Is AI Safe for Students? Part 2: It Depends on the Tool

So I wanna continue this discussion of is AI safe for students and discuss that research summary. Also, I know I have some parent listeners, part one focus fully on the teacher perspective, and I gave my three things that I look for in an AI tool to use in learning or classroom environments, guardrails, teacher dashboards for visibility, and using it at key instructional moments for normally very short, five to 10 minute periods of time. Just as a summary there.

Since I have some parent listeners, I wanted to kind of add onto that, and I created some dos and some don'ts when it comes to using AI tools with both students in a classroom environment as a teacher or instructor or educator, but also kids if you're parent, trying to help your kids with their schoolwork at home, for example. Let's get into it.

Stanford has an AI hub for education that has over 800 research studies on all things AI and education, and just a couple, two, three weeks ago, they released a summary. This research repository, and they called it understanding the evidence base on AI and K 12 education. It has some important takeaways around the topic of is AI safe for students? First of all, the majority of the studies that were indicated as being high quality causal studies, which were not very many by the way.

I'll talk about that more in a bit, focused on students as the users of ai, not teachers. That's really interesting. Moral of the story is that the number of studies deemed to be high quality causal were low, and of those that were deemed to be high quality causal, the majority of them were about students using ai. As a tool rather than teacher use. One important takeaway from this research summary was, I quote, the research base is growing quickly, but rigorous evidence is still thin.

A k, a, we have some research, but a lot more is needed. That doesn't mean that we shouldn't take. What the research is saying and look at it and use it to guide what we're doing, especially when it comes to emerging technologies where of course it's a brand new technology. The research is going to be a constantly increasing and changing field.

Of the high causal studies, one major takeaway from that research summary from Stanford was that learning improvements were noted when it comes to using AI as a learning tool, essentially with student and learners, but with one really big, but is the caveat there. So learning improvements were noted, but tool design matters. Here's some quotes from that study. Or I should say, research summary across the causal studies.

Reviewed evidence suggests that AI tools designed with pedagogical guardrails, such as tutoring systems that give hints or guide reasoning, show more promising outcomes than general purpose chatbots that provide answers directly. Here are some other quotes from the studies directly, so that one was from the summary. These are from the studies that were deemed as high causal as part of that summary. Not all AI tools function the same way in learning environments.

Tools that support reasoning may help support learning. While tools that simply generate answers may reduce the cognitive effort that supports durable skill development. In other words, using AI with students is not black or white. It depends entirely on the tool that you're using. Tools that guide and not give the answer are those that are linked to learning improvements. The research cites this as a Socratic style tool a Socratic style tool guides not give.

So for example, if you were to give a Socratic style AI tool, a math problem, it would say something like, let's work the through this together. Let's break down the steps. Let's take it one by one, rather than just giving the answer outright, which is what you're gonna get if you're gonna go into say, like a chat GBT esque type tool.

The research summary also says that learning science provides one way to interpret these findings, and it discusses various different educational theories on how an AI tool design would provide opportunity and also risk. If you dig down deep teachers that are listening, if you dig down deep to your teacher prep program, you likely remember your educational theory or educational psychology class. Every person that has a teaching degree has to take this class.

And you learned about all the different theories out there. You may remember, Vygotsky Vygotsky talks about the zone of proximal development or the ZPD. This is the idea that you have to meet students within their zone of proximal development or their ZPD if you are outside their ZPD. The learner's not prepared essentially for the information that you're giving them. This is a great way to think about using a guard, railed large language model with students.

They are not ready for the, give me the answer that these tools do us, that are adults with fully developed brains can also struggle with this. This can be we're overseeing that. This can be challenging for adults with fully developed brains as well. So it makes sense that it would be challenging for. Kids that do not have fully developed brain. Why don't I say that one more time? You get the point. You get the point.

So if you remember from the Part one episode I gave you those three requirements when it comes to using AI as a learning tool, the guardrails, the teacher dashboard, and the key instruction moments. So I told you that the first requirement for the guardrails, the guardrails have different functions. One is to make sure that there's no inappropriate or developmentally inappropriate outputs by the ai.

But another purpose of these guardrails is to set up the large language model that's running them in the background to be in that Socratic style, to guide, not give the answer essentially. That is kind of just a little bit of a look at that Stanford research summary, which I think gives us a lot to think about and really matches what my experience has been and kind of what I shared in that last episode all about. It's all about the tool. It's not black or white. It's not give or don't give.

It's give, but with a giant, but, and the tools are. Very, very important, but also how you use them. So the tools, having the guardrails, the teacher dashboard, but also kind of how you use them in an instructional setting is what I shared in that last episode. Based on this idea that tool design matters, I created some dos and don'ts when using AI for both teachers and or parents. And I know many of you are teachers and parents, so here we go.

I don't like to start with a don't, but it just really made sense to start with a don't here. This one's for teachers and parents. I do not want you to drop your students into an AI model, AI tool without any foundational AI literacy lessons and or conversations. Lessons can really just be conversations, and in this case, when it comes to AI literacy, they really should. Be a lot of conversations around it. So before the tool, the AI model, the AI tool ever goes in students' hands.

They need to know that this is an ai, this is not a person, it is not your friend. I'm gonna talk a little bit more about that in a minute. That's an important one for all ages of kids and adults. Pause, I'll come back to that. , They need to know, here's what it can do. Here's what it can't do. Here's how it can help you. Here's how it can't help you. Here's why you should question it and push back. I've been teaching and talking about that a lot lately. We need to learn to push back on ai.

It's not magical. It's not always right far from it. We know it hallucinates, right? So knowing when to push back on it and in appropriate ways is a really important part. Of that conversation. Now, you might be thinking, I don't feel that I am equipped or ready, or, I'm not a tech expert. I'm not an AI expert, so I can't have these conversations. But that is not the case at all. So much of AI literacy is having the conversations and just talking about all those things.

I just gave you what it can do, what it can't do, what it is, what it is not. You just need to model the right questions. Sometimes it's presenting and taking advantage of like a teachable moment to have the conversations too. So if you have. Surrey on your phone if you have a, I can't say it, A-L-E-A-L-E-X-A or Google in your house. Hey Google machine in your house, AI device. Use those as a way to kind of spark these conversations. Hey, you know, I see you're talking to Surrey.

Did you have, you thought about what Surrey really is, right? You understand, right that Sury is just an ai, not your friend, not a human, right? Like just an example there. I was working with a teacher with elementary age students and we were using a guard railed AI tool and we had had all these conversations with kids, with the kids in the classroom. Mostly the teacher. I was just kind of there supporting and uh.

They started using the ai and of course one student in the room said, it's my friend. And then you know how something like that spreads like wildfire in a classroom. Then all of a sudden, oh, that's my friend. So we were able to use that as a teachable moment and we paused the tool.

'cause that's another thing that these guardrail tools allow you to do is hit pause, and brought everybody together and we created a kind of t chart of what our characteristics of a friend and what are not characteristics of a friend. Okay? And we let the kids build the list and by the end they were going, oh, you are right. I don't, I don't think this AI is my friend. That was an example of a perfect, teachable moment of instead of that teacher. Just saying, Hey, nope, AI is not your friend.

That that was the natural inclination, that anthropomorphism. So stopping, having that teachable moment and letting students think through it and come to that healthy, responsible, ethical conclusion. Their self, moral of the story, curiosity, healthy skepticism, conversations, discussions, those are the foundation of AI literacy. Not being an AI expert, not being a tech expert. You already have the skillset you need to know to question it and just start having those conversations.

Okay. I said I would come back to this idea of companionship. This is becoming companionship with AI, in particular from humans to ai and humans developing really unhealthy relationships with AI and thinking that they are a friend, sometimes more than a friend. We're seeing this in all ages, kids to adults. You've likely seen something about this in the media. It is a major concern, but here's the problem. The knee jerk reaction is to say, as a parent especially, and a teacher, is to say, oh.

Nope, we are not touching that. We're not touching that technology. Nope, nope, nope, nope, nope. We're just not even gonna let it in our lives, and that's gonna shut down the option of that happening or the chance of that happening, I should say. That's not how I want you as a teacher and or a parent to react to this. . I know that seems crazy. You're like, Lindy, you just told us this is a major concern and it's a problem. Why wouldn't we just block it no more?

The problem is artificial intelligence is embedded in everything in our lives already, and it's not just technology, so not allowing the. A-L-E-X-A devices in your house or not allowing students into a large language model on a device, a laptop, a phone, doesn't mean that they're not going to be interacting with this technology. I always use the example I of walking down the street, you're likely having your face scanned by a video camera. And that video has AI built into it.

That , has facial detection, that's ai and that is just walking down the street. So this is no longer like, oh, we can just limit a device. We can limit screen time and we can keep our kids safe from this technology. And this is just the beginning of it. So instead, you need to take the approach of setting up. Responsible, healthy interactions with specialized tools. And I'm gonna talk more about that here in a second.

And having the conversations, and honestly, even if you don't wanna touch the technology, okay, that's okay, but you still have to have the conversation. You can't act like it doesn't exist because your kid, at some point in their life is going to come across an ai and you want them to understand that that AI is not human and you don't want to develop companionship with that ai. Okay. Enough babbling about that.

But hopefully that babbling led to you understanding that companionship is 100% a major concern. Another reason for guardrails is to help with that, by the way. I'm not sure I said that clearly. And the kneejerk reaction is just to not let your kids touch it. But that's actually the better way to go about it, is to teach them responsible healthy use.

Whether that is just through having conversations or ideally a combination of having conversations, interacting with some very specialized tools and technologies. And the great thing about that is it sparks the conversations, just like that example I gave you. Alright, so that is a don't, don't put your students kids into AI models, AI tools without first having some conversations. Those AI literacy conversations are so important and I should say.

If you are getting them into these tools, don't just let it in there. Use them to spark those conversations. Okay, now let's go to a do. This one is for teachers. I want you to use tools that have guardrails and teacher dashboards to give you full visibility into students' progress. These tools use the same underlying AI models like the GPT models from chat, GPT, the Claude models, the Gemini models, but they have this educational scaffolding kind of layered on top of them.

I gave you six of these tools in the last episode and talked much more in depth about the guardrails and the dashboard and the instructional moments and when and all of that. There are a few caveats here with older kids. There might be specific use cases, likely will be specific use cases, I should say, where you could use a Gemini gym if you have access to Gemini. And if you aren't familiar with a gym, it's really similar, pretty much the same as a custom GPT if you're a chat GPT user.

And what gyms and custom GPTs do is they allow you to basically in the background prompt the AI to have a specific goal or have specific knowledge base that it's pulling from. So you can add files to a gym and a custom GPT so that it's pulling from those files. So imagine like a chapter from a textbook or something like that, that students are learning. That's something that you can load in the backend and it's going to use that information.

To give its outputs essentially, so gyms, I would be hesitant to use custom GPTs with students. Maybe if you have GPT for education, you would wanna make sure all the data privacy and check for check is all good. But if you're a Google workspace for education school, Gemini is part of your core service. It is under the umbrella of your , data privacy agreement with Google, et cetera. So it's much cleaner in terms of compliance, if you may.

That would be the way that I would go a hundred times over, over a custom GPT. But even if you're using a gym, you've got all that compliance checked, they're still not gonna give you that visibility that you need. You're not gonna get summary of insights and where to take the instruction next. So again, this is only gonna be used with older students and really sparingly and only in really specific instructional use cases. Now parents, let's talk about you.

There are some freemium guardrail tools that you can use with your kids. I would start with either school AI spaces or Brisk Boost, and that will allow you to do, and you'll just have one. Kid, maybe two, three, how many kids you have in there? Likely one or two in that space or that brisk boost session interacting with it.

So that's a really great option, but since you are working at more of a one-to-one or a few to one ratio compared to teachers that have more of like a 30 to one ratio, you have a couple of other options that are a lot more challenging to do in classroom environments that you can do with only one or two or three kids at home. Gyms and custom gpt, I already mentioned those. That's gonna be a great strategy for you to set up. Let's say your kid is trying to work through their science homework.

You could set up a gym or a custom GPT saying this is their age, this is what they do. You can even load specific information in it, like I said, and it's gonna pull from that. And you're kind of like prompting the AI to guide that learning experience. Claude and I think chat GBT too.

Yeah. Both have what are called projects and what they allow you to do is give instructions and or files, and then you can have multiple chats within that project that have the context of the instructions and the files that it's pulling from. Perplexity has something called spaces that functions very similarly. It's basically the same thing as projects in Claude and Chat GPT.

The other thing you can do, which is similar but not as repeatable as a gym or a custom GPT or a project or a space, is you can set up a general purpose chat bot. So chat, GPT, Jim and I clawed, likely perplexity, maybe especially if you're doing more research stuff, you can set one of those up to guide, not give. By priming the chat with information. So for example, I would open up a new chat in one of those tools, large language model tools, and give it a prompt.

Something like this, you are helping a 14-year-old complete school learning tasks. I want you to guide and not give the answers. You can break down the steps for completion, help brainstorm ideas, give steps to help complete the problem, et cetera. Ensure outputs are developmentally appropriate to age 14. Okay? So now one thing to understand here is when you go into a chat in any of these large language model tools, chat, GBT, Claude Gemini, are the three biggest ones.

Every time you hit enter on a new prompt, it rereads the entire chat. To a point. It does. There does come a point where you hit what's called a context window limitation, where it can only go back so far. But if you haven't overloaded the chat with a lot of stuff, basically it's gonna reread that chat every time. So every time your kid is in there asking it things, Hey, help me with this. Help me with this problem. I'm working through this.

It's gonna reread this prompt that I just gave you, and it's gonna know. I gotta make sure that I'm giving outputs that are appropriate to a 14-year-old. I've gotta make sure that I guide, not give, . This is essentially what a custom GBT and a gym and a project. Are doing, but just within one chat. And if the idea of creating a gym or a custom GPT or a project, you're like, oh, I'm not there yet. If you're newer to these tools, this is a really great way to start.

Just open up a chat, give it a prompt like this. But this is similar to kind of what the prompt could be on the back end of a custom GPT or a gym, if that makes sense. Lemme give you one other example, and I wanna do this one for a writing task, since that is the biggest pain point when it comes to the copy paste idea of large language models. So here's another one. You are a homework helper for a fifth grader. Do not give answers directly.

Instead, ask questions that help them think through their own response. If they ask you to write something for them, redirect them to try a first draft themselves first. The assignment is a blah, blah, blah, blah, blah. A book report on this title, a essay on this topic, . There's another example of what you could do within a chat or within like a custom GPT or a gym or a project. Now the AI has the context. I would test this in different tools. I constantly go between those four.

Chat should be Gemini, Claude Perplexity, well, chate, Gemini, Claude. I constantly go between the three of them and almost always have all three open in my windows. Sometimes multiple different tabs open with multiple different versions of those. Perplexity. I use it very specific times, usually related to research type stuff. I'll talk a little bit more about that in a bit.

So test it in the different ones, the different AI tools, general purpose, large language models, essentially, and see if one does better than another. And try to act like your kid and see what happens and see if you kind of gravitate towards one or another. Also, test different models in these different general purpose, large language models, tools.

So this is something that a lot of people don't understand, especially if you're only using the free versions of these tools, especially chatt PT, the free version of Chatt PT, at least the last time I was in there, I don't think this has changed, only gives you its default model, so you don't even have the option to switch models. I am pretty sure the free version of Claude lets you change between the three different models that they have.

I'm fairly certain, okay, so if you're on the free version of trash, GBT and you're not willing to pay, by the way, I don't wanna go there, but you're gonna get so much out of these tools if you pay for them. Maybe I'll do a whole episode on that. So go into Claude if you're on the free version of chat, TBT, and in the chat window on the bottom right, you see a little dropdown, it'll always default to the Sonet model. If you click that, it also gives you two other models. Sonet is their default.

That's kind of like the go-to for anything and everything. Um, I always call it kind of like my, my go-to default. If I'm not doing anything like particularly specific in my prompt. Haiku is a faster model. I never use it. Opus is their reasoning model and reasoning models. Essentially, they use what are called chain of thought prompting in the background to work through their outputs in more detail. It takes longer to get a response out of a reasoning model like Opus.

Then it does a default model like sonnet. But you usually get much more detailed responses and it lowers the chance of hallucinations because it's using that chain of prompting to kind of check itself. I'm gonna make sure I'm giving you the right information here. So, Claude, yes, you have sonnets are default. It says for most efficient, for everyday tasks. It actually tells you that. When you click the dropdown, then you have Haiku says, fastest for quick answers me.

And then Opus is most capable for ambitious work. So it doesn't state that it's a reasoning model, but from my understanding, it is a reasoning model. Okay. Then Jim and I has their, they call their models and it's in the same spot as Claude in the chat window where you put your prompt in the bottom right corner. If you click that, it'll likely default to fast. So it says fast answers, quickly thinking, solves complex problems. They also have a pro model, which is really good with advanced math.

So if you have a high school math student, or you're a high school math teacher, definitely check that pro model out. And it's also their coding model. And then chat GPT at the moment. They quit. , Notice that Gemini quit numbering and really naming their models. Claude is still naming them and numbering them. Sonnet, Opus Haiku and they each, I think we're at 4.6 right now, 4.6 for Opus and Sonnet, 4.5 for Haiku. Gemini is just calling it Fast Thinking Pro and then chat, CBT.

We have instant that's for everyday chats and thinking. That's for complex questions. So that was a little bit of a tangent sidebar there, but I think an important one to understand what I mean by go test different models and different tools to see which one is going to be the best for whatever task you're trying to help your kid with at home. If you're trying to help them with advanced math, absolutely do not use the default models in any of these tools.

It's not gonna likely be a whole lot of help, and it's probably gonna be full of hallucinations. Likely you never know, but I feel like the likelihood is pretty, pretty high. So I kind of summarize that the default models in any of those tools are going to give really quick, not super thorough answers, but that's sometimes what you want.

The reasoning models often called the thinking models are gonna take longer to answer, but typically have more thorough answers with , less chance of hallucinations. Just in general, the Claude models tend to be more creative and better helping with writing. So anytime I am ideating and wanting AI to help me think through an idea or build on that idea or help me brainstorm, I'm almost always gonna go use one of the CLA models.

Depending upon exactly what type of ideating I'm doing, I'd probably start with the sonnet models, and then maybe once I narrowed it down to one or two ideas, I might then go into the Opus model and dig a little bit deeper with it. Anytime I want AI to help me finesse my writing, maybe reword this a little bit. I always go into the Claude models 100% over chat, GBT, I would say Gemini's kind of in the middle there and depends a little bit on what type of writing.

Gemini's pretty good at more of the technical writing. Now if you're at a school that has Google Workspace for education, you're gonna have Gemini. It's part of your core service. That means it's compliant, all that good stuff. It has some amazing multimodal opportunities that the other. Large language model tools, chat, pd, Claude Perplexity don't have.

So for example, you can click a button, you can add the nano banana model, which is Gemini's image Generation model is really good at creating infographic type visuals so you can turn something they're trying to learn into an infographic. Notebook, lm, I'm gonna talk a little bit more about this in a bit, but Notebook LM is amazing and has all sorts of multimodal stuff in it. So more of the story if you're at a Google school, if your kid's at a Google school.

Jim and I has a ton of power and integrates really well into like Google Docs and the whole Google ecosystem. Alright, so to wrap that up. Remember, don't do any of that without first having those AI literacy conversations. Now for a don't. This one's for teachers and parents. I don't want you to drop students into one of those general purpose chatbots with no structure.

Don't put your students and or kids into a large language model with no prompting or priming or those guardrails set up without any of that set up. The AI is going to default to answer mode and answer mode is what makes it super easy for them to outsource their thinking, especially for young people. Alright, now a do, this is for teachers and parents. I mentioned it. Notebook, LM Notebook. LM is a big do. It is so, so, so good. So if you're not familiar, , it's made by Google.

So it's part of the Google ecosystem and it's built around Gemini, the Gemini models. And essentially it allows you to give it sources of information. So the, that can be a website, it can be a document, it could be a research study, it could be a Google doc that you've created. And then it pulls information. And in the chat, when you're chatting with the ai, it pulls its responses from those sources and even cites those sources of where it came up with the response.

This is called a RAG Model Retrieval Augmented Generation. There you go. There's a mouthful for you. This idea that you give the AI a knowledge base, and it's using all of its large language model knowledge and training data, essentially, but it's specifically pulling its answers from that knowledge base. Okay? So that's what Notebook LM does.

So it's a really great way to teach students how to research and check and verify information from large language models, but making sure the responses are specifically coming from the sources of information that you give it. It also makes this multimodal learning a breeze. So if you're in regular Gemini, you can use nano banana. You can tell it to create an infographic from the information.

You can do that in Notebook L limb as well, and it's gonna take all of that source documentation, that knowledge base, and create an infographic. And there's actually a button that specifically says infographic. You can also create an audio summary that creates kind of a podcast style explanation of the sources. This is fantastic because I always say kids always have earbuds in their ears, so take advantage of it. Right? It'll create like two people talking through whatever content it is.

It'll create a video of whatever the source knowledge is or whatever prompt you get it. It creates flashcards and so many other things. It is so, so good for both teachers and students. Teachers of course. You need to be at a Google school. If you're going to use this with your students, it could still be something that you can use as a teacher tool and maybe you use it to create an infographic and then share it in your learning management system.

But you don't wanna put your students in notebook LM unless you are at a Google school that has approved it. All of the good stuff. Alright, another do, gonna do two dos back to back. And for this one I want to talk about research with one big caveat. I have an entire session that I do on basically the new research skillset with AI and large language models in particular, research is changing a lot and there's a lot that. Humans need to know how to do research using these large language models.

I'm gonna pack the essentials of that session. I do. That's easily 1, 2, 3 hours into just a few minutes here, so I'll do my best if you teach high school and maybe, maybe, maybe, maybe, maybe middle school. It really depends. With middle schoolers on their AI literacy level, there are really specific use cases where teaching them how to research.

Using these AI tools and large language models is really, really important and also can be really appropriate, especially for those high school aged students. This is an example of when you might want to put them into a general purpose, large language model, not necessarily a guard railed model with the teacher dashboard, all of that stuff. But I wouldn't just do this willy-nilly. This would be after tons of AI literacy conversations.

It would be after I've used those guardrail tools with them, I would start with notebook L and teach them how to follow citations. So kind of like scaffolding in, and then before I put them into it, I would've put these large language models up on the big screen and walk through the process and talked and showed them how you verify information, how you follow cited sources, which tools have cited, sources, which tools don't have sided sources, et cetera. That's my big caveat there.

I recommend either Perplexity or Gemini right now for research after notebook L. So Notebook L is number one, but the problem with notebook L is you have to have the right source information loaded into it, so it takes potentially a little bit of adult help to do that. If you wanna teach kids how to research and find appropriate source information and how to verify that's correct, then load into notebook themselves, that's where this comes into play.

So Perplexity and Gemini as of today are the two of those four big tools. Parity, Gemini Chat, GBT, Claude. That automatically default to provide cited links that can be followed to verify and evaluate sources, in other words. So hopefully everyone has at least gone into chat t and done one prompt. If you haven't, please go do it right now and you're gonna see that it's gonna give you a response with the default. And don't change any default models, don't change any default settings.

But what it doesn't do, it gives you the response, but it doesn't tell you where that response came from. It doesn't give you cited websites that you can use to verify that that information is correct and it's not hallucinating. That's what Perplexity and Gemini do as of default. And Gemini, this is a newer ish thing. It hasn't always done this. So if you haven't been in Gemini for a while, it does do this by default now. So when you put a prompt into Gemini,.

When you get that prompt back, you're gonna see, and sometimes it'll look like a little link. Sometimes it'll look like a little number, like a citation and a research paper. When you click on that, it'll give you a website that you can click and open to say, Hey, look, this is the website that is showing the information that I'm also giving you. Part of that is going and looking and saying, okay, this is a credible website, or This is not a credible website. Also, both perplexity and Gemini.

At the end usually, or off to the side, there'll be a sources button and you can click that and it'll show all of the sources that it cited throughout the entire response that it gave you, essentially. So what I do is I put it up on the big screen, I model it. We have lots of discussions and conversations. Around. I talk about the technology, all of the things we compare outputs across the different models.

I show them and I put up chat, GPT and Claude, and I show them, at least as of right now, that those tools do not automatically provide sided information.

There is a web setting that you can turn on in both of those tools that if you turn it on, then sometimes, not always, always, most of the time it'll give you setted links, but you have to know to turn that on and you have to know where to find it, so we talk through all of that and yeah, and then just show them, Hey, look, see how perplexity and Gemini give you these citations. I give them, I kind of have four different strategies that we talk through of how to validate information, et cetera.

One of them is ask different models. Go give the same prompt to different models. These are essentially AI literacy lessons that students must know and they must know prior to getting into an AI model that does not have guardrails. So Notebook LM is a great scaffold to learning how to research responsibly using large language models. Start there. Then we get into Gemini and perplexity. Lots and lots and lots of scaffolding. AKA. We're not just throwing them in there, even high school kids.

Alright, next. This is a don't for both teachers and parents. I don't want you using a general purpose, large language model for research tasks without that source verification built into it. So if you're able to know how to go into Chatt PT, and this is specifically for research tasks. Specifically, you're trying to find information, verify that information, compile that information. That's very different than just going and asking it, what should I have for dinner tonight?

That's not the greatest example, but we'll go with it, so specifically for research tasks, you need that source verification. Ideally, all large language models in the future are going to have this kind of like source verification built into it. I hope we're not there yet. There's gonna be times where you're gonna use it. There's gonna be times when you don't, but research is when you do want to use it.

Now, a do for both teachers and parents do teach your kids to prompt AI in ways that actually help them learn. These are five different strategies, and I've got these in a blog post, which I'll put in the show notes that I teach to personalize a task. So one, if you need something simplified, something like explain this like a third grader, explain this like a kindergartner. I've even gotten down to explain it like a preschooler before.

And it'll take really complex tasks and try to explain them as simply as possible. And then you can build back up from there. But if you're not understanding the complex tasks to start with, it's really hard to do anything from there until you understand it. And then you can go back up. You can have it personalized reading level. So change it to a 700 Lexile score, change it to a 500 Lexile score.

You can use it to translate if you're a multilingual learner or your kids are multilingual learners, you can have it translate to Spanish, you can have it give outputs in. Two languages. So English and Spanish.

If your kid or student is not interested at all in whatever it is that they're learning, and maybe you're a parent and it's 7:00 PM and you're trying to get through this so everybody can go to bed at night and there's no motivation there, you can have the AI related to something that they aren't interested in. So how does whatever concept you're learning relate to soccer or Fortnite or cats or curling? I've been doing that a lot lately. I'm like, tell me how this relates to curling.

Got into it at the Olympics, whatever it is that your kid is interested into. Or you, maybe this, this applies to you as well. Now this is one that I hadn't been doing, but I am now doing more and more for task initiation. So have the AI take. Whatever the learning task they're doing or whatever the homework is for that night and say, break this task into steps to help me complete it step by step. You can even then have it break it into more steps, or maybe it breaks it too far.

You can say, oh, maybe not that far. Bring it back. Wherever your kid is at in terms of their executive functioning skills, have the AI help you break it down step by step by step by step. I'm digging into this more and more of how AI can support executive functioning. I hear from almost every educator I talk to in every school that I'm in, that one of their biggest challenges with students is executive function or the lack of executive function usually.

So when AI tools break tasks into steps, guide students through the process, they're building those executive functioning skills, they're helping kids build those skills, task initiation, cognitive flexibility, metacognition. These are skills that many students, especially neurodiverse learners, generally struggle with, and a well designed. Remember, tool Design Matters. A well-designed AI tool can help kids learn how to approach problems and accomplish big tasks with lots of steps.

This is really hugely helpful for a teacher that has 30 kids in a classroom. You can't sit down with every kid and walk them through each step every second of every day. And some need more steps broken down than others, but this is also a huge help for parents trying to get homework done at night. Maybe you have multiple kids. You're trying to get all their homework done at night, and you can't sit one-on-one with each kid. So executive functioning, helping with task initiation.

Huge. But notice, so I gave you those. What were those five examples? Simplification, changing the reading level, translation, personal connection and task initiation. Every one of those prompts puts the student in the driver's seat, the learner in the driver's seat. They're not asking for the answer they're asking to be met where they are in their zone of proximal development, their ZPD, ? That is using this technology in a powerful way. That is AI literacy and action.

And it's a skill that they're gonna take through the rest of their life, far beyond the homework assignment that they're doing at that moment. On that note, please don't assume that students know how to prompt effectively most. Absolutely do not. In my experience, this goes back to this idea of digital natives, which I really wish we would get rid of. I always say that yes, we may have quote digital natives in terms of they're not scared of the technology.

They pick it up, they start pressing buttons versus adults, typically older adults are quite a bit more hesitant, right? And they're not just button clickers. Kids typically are button clickers 'cause they're used to it. It's what they grew up with. But that does not mean that they know how to use these digital tools and devices for learning and ethical, responsible, productive use. They know how to use it to play Fortnite or Angry Birds or Candy Crush. I don't know. I don't play digital games.

I don't even know what's in right now. I don't think Candy Crush and Angry Birds are in anymore, regardless, whatever it is. Roblox there. That was a much better example. They know how to use this technology to play Roblox in Fortnite. That is very, very different than leveraging technology as a learning tool to help you organize, to help you accomplish tasks. Very different. And the latter has to be taught, has to be modeled specifically.

They pick up on the former pretty quickly 'cause they are quote, digital natives, but they don't pick up on the ladder. The same is now going for using AI tools. They don't know how to use them effectively. The number of kits that I've sat and watched just take the very first. Response that comes from a large language model and copy and paste it, we all know they're doing it, but because they don't know any other, they've been taught to do it differently.

They also don't know the right tools to use or how to prompt correctly to use it in ways that support your learning and your thinking and don't outsource your learning and your thinking. So proper prompting of AI takes teachers, building that into a lesson. Parents, you have to model it at home and both teachers and parents having the conversations no matter who or how it has to be taught. Alright, that is the end. Our dos and don'ts.

Let me give you a quick little recap here because I blabbered, I know I gave you a lot of information and I blabbed a lot for that. There is, did you notice there's a lot to talk about and cover here and the answer of is AI safe for students is not black and white? I hope with all of my babbling that that is the conclusion that you have hopefully come to after the last episode, part one and this episode. Okay, so here we go. Here is an overview of my dos and don'ts.

Don't for teachers and parents. I do not want you to put your students into AI tools without any AI literacy lessons and conversations First. Next do for teachers. I want you to use tools that have guardrails and teacher dashboards to give you full visibility into student progress.

There are a few caveats and special use cases and circumstances with older kids where you may put them into like a Gemini gym, but that is not the norm and not where I want you to start either parents, you can use these guardrail tools with your students. Actually, I did a webinar on this for varsity tutors and I showed them how to use school AI spaces, notebook, limb. And one or two others. But those were the two big ones that I showed them.

So you can use these tools too as long as they're free meal. So I mentioned school AI spaces and brisk boosts were two good ones to start with. You also have a little bit more flexibility than teachers 'cause you're working with much smaller number of kids, so you can use some gyms, some custom GPTs projects, spaces, you can quote unquote prime those gyms and custom GPTs and projects and spaces with instructions that say, guide not give. This is the age of the student.

This is what they're learning. You can also, if you need a baby, step into creating a gym or a custom GPT. You can also just do this within a chat in a large language model and give it a prompt and say, this is what we're doing. This is the age of the student. 'cause you wanna make sure the outputs , are age and developmentally appropriate. This is what I want you to do. This is what I don't want you to do.

If you're gonna do that, test different tools, test different AI models and find the one that works best. Also, I'm not sure I specifically said this, but the one that works best isn't always going to be the same. It depends on the task that you're doing. I gave the example of math. If I'm helping with math, especially more advanced math, I'm gonna use a very different model than if I'm helping with writing. So a don't for teachers and parents.

Don't just drop students into a general purpose, large language model or chat bot with no structure. Maybe, maybe, maybe like upper high school kids. That would depend on a lot of stuff, a lot of different caveats there. But ideally, at minimum, go into the chat and give it some background and kind of prime that chat with information about what you're wanting the student to do.

Also, teacher, I don't know if I specifically said this, but teach your kids how to do this, especially the older kids, how to prime their own chats to tell it. Do this, don't do this. Guide me. Don't give the answer a big do notebook, lm. For both teachers and parents, depending upon for both, really, both what age student you're working with.

This might be more of a U tool where you go into it and maybe generate an audio summary or an infographic, or it might be a tool that you can put students in and have them create infographics and videos and flashcards and audio summaries. Also, notebook element is really great as a first scaffold into teaching students how to research correctly using large language models. That leads to my next do, which is for parents and teachers about research.

It takes a lot of scaffolding, a lot of AI literacy lessons and conversations. Putting them in notebook, lm, putting it up on your projector or your screen in the front of the room, showing them perplexity, showing them Gemini, showing them notebook L, and teaching them how to verify information and evaluate information for research purposes. Especially.

I don't, for teachers and parents, don't just put them in any general purpose, large language model for research tasks without source verification built into it. A do for teachers and parents. Do teach your kids how to prompt AI period, but also specifically how to prompt AI in ways that actually help them learn to simplify, to translate, to help them with task initiation, that's it. That's it. That was a lot.

We covered a lot of ground today and I could talk about so much more, but here's what I want to leave you with. You don't have to be an AI expert to do this. You don't have to be a tech expert. You don't have to have the perfect lesson. You don't have to feel fully comfortable. You're likely never gonna feel fully comfortable because this technology is moving so fast. What matters most, more than the tool or anything else is the conversation. Have the discussions. Talk to your students.

Talk to your kids. Ask them what they're using, ask them if they think it's always right, those conversations are AI literacy, and that is more important than anything right now to help avoid the bad things about this technology. Like companionship, conversations are available to every teacher, to every parent. No login required, no tools required.

Even as the tools keep changing, those conversations are the foundation to making sure that we're using this technology in healthy, responsible, ethical ways. The research will keep coming, but building curious, skeptical, thoughtful humans, that is what we need to be focused on right now.