Why AI Can't Replace Tech Help (Yet)

Hi this is Gary with MacMost.com. Let's talk about using AI for tech help. So there are a lot of different ways to get tech help. Besides official sources, like going through Apple Support, there are lots of online forums, you can ask friends and family, and you can also now use AI. If you ask large-language model AI, like ChatGBT or Google Gemini, questions about tech, you get answers right away. And often they're right, but sometimes not. So it's important to understand how AI can still get things wrong, at least for now. So I'd say most of the time when you ask something like ChatGBT how to do something on, say, your Mac, you'll get a good answer. But not all the time, and it's important to realize that. For instance, here's an example where I asked ChatGBT about how to get rid of the margin that's around windows when you do things like, say, fill the screen. The correct answer is that there's a setting for that in System Settings, Desktop, and Dock. But ChatGBT didn't seem to know about that setting. It suggests other workarounds, but not simply just going to that setting and switching it. Now here's another example where ChatGBT seemed to know about the subject being asked, but didn't know the right solution. This is about adding things like duration and dimensions as columns in ListView in the Finder on the Mac. And it seemed to know that you could do that, but it didn't know how to do it. Here it suggests simply having a folder filled with only videos or only audio to be able to get this. But the solution is in fact to rename the folder something like movies or music to get duration and pictures to get dimensions. But ChatGBT just didn't seem to know that. Even when prompted further, it still didn't know the correct answer. Even though there are plenty of online sources that talk about this, including my own videos. As a matter of fact, it goes and invents an entire idea of like a movies or music template or something like that. It's really much simpler than that and pretty easy once you know how. But someone trying to do this just using ChatGBT to guide them won't get there. Now remember, all these large language model AIs are all based on information that's out there. So things people have written, videos people have made, things like that. So it's just drawing on a body of knowledge that's already there. And sometimes that body knowledge is old. You get the same thing when you search the web for a solution and maybe you get a post from three years ago about it that doesn't know about something new. AI falls into that same trap, sometimes taking the old information and not looking for anything newer. So for instance, here I'm asking about adding more things to Control Center. And a great new feature of macOS Tahoe is the ability to have multiple Control Centers. But there's so much knowledge from before Tahoe, where you can't do that, that the answer you get from ChatGBT is inadequate. It just doesn't mention you can just add another Control Center. This happens all the time, giving you advice about older versions of macOS or iOS. Or if you're a programmer, older versions of Swift or PHP or whatever it is you're using. Now probably the biggest problem is that sometimes large language model AI will make stuff up. People call this hallucinating. And you could see it all the time in tech questions. So for instance, here I'm asking for some terminal commands, some things I can use in the terminal to say change how the desktop looks. And there are some commands. And when asked this, ChatGBT did actually return some useful things. But if you scroll through here, you can quickly come to something that's just made up. It doesn't actually exist. If you search for this FX enable snap to grid, it's not there. Nobody talks about it online. Somehow ChatGBT decided, well, it would be neat if something like this existed. Or maybe it just does. And this might be what it's called. And it just presents it as fact. But if you try it, you'd be really confused to see that doesn't work at all. Now I want to pick only on ChatGBT. So I've tried some others as well. So here's Claude. And I've asked Claude about icons in the dock, like, notice that the calendar shows the current date. What else can do that? And it gives some correct answers, like for instance, activity monitor, it will actually show things in the icon in the dock, you got that right. But then things like FaceTime and Time Machine don't actually do anything special with the icon in the dock, yet it throws them in the list here anyway. Now large language model AIs don't like it when there's simply no answer. If you ask it a question, you'll almost never get a nope, you can't do that. And that's it, which is what somebody like me would tell you. So for instance, here, I'm asking a common question about how do you get the width of a table in the notes app to be something smaller, right? It seems that when you have a lot of columns, it just fills from left to right in the note. Well, of course, the answer is you can't do anything about that. There are no controls for that. But asking Google Gemini about it, it makes up an entire menu item view note body width, and setting it from standard to full width. That doesn't exist. That's not there. It doesn't like the fact that the answer to this is you simply can't do that. So it makes something up. If you find these videos valuable, consider joining the more than 2000 others that support MacMost through Patreon. You get exclusive content, course discounts and more. You can read about it at macmost.com slash Patreon. And one last problem I want to talk about is that AI likes to agree with you sometimes to the point of giving wrong information. So for instance, look how this question is worded. It's fairly typical for somebody to ask something like this. There is no way on a Mac to put the time in the doc. Where else on the screen can I see the time? Now, it's pretty common to make a statement like that as part of your question. And somebody knows what they're doing will say, Well, wait a minute, actually, you can do that. But here chat tbt takes that as a statement of fact, and neglects to tell you that that's wrong. And the obvious solution is to simply say, Oh, you can do what you said you can't do. So here, it's given me a bunch of other options, ways to see the time, never telling me that you can actually put the clock app in the dock, and the icon will actually show you the time. And when I actually called chat tbt out on this, notice how it gave a very human response, reminding us that what large language models do is imitate human responses. They're not fact machines. They're basically response machines. So when I said you can put the time in the dock, just put the clock app icon there. It says correct, putting the clock app in the dock does show the current time. And then it says, Well, there's clarifications and what this really means. And the accurate statement is that macOS does not let you place the menu bar clock in the dock. But if you go back to my original question, that's not what I'm asking at all. I never even mentioned the menu bar in the original question. Now all those flaws also exist with human tech help. Humans can get things wrong. They sometimes make things up when they don't know. They don't like to just give no answer or you can't do that. So they'll look for things that maybe don't quite fit. And they'll sometimes agree with your assumptions, thus making it harder to find an answer. So it's not like these flaws are new. We've always had these problems in getting tech help. But there are some ways where AI help is actually better than human help. Like for instance, you get an instant response when you ask a question. If you post something online, it could be hours or days before you get the first response and maybe longer before you get a correct one. At least with AI, if it gives you the wrong answer and you try it out and find out it doesn't work, you only wasted a few minutes instead of waiting until tomorrow to get the wrong answer. Also with AI, even when it gets it wrong, at least it's usually polite. A lot of times in online forums, you can find people are just downright rude. And with human tech help, there's probably not going to be any improvement in the future. Whereas with AI tech help, I think a lot of these problems will eventually be addressed and we'll find a higher percentage of our questions get answered and answered correctly in the future. But there is one big problem that looms out there and that's that AI gets its information from what people like me post online. So once AI starts answering all the questions and people just go to AI for the answers, well then where will all the new material come from? When macOS 28 or 29 comes out, how many new articles will be posted online about it for AI to be trained on if everybody who posts such articles can't really do it if it's not a viable business model anymore? It's probably going to be that the future internet is going to be kind of a hybrid. Like you may actually go to an encyclopedia or a database for an answer to a specific question, but you may get a book on a subject if you just want to get general knowledge about it. Maybe the same thing for tech help. People like me may still keep posting how-to videos, lists of tips and advice and things like that while AI takes over the direct question answer part of it. But until AI tech help gets better, keep in mind sometimes it could be wrong, sometimes it can give you an incomplete answer, or sometimes it simply won't know but won't admit that. Hope you found this useful. Thanks for watching. Want to know about every new MacMost video each week? Sign up for the absolutely free MacMost email newsletter at macmost.com slash newsletter.