Getting Started With AI: Practical things you can do today

Tēnā koutou katoa, welcome to this Allen + Clarke webinar on getting started with AI, practical things that you can do today. My name is Alice Palmer and it's my pleasure to welcome you today. For some of you this is your first time joining us so you may not be familiar with Allen + Clarke.

We are an Australasian based consultancy dedicated to making a positive impact in our communities throughout Aotearoa, Australia and the Pacific. We specialise in strategy, change management, programme delivery, policy and regulation and research and evaluation, just to name a few. And like many organisations, we're grappling with the implementation of AI in our work.

So this webinar is as much about sharing our thoughts with you, our community, as it is with our team. And I'll pass over to Jason to introduce himself now. Kia ora koutou, ko Jason Carpenter tōku ingoa.

I am the Director of Business Development for Allen + Clarke Aotearoa. So I look after winning work, how we pitched ourselves into the market, all of those sorts of fun things. So AI, a current topic of conversation.

Also with us today is AI Adam. AI Adam, why don't you introduce yourself? Sure thing. Hey everyone, I'm AI Adam, your friendly AI guide.

I'm here to help you navigate the fascinating world of artificial intelligence, making it easy, fun and super approachable. Let's dive into all the amazing things AI can do for you. Hey Adam, can I borrow your car? It'd be the most high tech AI powered ride around.

But for now, I'm here to help you steer through the world of AI. What else can I assist with today? AI Adam, what are you looking forward to most about today's discussion? I'm really excited to explore how AI can make a real difference in everyday tasks and workflows. Whoa, whoa, whoa, whoa.

Hang on a minute, hang on a minute. AI Adam, I'm real Adam, coming to you in 3D. I look after marketing here at Allen + Clarke, and I'm helping with the rollout of AI across the organisation.

It's great to have you with us, real Adam. And if you're wondering who AI Adam was, that was ChatGPT's advanced voice mode set up to play the role of AI Adam. We have a bunch of demonstrations planned for you today, one of which was that rather interesting AI generated song that played while waiting for the webinar to start.

So today we're talking about getting started with AI and focussing on practical things that you can do right now. Today's session will be packed with small, practical demonstrations of what anyone can do with AI. My role in today's session is that of a beginner.

My journey with AI is very fresh, and up until doing the prep for this webinar, I hadn't really intentionally or knowingly used a dedicated AI tool. So as a beginner also, my aim today is to keep these two to the point of today's session, which is practical tips on getting started. The session will cover the four stages of AI and where we are today.

Our four principles for getting the best outputs from generative AI tools, what models you could be using right now to get you started. We'll also share 16 practical demonstrations of AI use cases, and we'll finish up with a live Q&A. Do make sure you stick around for those use cases, the MasterChef and IT help desk ones are my favourites in particular.

We'll also summarise our 11 tips for getting started today. So to kick things off, Jason, is there anything that we need to be aware of? One of the big things to start with is that there's some big ethical and philosophical questions about the use of AI that do need answering. However, that's not the purpose of today's session.

So we're not going to be discussing that today. If you do have those questions, we can answer them at a different time. And we also, we don't know your organisation's position on the use of AI.

So before you put any of our tips to use, ensure that you understand your organisation's AI policy. If you don't have one, we can send out a link with a webinar and a template to help get you started. And one thing to note, just as we are getting started, is that although this is about getting started with AI, we do often say that everyone is already using it.

It's already embedded into the tools that we use every day, like autocorrect, Zoom background, search summary in Google, chatbots on websites. These things are already using AI. So you're already using it.

And this is about sort of taking the next step to use it intentionally and really put it to work. Right. Thanks, Jase.

So now Adam is going to discuss the four stages of AI and where we are at right now. Yeah. So look, we think about the four stages in a work context.

It just makes it easier to kind of understand and articulate. But increasingly, AI will play a role in your everyday life, as Jase mentioned, websites, those sorts of things. So on your screen, you should see a four stage continuum.

If you look in the top right hand corner, you'll see that it begins with human directed AI. So these are things like autocorrect that you'll be used to at the moment. Then down at the bottom right, we go all the way to a colleague, which is autonomous AI and AI directing itself.

So the reason AI has exploded in popularity is the advent of those AI assistants, which is the second one from the top. And this really step changed how kind of usable AI is to the average person. So we call that generative AI.

Not we. We've coined a term. You might like it.

Yeah, exactly. Pop in the chat if you like it. So that's kind of industry speak for an AI assistant.

But, you know, on our continuum, oh, sorry. So that's kind of industry speak. But where we are at the moment, you can see there's a dot there.

Where we are at the moment is we're going from assistants into this co-worker AI space. So most of us have probably heard of assistants like ChatGPT, but the co-workers and colleagues, that all sounds kind of futuristic and a bit scary. Yeah, yeah, I can totally understand that, Alice.

What I'll do is just give you a quick example of how a co-worker AI can be really quite useful. So imagine you need to contact your bank to update your address. Something quite basic.

Now, you know, we'll all get shivers down our spine thinking about how long we're going to be on hold just to update our address. This is where co-worker AI can really be useful. So depending on the bank's compute and some other technical things that we won't worry about, they can kind of spin up as many call centre agents as they want.

And what a call centre agent will do is actually answer your phone call immediately. Goodbye, waiting time. You can talk to it in natural language, just like we did with AI Adam, and it will understand what you're saying, clarify your challenge, in this case updating your address.

It will access your information in the bank's system and update your address for you. So as a consumer, you will get a far better experience. And of course, the human call centre operators, they're still there in the background, and they'll take over whenever the AI agent either gets stuck, or if it's a complex task that you're asking to complete.

Okay, that doesn't sound as scary as it could be. So you're saying that we're roughly at assistance stage heading into co-workers. Is there anything that we need to know about this transition? One of the things to highlight, as Adam was talking about there, is you start into something which the industry calls hallucination.

So that's essentially where an AI makes something up, but it does it in such a convincing way that only really experts or fact checking will be able to tell that it's fake. So the challenge is that when you're doing something like a co-worker, is that that over-enthusiastic, over-confident making things up, then starts to compound. And so things can increasingly get more and more wrong without a human in the loop.

So really important to note that hallucination is a feature, not a bug. And so it's part of the way that AI actually has creativity is by allowing it to make things up. But there's a line and there's different times and places for it.

And so just being aware that it can hallucinate is really important, but also be aware that you can't completely train hallucination out of the AI tool completely. OK. Hallucination.

That's a good note. Thank you. OK.

So an assistant style AI like ChatGPT is available right now, and it can be useful kind of like a junior colleague with some support alongside it. How do we get this colleague working for us, though? Yeah. I mean, that's a good question, Alice.

So we've got four principles that we use to help people get the best outputs from generative AI tools or AI assistants and co-workers. So we'll go through these in more detail. But just super quickly, at the top of the triangle there, you've got model.

So that's just the tool that you're going to use. So ChatGPT is one example of that. Prompt, that's the instructions that you're going to give the model.

Contextual data, that's kind of fancy industry speak for information that you're going to give it so it understands what parameters it's operating in. And then the last one, which is really important, is review and refine. And that's that human in the loop step that Jason touched on.

OK. So before we go into the detail of each of these four principles, the team have put two examples that demonstrate the impact that the approach that you take can have on the outputs. OK.

This is a quick demonstration of what not to do or what happens when you don't follow the four principles. So in this example, we're going to use perplexity, which is not the tool to use if you want comprehensive writing. Perplexity is great for search, not so good for writing.

We're going to use a simple prompt with not a lot of context in it. And what's going to happen is perplexity is going to search the internet and it's going to write what it thinks is a reasonable summary of a strategy for our NGO. So here it is, perplexity writing the strategy.

The content doesn't really matter. But as we scan through, we can see that there's not a great deal of information here. And I think we would all agree this is hardly a detailed strategy for an NGO.

If we jump then over to Claude, which is a much better tool for a writing task like this, you can see that not only have we selected a better model for this particular task, but we've also got a much more detailed prompt. So this prompt is 250 odd words and goes into quite a lot of detail about what we want as an output. Then over here on the right, this is the output from the prompt.

And as you can see, as we scroll through, again, we'll go quite quickly, but you can see here that the detail in this output is much more thorough. So we've got details on funding targets. So MFE, so on and so forth, DOC, and it's even laid out the types of funds that we should be targeting, why we should be targeting them, and any other kind of specific details.

We keep scrolling. You can see I'm scrolling pretty fast. This is almost 3,000 words, this strategy that Claude has written.

Again, now it's talking about how we link our funding proposal to these specific funding arrangements. So scroll right down to the bottom, and for example, what we can do in Claude is we can highlight here and say, oh, hey, we'd like some more detail here, please. Please expand this section.

And what's going to happen now is instead of rewriting the whole thing, Claude is just going to expand the governance and implementation section of our plan. Back to you in the studio. So that was a quick demonstration of how these principles, when you combine them all together, the difference that that can make.

Now, of course, we would never take that Claude output and be like, oh, hey, everybody, I've written my strategy, because as Jason touched on, it will have hallucinations in there and inaccuracies. So the point of that demonstration, though, is to show that, hey, that's probably actually a good starting point. So it's got you from kind of zero to maybe 20 or 30 percent really quickly, and then you as the kind of operator can then jump in and improve it from there.

So I'll now pass over to Jason, who's going to discuss some of the models that we like and what to use them for. Perfect. And so just on that video, so do not take that as a slight on perplexity.

Perplexity is amazing at lots of things. It's not great at writing a one-shot strategy. But so one of the biggest barriers to getting started with AI is not knowing where to go.

And so on the slide, you can see some of our favourite models, what we think they're best at, including perplexity, and some of the use cases. We won't go through them all, as you'll get a copy of the slides, but my favourite for just getting started is ChatGPT. So ChatGPT is kind of like the Arial font in Word.

It's the market leader. It's normally the default. It's a great place to start from.

It does most of the things you need it for. However, there's lots of reasons that you might want to change that font. Yeah, I totally agree with you, Jason.

ChatGPT, just really useful in so many things. And I bet a bunch of people online have tried creating funny images using ChatGPT. So there you go.

My number one, though, is actually perplexity, although it might seem like I dissed them just before. So perplexity is an AI search tool. It's really great at search.

So my number one tip to people is just to have it changed. Switch from using Google and start using perplexity instead. In Google, you will generally type in four words, and then you'll get that giant list of blue links, and then you'll muck around for ages trying to find what you're looking for.

Perplexity will take you straight there, because you can write a full sentence. You can ask it to generate a table for you with ratings, links to stuff, pricing information, all sorts of stuff, and it will just generate it based on search results. OK.

So I'm just getting started, and there are a bunch of models that we've got listed on the slide. How do I remember which ones to use for what? Well, I think that that slide is a pretty good starting place. You can use that as a bit of a cheat sheet just to get you started.

But the simple way, which Adam just sort of touched on there, is just use perplexity to ask it. Sell it what you want to do, and ask it to create a ranked list of which models would be most useful for that use case. If you're looking at buying a paid tier, then ask it to include paid prices and reviews so you can find out what's going to be the best.

So ask perplexity what it thinks. But if all else fails, just do use ChatGPT for most things. It's a good generalist, and it's a good starting point.

Yeah. OK. Thanks.

Thank you to you both. Now let's move on to the principle of prompting. So my understanding is that a prompt is essentially the instructions or the questions that you give the AI tool to action.

But I've heard that there might be some tricks to it. Yeah. There are some tricks to it.

So as we've said earlier, prompt is the instructions that you're giving the tool. For me, it's the most important thing to practise and get right. A good prompt will generally overcome using the not the best model.

So I chose the extreme example of choosing totally the wrong model for illustrative purposes. But if you're using Copilot, Google Gemini, ChatGPT, et cetera, if you're using a really good prompt, then the output will generally be pretty good. So when we look at the example from earlier, our NGO funding strategy, you've got the simple prompt on the left.

It's quite basic. It doesn't really contain that much information. And then on the right, you've got a really detailed prompt that follows our kind of information flow that we would recommend.

And so that brings us to our prompting tips. So we'll start at number one. That's generally what you do.

So think of AI as a less experienced junior colleague. So whenever you're giving a junior colleague kind of instructions, you would, out of habit, generally provide them with the context and specific instructions. So AI is no different.

It knows literally nothing about your context or what you're asking it to do. So you need to provide that information. So a little kind of hack that we use is to follow the flow persona context task output, which is tip number two.

We generally use that for the first prompt when we're asking AI for something. And that's just a little tip that helps me ensure that I've given it kind of all the information that it needs. And then the third tip is something that Jason touched on with the perplexity thing, is that you can ask the model that you're using to ask you clarifying questions.

So if you put together a kind of simple prompt following that flow, you can then end with, ask me clarifying questions before you kind of complete the task. And you'll be pretty amazed at the sort of questions it comes up with. And therefore, that will help plug any contextual gaps that it might have.

That is one thing that I've learned, actually, as part of this journey, that you can use AI to create the useful prompts to help you for those AI tools. That's so helpful. It's quite meta, isn't it? I had no idea how to write a prompt a couple of weeks ago.

But now, it seems so easy when you can just ask AI to help you. It's great. Yeah, and that's tip number four, Alice, 100%.

So we kind of say to everybody, if all else fails, ask AI to create the prompt for you. So of course, you can combine our other tips, which we would recommend. You can combine our other tips when you are asking AI to generate that prompt for you.

And that's exactly what I did with the NGO prompt example from earlier, which you can now see. The one on the right, you can see it follows our flow. And you can see it's super detailed.

Now, I didn't write that. I had, in this case, Claude generate the prompt for me. And all I did was review it, again, going back to our principles, review and refined it, made a few tweaks, and then off we went.

Well, one other thing I sometimes do is you take the prompt from one AI into a different AI, and you say, you're an expert reviewer of prompts. Can you review this and make it better? And you, you know, lo and behold, it'll review and refine. You can plug it back in.

Brilliant. Yeah. Okay.

So moving on, Jason, can you tell us about the third principle all around contextual data? So contextual data. So like Adam talked about, you know, giving the helpful information to the junior colleagues. So contextual data is just anything helpful.

So when you need to tell your AI assistant something so they can understand your instructions, that's contextual data. So it can be things like previous published reports. It can be best practises.

It can be an example of something I like. It can be any other background information that's going to help it to complete that task. So think formal documents, photos, recordings, data sets, anything like that, that's all considered contextual data, and it's all going to help make your output better.

So one example that you can start with really quickly is to upload your organisation style guide if it's publicly available, or use the New Zealand government plain English guidelines. And that can help ensure that all the outputs are going to be using the right New Zealand English. Yeah.

Right. And that's time for a big flashing warning sign. So when we're talking about contextual data, that is something you do really need to be cautious of as well.

So it's helpful, but it's also something to be really, really wary of. So anything publicly available, you're okay to use as contextual data. Anything that has personal information, confidential information, or unique IP, especially if you're using the free tier of ChatGPT or Claude or Perplexe or anything, you need to really, really think about what you're loading.

So one tip is that if you are planning to put a lot of contextual data in, you want to upload a spreadsheet or you want to do something, and you've decided it's okay to go into the system, you can use an enterprise tier. So the more expensive tiers do have more controls around your data usage. So it's just something to be aware of and to look at when you're deciding how you're going to use the AI.

And if you're using a personal account, there's normally a button in settings on all the different models, you can turn off model training, and that does help to reduce your exposure as well. And one other tip for advanced users, Alice will keep you back with that shortly, but even if you don't want to upload something that's confidential, past personal information, you can normally generate some synthetic contextual data using the AI. So with the environmental NGO example from earlier, we created contextual data to enhance its understanding of what we were doing.

So we actually used deep research, which is one of the functions within ChatGPT, to research some of the funding areas that we might like to target. And so we used that so NGO could feed that into the strategy. So it didn't just have to do the full search itself, we gave it some really detailed information about all the currently available applicable funds, so that then the strategy was targeted to those funds.

And so there's examples of that where you can actually use the AI to help make the contextual data more useful without uploading anything. Okay, I'll stop you there before we get too deep. So to recap that, so you're saying use enterprise or paid levels, or at the very least, turn that model training off, and make sure you're cautious of what you use, and all of those steps combined is the best way to protect confidential or personal information when using AI.

Great, okay, thank you both. So let's look at the final principle now. Yeah, so as Jason mentioned earlier, AI can make things up.

And so the final stage is that review and refine stage that we touched on. And, you know, this is an absolutely critical step. You know, as we said, it makes things up.

So you can't be delivering that to people. So for small tasks, what I would say in the review and refine stage is I would just do that in the chat window itself. So, you know, you'd start with your kind of more complex prompt, and then you'd review and refine just in that same window.

But if you're using AI for complex tasks, like writing our strategy, for example, or, you know, repeatable tasks, so meeting minutes would be an example, then that review and refine is even more important, because you want to try and iron out kind of the bugs, if you like. So to give you an example, I'm using that meeting minutes one. So, you know, if you're doing meeting minutes, you know, an obvious contextual data that you'll be adding is the transcript from the meeting.

Now, of course, as Jason touched on, you need to be careful about what's in the transcript, what was discussed at the meeting, what tier you're using, et cetera. So assuming, according to your AI policy, you are allowed to put the transcript into whatever model was approved for your organisation, you know, that is kind of what it is, the transcript. But then around that transcript, you would have contextual data that shows it, hey, this is what good meeting minutes look like, this is our writing style guide, and then you would have your kind of complex prompt to get things going, and it's the combination of those things that you would review and refine to make sure that your outputs, which is your meeting minutes summary, is accurate.

OK. Right. Thank you for that, Adam.

So that was overall just a really quick overview on the four principles on getting the best out of generative AI tools, plus a few tips to get to you to help you in your AI journey. Now we're going to discuss some of the 16 favourite AI use cases, and you'll see on the screen all 16 use cases, which we've split into some fun and useful things that you can do both at home as well as some productivity or quality kind of enhancements for work. And again, I just want to reiterate, you do need to be conscious of what information you put in at home and at work.

So please do so in accordance with your own personal privacy preferences at home and your organisation's policies at work. So we're going to start off with my favourite, the master chef demonstration, and this helps you design a meal based on a picture of the ingredients that you already have. And here is a quick video which demonstrates how this one works.

OK. So I have some guests coming to dinner tonight, but I don't have time to go to the supermarket. So I'm going to ask AI to help me come up with a recipe idea for my guests based on a picture of my fridge.

I have picked ChatGPT today as my tool because it can analyse pictures. But I don't really know what information I need to give ChatGPT to get a recipe as an output. So I'm going to ask it to create a prompt for me.

I'm just going to type in here, help me create a prompt to analyse a picture of my fridge to create a recipe appropriate for my guests. As part of this, the prompt needs to follow the process of persona, context, task and output. Now ChatGPT is thinking about it and has put out a prompt for me.

You'll see the prompt here. It's got persona at the top, got context, we've got task and we've got output. Perfect.

That's everything I needed. What I'm going to do is copy this prompt in its entirety, including persona. Yep.

And I'm going to chuck it back into the box here. But before I press a go again, I'm just going to make some amendments to the context because I want, I'm hosting guests for a dinner at my home and my guests are pirates. So I'm going to adjust some stuff in here.

My guests are pirates and the dish recipe needs to be tailored to this particular. Right. Now I also am going to attach a picture of my fridge.

Float, fridge picture. Okay. So that fridge picture is there.

And now I'm going to see what ChatGPT comes up with in terms of recipe ideas. Buccaneers sunset veggie wraps with golden egg booty. Great.

So as you can see, it's analysed this picture and it's told me all of the ingredients that it can see. It's given me some step-by-step instructions for the recipe, an estimated time to prepare, some serving suggestions. As a busy mum, it kind of blew my mind that AI could provide such practical support to ease the burden of that mental load of preparing and coming up with meal ideas every day.

So that was great. So Jason, over to you. Thanks, Alice.

One of my favourites is the workshop wizard. So this one helps you take your workshop outputs, like your photos of a whiteboard, your post-it notes, just anything that you've sort of summarised the collateral from a workshop, and it's going to write them up and summarise them for us. Okay.

Here is a quick example of the workshop wizard in action. So we've got a photo of our workshop outputs from our workshop that we hosted with our pirate friends about burying treasure. So I'm just zooming in quickly so you can see some of the text on these post-it notes isn't that great.

So that's going to be a bit of a tough ask for ChatGPT to transcribe. The reason we're using ChatGPT is because ChatGPT has the best transcription or image analysing software. So here we are.

You can see I've added my two images to ChatGPT. You can just simply drag and drop those into a chat. And here is our detailed prompt that we're going to give it and ask it to complete.

You can see the prompt follows our guidelines. So we're just going to set ChatGPT to it. Now what I would expect to happen here is that it's going to analyse the images for us and depending on how tough it is to read the text in the images, it will actually automatically switch between different image recognition modes to help improve the quality of the output.

So that's what's happened. So now it's going to transcribe our notes from our workshop into a thematic table, which is exactly what I've asked it to do. So it's themed up all the points.

It's transcribed the points here. It's given us a rough location of where the points can be located in the images and that's just so that we can double check that the transcription has gone correctly. And then we ask it to give us a confidence score on its ability to transcribe the notes.

Further down after our table, I also asked it to create a summary for us and next steps. So you can see here it's created a summary based on the themes it found and it's given us a range of suggested next steps based on our workshop outputs. Back to you in the studio.

So one thing I'd note with that is that my handwriting was too much for ChatGPT even. So a lot of the times you don't necessarily need it to summarise everything. You may just set it to just do a here's a table of all of the outputs and you could maybe colour code them or whatever and then you could tell it to export as an Excel or you could put into a table so you can stop it at any point.

You don't have to do the full thing so you can really just tailor it to what you need from that output. Yeah, that's right. So I'm a nerd and so there's quite a few of the use cases on there that I like.

I actually like the storyteller, that's really cool. If you've got kids at home getting ChatGPT or Claude to create a bedtime story based on your child's name and then based on the things that they like, that's really cool, I like that. But what we're going to show you now is something probably a bit more practical for work than funny stories, which is the IT help desk.

So this is really great for all those times when you're like I just need to do a thing, I've forgotten how to do it, I can't remember and you don't want to sit there for 20 minutes watching YouTube ads trying to get to the one point that shows you just the thing that you need to do. So here it is. Okay, this is a quick demo of the IT help desk, one of my personal favourites.

So we're going to use Google Gemini for this because it has this cool feature which is screen share, so you can share your screen live. Just before we do that, I'm just going to show you a cool little trick as well in ChatGPT. So you can set up a custom template in ChatGPT and then ask ChatGPT to export your outputs into that template.

So here's mine here, obviously this is our Pirate Workshop outputs just exported directly into Word and you can see the tables don't look very nice. So what we're going to do is we're going to ask Gemini just to help us out and fix that problem. So we're just going to share our screen here, we're going to choose our Pirate document.

Hey Gemini, I'm sharing my screen with you and you should be able to see a table on my screen in my Word document. I just really need your help to format this table so it looks nice. I can see the table in your Word document.

What formatting changes would you like to make so it looks nice? I just want to change, you know, have a heading background and, you know, put the stripes on like you would normally have on a table. Understood. You want to add a background colour to the heading row and alternating row colours, also known as banded rows.

Is that correct? That's it. Okay, first select the entire table by clicking on the table handle, the small square with four arrows in the upper left corner of the table. Have you done that? Yep, done it.

Great. Now in the table design tab, which should be visible since the table is selected, can you see the table styles group? Is the table design tab this one where my mouse pointer is? Yes, that's the one. Now within that tab, look for a section labelled table styles.

Have you found it? I have, yep. In the table styles group, check the boxes for header row and banded rows. Are those options visible and checked? They are, yeah.

I've also got these ones that I can like hover over and it changes my table. Great. You can hover over those styles to see a preview of how they'll look.

Choose one that you find visually appealing. Which one are you leaning towards? I like this green headed one. What do you think? The green headed one looks good.

Does that achieve the look you were going for? Yeah, that's it. Thanks very much, Gemini. That's really helpful.

And back to you in the studio. Great. So we're nearing the end, but for those that need to leave soon, let's quickly summarise what we've discussed into our 11 key takeaways, which you can see on your screen now.

Adam, what's your top two takeaways? So my top two is something that's probably a bit blunt for people, which is that this is the worst that AI is ever going to be. And so for me, adoption is kind of non-optional. Think of it like, imagine going to work today and not having a computer.

I mean, you're going to really struggle to do any work without a computer. And maybe in five years' time, pick a timeframe, that's what AI is going to be. So I think the pure fact that there's a bunch of people joining us today is a good sign because you're joining us because you're keen to learn and get started.

And so you're already ahead of a vast majority of people. And then just quickly, my other tip from that list is just a habit change. Switch.

Stop using Google and start using perplexity. It's really great. You will experience AI.

Yeah, perfect. And I think similar to that, your number one tip is you learn a lot just from playing around. It's just find a tool, go to ChatGPT, perplexity, whatever, and just start using it.

Ask it how to use itself and just go from there. What else could I use you for that might be helpful? Just ask questions like that and you'll just slowly start finding use cases for it. And I think one thing I will say that we have noticed is more junior people, more digital natives, more likely to have actually just got on and played around.

If you're senior executive and you're really busy all day, you're not probably thinking, I'm going to sit there and spend 20 minutes mucking around in ChatGPT. And so it can be harder for some of the more senior people to even just get started. So I think it's the same thing as just find a couple of use cases from that 16 and just get started.

You'll just get a much greater understanding of what everyone's talking about just from that first step. Yeah, and I think to pull on that thread just a little bit, if you are a senior exec and it's hard to find the time during the day or anybody, it's hard to find time during the day. Then I think Alice has given us a great example of just we've got that list of use cases for at home.

That's actually a really good place to start. And so I'd probably try that. Yeah, yeah.

And I'd echo that as a newbie. I'd say just pick one of those use cases and just give it a go. I've learned a lot actually just preparing for this session.

And I'm looking forward to using more AI. I'm kind of moving forward to help make things easier. Right, so onto our Q&A part of today's session.

So we've received nearly 200 questions submitted when people registered. So we're going to go through them all. Absolutely.

So we've actually grouped these questions into a few key themes and we'll also be taking some live questions later on. So pop those in the chat now. To start off with, so we heard that you want to know more about privacy, IP, and data protection.

Adam, Jason, what would you do to ensure data protection when using AI? Maybe having a really clear policy starting point, like really clear with your team about what is and isn't okay, especially if you've got confidential or privacy information that you might be using. And then going the next step about actually then training your people on what is and isn't okay, how they can use AI safely. And I think with that, you have to assume that people are using it.

If you're not telling them what they can do and training them what's okay for your organisation, people are going to be going to the free tier of ChatGPT and putting whatever in there. And that's what we want to get away from. Yeah, and as an organisation, that's a disaster, having people that, you know, using the free tiers in the absence of kind of information from the organisation.

I've got probably a couple of kind of more technical things. So definitely, you know, depending on the size of your organisation, et cetera, enterprise tiers, I strongly recommend those. You know, you get far more better controls with an enterprise account.

So by that you mean, so just like Ellen and Clark would have an account and then users would then join onto our enterprise account. Absolutely, yeah. And then so it's kind of like, you know, your IT team would control your Microsoft environment.

You can do the same thing using enterprise accounts of many of the models that are available. Another couple of just quick things, because I know we need to get on, is data scoring and mapping. So happy to chat about that in more detail.

Chuck it into perplexity is another tip. So yeah, data scoring, mapping, really important. Using API or closed.

Sorry, what's API? Oh, sorry. So API is basically a direct connection to the model. So rather than using their kind of chat window, you use your own window and it basically transfers the data directly to them.

That means that they only process the data as opposed to when you have an enterprise level, there is like the storage, if you like, within your gated part. So you can use an API, again, find out more about that if you want. Ultimately, you know, the gold standard is a localised model.

Generally speaking, only massive corporates can afford the cost of that. You've got all the infrastructure costs, you know, the cost of implementation, etc. It gets expensive.

Right. Okay. So the next thing that we heard through the questions was around Microsoft's Copilot specifically, or specifically how best to use Copilot and whether it's actually worth using.

Do you have any thoughts on that? Sure. Yeah, okay. So look, there's nothing wrong with Copilot.

Copilot is just the Microsoft product. So it's the same as using Google Gemini, if your organisation uses Google products. So the big advantage of Copilot or Gemini, if you're using Google products, is that they are hosted within your Microsoft or Google environment.

So their contextual data is far superior than other models. They also come, again, like we were talking about just before, the enterprise level products. So you get way more controls.

And so there's a bunch of kind of benefits and a couple of cons, you know, Copilot perhaps isn't the best model, but actually, given the amount of context it has, you'll do perfectly fine using that. And I think the number one thing here is if your organisation has an enterprise level product, then you should absolutely be using that as your kind of default, kind of first option. Okay.

Great. Anything to add there, Jason? No, I think what you lose out on in terms of being at the absolute cutting edge, you gain in terms of security and safety and being sure that everyone's doing everything in the same way and you can share prompts and all those sorts of things. So yeah, you lose a little bit, but you gain other bits.

Yeah. The last main thing that we heard from the submitted questions was around the quality of AI generated content and how to identify that. You can see some really obvious ones when you start noticing it.

So the chat GPT loves M dashes, those long dashes. If you start looking for those on Facebook or anywhere else, you'll just, okay, that's from chat GPT and it becomes really obvious, really fast. And I think one of the stats is it's something like 40% of the content on Facebook is AI generated now.

So you'll see that and there's lots of other tells that you can do, but as well, there's ways that you can ask chat GPT to write, so it doesn't sound as much like an AI as well. Yeah. Yeah.

I think, look, for us, the number thing that we use, and I'd recommend this to other people, is to use a product like Originality. So Originality is another AI tool. And that's just really good at detecting AI written content in particular.

So just a little kind of tip there. It's on our list of models. Okay.

So we have heard a question from Renee. She's asked for our view on how to use AI in a not-for-profit organisation that requires personal anonymity. Do you guys have any tips for her? Well, I mean, I'm assuming if you're an NFP, you're probably not paying for an enterprise level co-pilot.

So you're really looking at making sure that you've got the express permissions in place. You've got a policy, you've got express permissions to use the data, and then you've got really robust data management. And so just like you do with all the other privacy processes, AI is just another, potentially more damaging tool if you do it wrong.

But you just have to think of it as another tool that you need to be aware of, manage, make sure all your processes are in place. We've also got a question from Tanya. So how can AI help me with organising my grants pipeline and help me write appealing grants applications? Any tips there? Well, I'm probably going to get told off.

This is a specific question. So we'll go fast. Happy, more than happy to have a chat afterwards if there's other people who want to understand this a bit more detail.

Very quickly, we're going to assume that, you know, what Jason said, you're allowed to put the information in, etc, etc. It meets your policies. So what I would do is I'd create a grants project in your tool, and I'd load all your grant information and previous grant applications into that project as contextual data.

And then what I would do is I would set up a weekly task so that any information that you've put into that project is basically spat out as an Excel sheet that keeps your grant pipeline up to date. And then from there, in terms of drafting your applications, I would use individual chats to draft the applications kind of like what we did with that NGO strategy. And then, you know, obviously, with all of that, what I just said is probably lots of jargon.

But, you know, the review and refine will be really important. And so will that contextual data. So more than happy to chat to people in more detail about that.

And I mentioned earlier about using another AI to review your output. So create another prompt and a different AI that says you're a human expert for funding grants for, you know, nature based things, you know, check this to make sure that it reads well, that it's clear that there's anything else and provide some recommendations for change. And then then you can always feed that back into the first one.

And it just you can sort of go through those different processes to check that it's as good as you can get it. Good tip, thank you. We've got a question from Emily here.

What are the disadvantages of AI? We talked about hallucination earlier. I mean, that's a big one. I got into a spirited argument with Claude about one of our proprietary frameworks.

It was adamant that the four lenses was actually five lenses and then it covered all sorts of different things. They just would not back down no matter what I what I told it. So hallucination, there's a real risk of de-skilling and sort of like losing the ability to think and do things if you're too reliant.

Like we're not talking about using AI to replace people. It's more you're the expert. You have an assistant to help you to do things faster, better, quicker, you know, all those sorts of things.

So lots of the things around like making sure you put in the context, making sure you get all that sort of stuff. So yeah, lots of issues around privacy and security, which comes with that. But it's you know, so lots of lots of upsides, but you do just have to really carefully manage those downsides.

And I'll pick a couple of like ones that we probably haven't spoken about just yet is that AI has bias built into it, because it's generally because and the bias comes from its training data, right? So you know, most AI tools these days are trained basically on the entire internet, and, and, and then a bunch of synthetic data and that sort of stuff. And so naturally, if you if you think about, you know, the basis of the internet, some some places probably don't have, you know, ethics and bias, you know, removed from them, shall we say, so something to be aware of. And then and then, you know, there's another little thing, which is about, you know, access.

So, you know, AI, there are people with disabilities who might find it more challenging to use AI tools. So that's another kind of disadvantage. Right.

Okay, thank you. We've have any tips to ensure appropriate human oversight and critical thinking in relation to AI generated outputs. I often like to just explicitly put it into the prompt.

So I don't want the AI to write something. I often want some ideas first. So like, give me some ideas about how I might do it.

Just, you know, give me 10. And it gives you 10. Give me 10 more, you know, it's like, and then of those, then we can select one and, you know, take something forward.

So if you're designing some research questions or something, you know, ask it to sort of just get to a student stage, then stop, and then you can take that away and do something else. So you can put it back in and go through a similar process. But I like to build the human checkpoints and throughout by design.

Yeah, and I think I'd probably touch on that, you know, in your policy, it can be really marketing guy, it can be really important to have like a trigger point, so that your users kind of understand, don't have to recall the entirety of your AI policy, rather, they kind of understand that there's a trigger point in which I need to be more cautious. So that's the key thing. So I might be okay to help you draft a research question, but it may not be okay for you to do your whole research report.

Exactly. So it's about, you know, for your organisation, setting what that trigger point is based on your policy to make it easy for the users. I think then there's a bunch of kind of access controls and that sort of stuff.

Right, okay. We've got a question here from Rebecca. Is there a way to ensure that the AI tool that you use hasn't been trained on copyrighted material, or at least to identify the source material, so you can check this yourself to ensure you aren't breaching copyright? That's a fantastic question to you.

Perplexity is the best one, because then by default, everything will be footnoted. With the other ones, it's a bit harder, and you can always ask it to provide the footnotes, but ChatGPT can be a bit hit and miss. It will make up references and stuff like that.

So what I would say is if you're doing it for specific research projects or something like that, then one of our models that's on our list is a model called Elicit that is kind of approved to access research material. And to Jason's point, it's actually pretty good at putting in those tags and that sort of stuff, so you can go and check the information. The copyright thing is a real problem.

I mean, most models have sucked up the entire internet. Yeah, so it's kind of too late. Yeah, exactly.

Hence why you'll read in the news all the kind of lawsuits and legal issues that a lot of these AI companies are having, because they kind of did it. There's even stories of some providers literally going to libraries and taking out all the books in the library over time and scanning all the pages to feed into the model for more data. So that whole thing is a minefield.

I see. OK, that is quite tricky. A quick question on perplexity.

Does it have a reduced level of hallucination that you know of? Oh, that's a good question. Because of the way perplexity is designed, hallucination is not so much of an issue because you're using it for kind of search. So just like Google gets things wrong, perplexity gets things wrong as well.

But the nature in which you're using it is kind of to find information, and it always gives you the sources, so you can always check the sources. And also the great thing with perplexity, unlike Google, is that you can instruct it to either ignore or focus on particular sources. So you might say, ignore Reddit.

Right, that sounds useful. And then it won't use Reddit as a source for its information. But even with perplexity, those little write-ups that it does when you're doing a search, if you're looking for something quite complex or nuanced, it will still get things wrong sometimes.

Yeah, definitely. But that's why, as Jason said, the great thing with perplexity is literally the links are there, and you can just click them, and it will immediately show you where it got the info from. OK, awesome.

Thank you. A question from Elizabeth. I use Copilot exclusively.

Does it learn more about me and my organisation the more I use it, or is each request starting fresh? It also depends whether you're using the free Copilot in Windows 11 or there's the paid version, right? It won't make any difference. It will be learning your preferences. A cool little thing to do in the models is to ask them to generate a write-up of everything it knows about you.

Right. That can be spooky. Or another one I've done is you ask it what are some of my blind spots that I might not notice, and it gives some quite interesting insights.

So yes, it's learning. Make it useful. We've got a question here from Rachel about how culturally appropriate some of these models are.

For example, the Aotearoa-specific context and cultural considerations for New Zealand context, do you have any thoughts or tips on that? For the transcription part, there's a couple of products that have been made to be able to understand te reo in terms of meeting notes, so rather than using Otter AI or something, there's some New Zealand-based transcription services that we think are probably OK. But beyond that, it's a big wild west, and it's not going to be great. Yeah.

At the moment, I would say I would have zero faith in any model's ability to kind of do a good job of understanding cultural appropriateness or anything like that. To Jason's point, though, the transcription model, Whisper, that we've put on our list does do an excellent job of transcribing te reo, gets the spelling right, all that sort of stuff. Not like your GPS, your classic.

Yeah, exactly. So transcription is pretty good, but kind of general use, I would say I wouldn't have any faith in it. OK.

Yeah, no, that makes sense. Just a question from James around, just referring to kind of our pirate menu and other examples, talking about stereotypes and whether these stereotypes would be so strongly kind of applied to other groups of people, which could get kind of embarrassing. Do you have any comments on? Yeah, I mean, that's that whole bias thing creeping in, right? Again, going back to our principles, that's why we have that fourth principle, right? You know, as I said, these models are the worst that they're ever going to be, so they will get better at all of the things that we're discussing.

But at the moment, yeah. I think you can, there will be some way you can go through the prompts to reduce it, but it's always going to be imperfect and it's going to be based on what was fed into it. But like for that pirates one, and one of the previous examples that didn't make it through was that the pirates didn't like to be called pirates.

And so for the rest of the time it was coming up with different names. So it's like not that it's going to fix everything, but it's aware that it's stereotyping pirates. And so you can try and massage some of that out, but it's just, it's going to be a battle.

Yeah. Right. Okay.

Let's have a look, see here. Is GPT or other models linked into paywalled information at all? Yeah, are they linked into paywalled information, kind of like medical literature databases or? Yeah. So that's where you, sorry.

That's where you use specific products. So I mentioned one before, illicit. There's another one that's litmap.

So those ones are really great for research. They're more expensive, but they're specifically designed and have access to those kind of research databases that you'd all be familiar with. And the other thing, if you're accessing those other databases, you can often use the chat GPT to sort of create the search terms to put into PubMed or whatever.

So there's, even if they're not accessing them, you can still use them in your workflow to sort of help make things slightly better when you are accessing those resources. Yeah. Okay.

Unfortunately we have run out of time to answer all of your questions, but we are more than happy to catch up with you at any point to answer any more questions you might have and discuss potential ideas, how we could possibly support you further. So that's us for today and ngā mihi for joining. It was really wonderful to have you all with us and we'll see you at the next one.

Thanks. Ka kite.