um the title of the thing they’re talking about his response uh to that is just a little weird he’s like you know not he doesn’t really respond like your average human right so okay there we go okay now let’s talk about the last thing uh this will be fairly quick um music detecting ai music so uh it can actually be somewhat hard in certain cases to detect AI generated music. But for example, if you go to Spotify, if you haven’t heard of an artist and their music, and you can’t find a website for them, that’s a good sign that they may not be real. But here I will play some samples. Okay, there’s that one. And then here’s the next one.
Voice Sample 4 (Singing) • 01:43:23
That porch swing still leaning on that same cracked stone Mama’s voice on the screen door saying, boy, you ain’t grown I left in a hurry, chasing some far off shot. Every mile between us feels like I crossed a line. Back road porch light, burning like…
Aric • 01:43:52
Okay, we’ve heard enough of that. And then the final one.
Jamie Wyman • 01:44:15
If I told you you are powerful, would you believe me? Would you know that you have everything you could ever need?
Aric • 01:44:26
Okay, that’s enough of that. Okay, so how do you detect AI music? Though it can be hard to detect by someone who isn’t trained to play or read music, AI music doesn’t always do things the way a human would when putting music together. Bad chord progressions, overly repetitive music, which may sound odd given the state of modern music, but it can be even worse than some of the stuff you see from humans. There can be odd artifacts in the music, sudden odd sounds or a bit of an instrument that comes out of nowhere, a singer that doesn’t breathe, etc. Also, AI music with AI-generated lyrics can be a bit odd. Things that sound good but don’t mean anything, or that mean something but aren’t really used correctly in the context of the song, or maybe a song that seems to be building up to a specific theme or idea and then just doesn’t go there. There’s no payoff. That’s not uncommon for AI music. AI-generated voice singing can have similar issues to AI voice in general. AI singing tends to have an overly processed sound, like someone recorded the whole thing with pitch correction always active. AI music is mostly trained on pop music genres and languages, so it struggles to generate convincing music for stuff that it hasn’t been extensively trained on. So, for example, if you want an English language pop song or something like that, like Taylor Swift or Lady Gaga or something, AI can probably handle that pretty well. But it doesn’t do such a good job if, for example, you wanted heavy metal Mongolian throat singing sung in Mongolian. Because the chances of it being extensively trained in that is somewhat unlikely. And you’ll just have to listen to The Who instead. And that’s H-U, not H-W-O. Also pay attention to the artist’s name. If it’s a popular one, is the song you’re listening to part of the artist’s main body of work? Or is it hard to find? If it’s an artist you’ve not heard of, check the web for the band or the singer’s presence. If they have an album, they should have a website somewhere with their contact info and discography. Because again, they’re going to want to sell their stuff. So they’re not going to, you know, just have disappeared off the face of the earth. If you can’t find any information about them, it’s probably AI. If you have the track file, check the metadata for the song or the album. Because not everyone is good at cleaning up the metadata from AI song generation tools. And a lot of the song generation tools like Suno put metadata in the track that denotes that, like, hey, this was generated through Suno, which is a popular, partly free and partly paid music generation service. So all this being said, let’s listen to the samples again. Now this first one sounds sort of classical or maybe new age. Listen closely to that piano. This is AI and it’s not real. And it’s obvious if you listen closely to the piano, because there’s sort of a buzz or a hum in the piano as it’s playing. And it comes and goes in ways that, like, even if it was a bad recording, you wouldn’t expect it to act that way. Okay, so let’s look at the second one. This is the country music sounding.
Voice Sample 4 (Singing) • 01:48:28
name on that same cracked stone mama’s voice on the screen door saying boy you ain’t grown
Aric • 01:48:38
okay is that real or is that ai this is ai yep it’s ai uh entirely ai uh uh for example even the lyrics were generated by ai for this one um and you can tell because it says like Mama’s voice on the screen door, which you would say through the screen door or something similar, but not on the screen door. It’s not like her voice isn’t sitting on the screen door. But in addition, the voice is a bit over processed. And frankly, the song doesn’t really go much of anywhere. I mean, it sounds like it could be kind of a country song, but again, there’s no clear point to the song either. So that is also AI. Now, let’s go to this one that’s sort of like, I don’t know, a little bit more poppy, but not quite. Maybe I would call this alternative.
Jamie Wyman • 01:49:53
Okay,
Aric • 01:49:57
is that real?
Frank • 01:49:57
No, I don’t think so.
Aric • 01:49:59
That is in fact real.
Frank • 01:50:00
It is.
Aric • 01:50:01
Yeah, it’s by an artist called Jamie Wyman. It was a fairly recent release. She’s not a big name. I intentionally picked a small indie artist because I didn’t want you to be able to like, oh, that’s so-and-so. So, yes, this is from an album, like a concept album that she wrote that’s sort of like a rock opera. But, yeah, so this is, in fact, real. The entire thing she used, she didn’t use AI for any of it. So that’s good. All right. So that’s it for the music. Now let’s sort of wrap things up here. While I can’t cover every possible use of AI, you know, hopefully I’ve covered enough of the important areas that you’ve got a better grasp of how to approach and consume content so that you can detect when AI is being used. Use of AI isn’t always bad, but finding ways to reliably tell AI content from human-generated content is becoming increasingly important. and then I will just relink to the PDF and to the related files, so you can play with them yourselves if you want. Okay, let’s go. There we go. All right, now, anybody have any questions?
Frank • 01:51:22
Yeah, I do have a question. If you were a lawyer and you want to have them write your brief, why couldn’t you tell it to prove, to go to the site and verify that it’s a real site, So that they aren’t fighting.
Aric • 01:51:35
Here’s a key problem with LLM models, the chat bot models these days. They do not have the ability to fact check themselves.
Frank • 01:51:49
Oh.
Aric • 01:51:49
Because they don’t know whether stuff is true or false.
Frank • 01:51:54
Right.
Aric • 01:51:55
They just know what they were given. And they’re restringing stuff together based on words they’ve seen. Essentially, AI chat models are just going, okay, given the input I was given, the question that you asked, what’s the likely next word? And what’s the next word after that? What’s the next word after that? There is no fact-checking whatsoever. So you can try to tell it, like, don’t lie to me or don’t hallucinate or, you know, stuff like that. But there’s no way that you can be sure that it hasn’t hallucinated. Some models will actually link to the information that they use to generate the response that they’ve given you. And you can, in theory, go to that link and check to see if what it’s saying about the topic is in fact what was actually meant by the content that they’re using. But very often they link to stuff that has nothing to do with the content that you’re working with. And if you’re a lawyer, being precise is your job. Like that is the one thing you have to do. And it’s not an easy thing. But, you know, lawyers have tools. I’m sure eventually, I mean, there are tools that will help with discovery for lawyers. But in terms of writing legalese, none of it’s trustworthy at this point. Maybe that’ll change in the future, but it’s definitely not now.
John S • 01:53:36
Aric, speaking of fact-checking, early in the presentation, you had mentioned some fact-checking sites. Could you repeat those?
Aric • 01:53:44
Oh, sure. Yeah. So, for example, you know, Politifact is one site that, you know, they while they do focus mostly on political content, because obviously that’s where a lot of people are concerned that whether they’re getting, you know, an accurate picture or whether the politician is lying, etc. Oh, I hadn’t noticed. yeah yeah so there’s that uh and then another one that is very popular is uh snopes s-n-o-p-e-s um it’s been around for a long time they are an independent uh basically media group and that’s all they do is fact check uh they don’t they don’t um they don’t take ads uh you can sign up basically for a membership with them if you want. And that’s basically how they get funded, is that people buy memberships. But they don’t take ads because they don’t want to appear biased.
John S • 01:54:51
Okay, thanks for that. This was a very helpful presentation.
Aric • 01:54:54
Right. And there’s another one called Pointer. It’s like P-O-Y-N-T-E-R, I believe. That’s another group that does fact checking. So if you’re interested, that’s another place you can go. So they do a wide variety of different types of fact checking.
Jim • 01:55:10
Are there any AI tools for facts that give you some kind of idea of how likely it is? For example, it might not be great on political facts because one side or the other is going to have voluminous information on it. But if you’re looking at information that has very sparse information on the Internet, like 16th century American history stories or something like that, or Bulgarian ones, it might say it presents information from that with the same strength as if there was huge amounts of information for it. Or am I missing something that there are sites better than what I’m thinking of?
Aric • 01:55:57
So here’s the thing. Like you’re talking about like double checking an AI’s response.
Jim • 01:56:04
No, I’m thinking I’ve got a lot of data that, you know, I’ve got a strong sense of this is true. It’d be kind of like if you’re thinking of neural networks in the past where you had a strong synapse or a weak synapse. And you might go, oh, well, I’ve got strong synapses for this. So it’s pretty likely to have data for it. So it’s likely to be true. Whereas you’ve got one single thread that gave to that and it might be just bad data. It would warn you that it might be hallucinating this because it doesn’t have very much of a strength.
Aric • 01:56:40
Yeah. So, you know, the key weakness for any AI model is the strength of the training data. The better the training data, the better the results are going to be. The problem with AI chatbots is you can literally ask them anything. They’re not experts in any given subject. You can literally ask them anything, and they’ll try to string something together based on the data that they have been given. But, for example, if in the training data, you ask it like, hey, is Goodfellas a good movie? and all the training data that is seen is just a bunch of negative reviews of Goodfellas for some reason, it’s going to tell you, no, Goodfellas is a terrible movie. You shouldn’t watch it because it’s got all these bad reviews. That doesn’t necessarily reflect reality, but it reflects the training data. And here’s the problem, too, is that the training data is not always carefully fact-checked. you know if you’ve got you know 70 terabytes of data uh there’s not humans fact checking all of that stuff to make sure it’s right right uh it’s just there so you know the chat bot will happily pick up you know uh you know flat earth content or you know stuff like that or we didn’t land on the
Jim • 01:58:11
moon uh in your earlier example i was kind of going if there was one review online that was was negative against Goodfellas instead of a lot of reviews that were negative, you’re going to get the same result from the AI, right? It doesn’t care if there’s very sparse information about it or a dominant amount of information to the same effect.
Aric • 01:58:32
Yeah, if that’s all the data that it’s got, then that’s all it’s going to give you is the data that it has access to. So if it’s only got one thing that seems to match the content you’re looking for, then it’s going to parrot back that information to you right or wrong i’m often looking for stuff that may not even be on
Jim • 01:58:51
the internet information that nobody is bothered to put on there but i’m curious about history or
Aric • 01:58:57
something yeah and ai is not going to be able to help you with that yeah because uh you know if if you if humans can’t find it on the internet uh and it actually doesn’t exist then it’s not of the training data either. So AI is not good. But AI may well give you an answer. It may not be the right answer, but it will put together that something that feels true.
Jim • 01:59:24
What I love in following your example is if you had the Goodfellas, it’s like Goodfellas is a great movie. All the reviews we have are negative on it. And I get like the AI responses for a Google search or something, I sometimes see that kind of response.
Aric • 01:59:41
Yeah, yeah. And that’s literally a problem with the, because there’s not only do they give a model training data, but they usually give it a set of rules of like, here’s what we want you to do with this data. And how they give those initial set of rules affects how the AI model will then parrot back information later. So that’s why get some weird stuff like that. It’s just, you know, the training guide essentially was broken in a way that let the model do a thing like that where it really shouldn’t have. Right.
Jim • 02:00:23
My tendency on anything coming out of stuff is to assume it could be wrong and, you know, use it as clues of what to look for. But the other end of the AI problems, we’re detecting bad stuff, but all the there is denigrating the value of the good stuff that’s out there. Right,
Aric • 02:00:43
right.
Jim • 02:00:44
And that isn’t talked about very much.
Aric • 02:00:46
Yeah, I mean, it definitely is a problem. And that’s why, like, for example, I never go to an AI chatbot to learn things. You know, AI can help you sort of like, get a start. But you can’t know that the words it’s telling you are going to be in any way definitively true. I mean, there are certain things like, you know, if you ask it if the sky is blue and it says yes, then I mean, you can go outside and you can check it and be like, okay, yeah, the sky is blue. It happened to get that right. But if it’s something you don’t know, you can’t know whether the AI is giving you the truth or not. And again, it’s not like nobody programs the AI to intentionally lie to you. You know what I’m saying? It has no consciousness. It’s not trying to lead you astray. It is just stringing words together in a way that hopefully keeps you using the chat bot. So that’s why a lot of chat bots are also very complimentary. So, you know, they’ll be like, oh, I’ve never, you know, when you say things like that, I’ve never thought of it that way. You’re really on to something when, in fact, like, no, you’re off in the weeds and that doesn’t, that’s not a thing. Like, it’s not accurate what you’re saying. But it will be telling you that you’re, you know, the most incredible, insightful speaker ever. You know, because, again, it’s programmed to be complimentary. And when OpenAI recently made a change to one of their more recent models to take away some of that head-patting behavior, where it’s making you feel good about yourself, people got really upset. because they were like, you know, now the model is just saying things and it’s not interacting and it’s not being like a human and I hate it. So they very quickly took the model down and retrained it so that it once again started to, you know, stroke everyone’s ego because, again, people like it when, you know, the bots do that. So, yeah, it’s again, there’s nothing wrong with AI. You can do some amazing stuff with it, but you need to know how to use it properly and what its strengths and weaknesses are. And unfortunately for the chatbots, because they’re a jack-of-all-trades, a master-of-none kind of thing, you can ask it anything, but it can’t be a subject expert on every possible topic. But it will talk to you like it’s a subject excerpt on any given topic that you give it. Unless you ask it for something that has been programmed to not tell you. For example, if you ask it to make a bomb or something, you know, like that. It’ll be like, nope, sorry, I won’t do that. And of course, people have found a way around that to get around the rules that say don’t give people bomb making recipes. You know, so you just have to be careful how you approach AI. and far too many people that I know use, even my own brother uses AI like a crutch, the chatbots like a crutch. So he doesn’t do traditional searches anymore. He doesn’t look for his own information. He doesn’t evaluate the concept. He just takes whatever AI tells him and that’s the truth. Whatever it is, whether it’s a lie or not, doesn’t matter. The AI said it, so it’s true, which is not how you should approach those sorts of things.
Michael • 02:04:40
I have a question. You started off your talk by citing how there is a great RAM shortage right now because the RAM chips are going into the servers and all these server forms are being built to help with processing and AI. But you mentioned that Apple is discontinuing the low-end model of the Mac Mini and also the Mac Studio. I saw the same article on Mac Rumors. My question is, isn’t Apple a little bit different because isn’t the RAM part of the actual microprocessor?
Aric • 02:05:22
Yeah, but the thing is, is that while that is true, the process that makes the RAM is the, you know what I’m saying? It’s still RAM, whether it’s on the die or not. whether it’s on a separate stick or whether it’s built in to the package that is generated, that comes with the CPU, et cetera, it’s still RAM. And it’s made the same way as other RAM. It’s just closer to ground zero, if you will. So that makes it faster and more flexible. But it is still RAM. And they still have to follow the same processes. they need the same materials to make the ram um and everybody wants to make ram because that’s what ai tools need a lot of they need a ton of ram because that’s where most of the thinking if you want to call it that gets that’s how it gets done is in ram so there’s just a ton of everybody’s demanding like i need ram for this so there’s only a there’s actually only a few major manufacturers of ram uh and they’re just overwhelmed you know what i’m saying so it’s not like apple does not have the expertise themselves to make their own ram and they don’t have the the access to materials they just make a blueprint of like take this blueprint and make it please and then somebody else has to get all the parts together and make it a thing uh and so unfortunately even though apple’s a big player and they it took longer for them to be affected by the general ram shortage, they have been affected. Like the high-end models, you cannot get, I think the Mac Studio, I think, had like a 512 gigabyte version that you could get for a higher end. That’s not a thing anymore because, again, that’s a lot of RAM and it’s hard to get for prices that makes sense you know i’m saying because obviously apple is not you know bumping the price up to astronomical levels and then telling you well if you wait you know a year we can get you that model no they’re just like nobody’s going to do that nobody’s going to pay what they need to pay and they’re not going to wait the amount of time it requires to get this so a lot of those of an
Jim • 02:08:02
exaggeration. Apple doesn’t have trucks out on the beaches to get their silicon that they’re going to use for all their chips, and they don’t have their own galenium or germanium mines or whatever. At some level, they are buying stuff from other people. They found that the expense level is such that they got stuck in the middle of the loop someplace.
Aric • 02:08:26
Yep. And that’s why the Mac Mini and the Mac Studio So some of the models, even if they continue to offer them, like at one point, I remember it was reported that some models, you could still order them, but you would have to wait months for delivery. And then eventually it got to that point, and I think they just took it offline because they were like, people aren’t going to wait that long. Also, another reason why Apple has been having problems is because the Mac Mini and the Mac Studio are excellent AI model tools. So they work really, really well with AI. So a lot of people were buying them, particularly the Mac Mini, because the price for what you got was pretty incredible. you know it’s not being used in like in the some but being used by facebook or anything but a lot of people uh were buying them for like smaller businesses to do ai on-premises ai work um so that’s another reason you know they just they eventually run out of parts
Frank • 02:09:44
so okay any other questions before we uh call it a night yeah the general question uh i write a report on apple word but i need to share it with people that use uh the pc what font is usually the
Aric • 02:09:59
best font to use for that well um generally speaking stick to uh you know things like uh There’s a lot of fonts that are in common. So like anybody who has Microsoft Office is going to have Calibri because that’s the default font for Microsoft Office, whether it’s PC or Mac. Times New Roman is a thing that pretty much everybody’s going to have. Courier. Courier, yeah, is another one that just about everybody’s going to have.
Jim • 02:10:35
um palatino used to be the favorite at our work a dozen years ago or more uh yeah that’s still
Aric • 02:10:43
that’s still pretty common uh i’m trying to think of some of the other uh really common ones i don’t well i think i have enough okay but yeah if you stay if you stick to one of those yeah i guaranteed
Frank • 02:10:55
you’re going to be safe because when i transfer some of my stuff over to the turn it into a dock that it will say these some of the fonts aren’t yeah they aren’t present to the other person’s
Aric • 02:11:07
site yeah i mean you can avoid that by generating a pdf but also if the person needs to edit the document then that makes it problematic so yeah stick with uh times new roman or clebris or
Jim • 02:11:20
something like that one of the strategies to help avoid that is don’t go out there and say i’ve got hundreds of fonts on my machine in which parts of my article should i use each one of those fonts
Aric • 02:11:31
yeah you know honestly like i remember back in the early days of the mac that was super common uh you know you would buy like you know those cds full of those fonts and insist on trying every single one of them in in a document just to you know get your money’s worth right and to feel like it’s giving some visual interest but really what you’ve done is just basically the visual version of sort of vomit, you know, like you’re just throwing everything everywhere and seeing what sticks. So yeah, thankfully, that’s not as common now.
Jim • 02:12:12
Yeah, we had a similar issue in programming the machines that people use to run a refinery. We had somebody come in who was completely outside of that area. And he was like, oh, you’ve given us 16 colors to use if you include black and white. So I want to make sure on each page i figure out where i can use each of those colors appropriately to be the best uh intro keep the page interesting uh and it’s like no no no no you don’t understand the idea is how you can put stuff on the page so somebody can pick it up quickly and respond to it and they react to it appropriately not enjoy the big picture yeah exactly sometimes it’s about getting
Aric • 02:12:52
stuff done you know form over function kind of thing you know so yeah it’s it’s funny how some
Jim • 02:13:01
people will do that but uh okay well thanks for all the extra information and uh we’re going to be the creators of uh next month by asking questions yes yes yes because it’s been a while
Aric • 02:13:13
since we’ve had a good q a because most of these sessions have run uh kind of long so i’m sure there will be plenty of questions to be had. And then we’ll have a bit of a break and we’ll come back and finish the year strong. I can’t believe we’re halfway through the year already. It feels like it just started. All right, guys. Thanks a lot. Good night. It was a good one. Thank you. Thanks.
That concludes tonight’s podcast. See you next month. ♪ ♪