Transcription
Welcome to this month's AI webinar. And this month, we're focusing on the topic of fact-checking with AI and the hallucinations that AI can have as well, and just ways to try and mitigate that.
In this session, it won't be very long, but it will be jam-packed with a lot of information just to keep you informed on best practices and things to be mindful of when you're using AI. So, what we're going to cover in this session is just first of all, unpacking the risks. So, where AI can go wrong and how sometimes we rely on it a little bit too much. How it can lull us into this false sense of security, thinking that it's providing really accurate information, and what can happen when it is inaccurate and incorrect.
I have done some research as well, just bringing together a list of the major AI platforms, the mainstream ones that people are generally using, and their current hallucination rates, just so you can have a feel for really understanding generally how often AI tools that we use every day get things wrong. We know through our own experiments, I'm sure, but this is just some hard data on that, too.
I've also done a bit of a literature review, looking at what academic literature, peer-reviewed studies focusing on AI hallucinations, and what they've come back with in terms of ways to try and reduce it, and also new research that's happening in the space because it may not be this inaccurate forever, and this sort of, what would I say, inconsistent forever. So, there is hope, but still, we still need to be really mindful.
And then I'll explore some of my own research findings from my book, AI for Strategic Communication, about some things that you can do through the content generation process when you're using AI, again, just to try and reduce those, I guess, instances of hallucination, and then some things to do, some actionable takeaways that you can use in your own AI adventures to just to make sure that what you're putting out there is correct.
First of all, I want to share with you just some statistics that have come from my own research. As I said in my book, Artificial Intelligence for Strategic Communication, I spent last year really investigating how AI is impacting the industry that I'm from. And I did this because there are a lot of people out there using their own experiments as fact. And as we know, everyone's experiences are very personal with AI. Sometimes things can work generally, but I really wanted to stop using what I was doing and go out and see what a lot of people were doing, and then use that research to guide others.
In 11 months, I conducted 41 interviews. And these were with scholars. They were also AI scholars and researchers, also practitioners in strategic communication, marketing, advertising, and then there were representatives from AI tools as well. AI tool developers were in that mix of interviews, just so I could get all different perspectives. And then I conducted a survey with 400 strategic communicators, and they were across Australia, the UK, and the US. Collectively, the sample spanned eight countries.
What I found was something really alarming, and this is why you often see me, if I'm talking on a panel or something like that, I always talk about this issue about over-reliance and not fact-checking and editing AI-generated content because the findings from my study were just really alarming. Firstly, 40% of the people in my survey, and these were strategic communicators, deemed fact-checking AI-generated content as only moderately important, which is absolutely frightening when you consider that these are people representing organizations and their clients, and they're not fact-checking what AI is giving them. That, that statistic.
And then 49% were barely editing the content that they were generating using tools like GPT, etc. On the flip side, 51% were, which was good, but still, that's nearly half people are not doing any sort of editing. And it's a real concern. And it's not because their prompts were good, that they were generating exactly what they want first time. I could guarantee that. It was just that maybe they were strapped for time, or they're just relying on it too much to actually do that extra step and make sure that it was not only correct, also aligned with their brand voice and the organizational values and things like that.
And this sort of over-reliance, and I'm actually, funnily enough, just before this webinar today, I've been working on a paper about a research paper using some of the research from my book on this topic of over-reliance on AI. These are two examples that happened last year, but there's new ones happening all the time. So, this first one, LJ Hooker used AI to generate a real estate listing, and in the real estate listing, it said that this particular property was close to schools and things like that, and it wasn't close to schools at all, and they put that out into the public, and it was wrong.
And then a lawyer in Melbourne was using AI to come up with different cases to use in their proceedings, and it was all made up. I saw actually this time last week a similar thing that happened in Western Australia with a lawyer using AI, and it's just making up all these cases, and they actually got penalized because of that.
Another, I guess it's funny but not too funny, but I saw a TikTok yesterday about a woman crying at the airport because she missed her flight because she went to ChatGPT to ask when it was. Can you believe it? Rather than actually going to the website of the airline or even the airport, just to see when the flight is, she went to ChatGPT. ChatGPT just gave her a random time, and then she missed her flight. These things are happening every day, and you might think, why would people do that? But honestly, people are relying on AI more and more, and it's to their own detriment.
Where can AI go wrong? Really, if you're using it personally, of course, it can just make things up, and you're putting it out on social media. But when you're actually working for a business and in a team, and you're representing an organization, and or or yourself in a business sense, it can really jeopardize your reputation. And it can take years and years to build a positive reputation and build relationships, and it can really take one social media post that's inaccurate to really destroy that. This is why it's so important that there are policies and processes and everyone's aware in your workplace around these dangers.
The other danger is this thing called shadow AI. And if you don't have any guidelines in your organization around this, it means that people are going off and doing whatever they want with whatever tools they want. And that really puts your organization at risk as well. It's really important to make sure that you have these guardrails in place. And it matters more than ever because, as we've said, AI is making things up and false claims, and it can actually put you in a legal predicament, exactly what's happening with the lawyers that I mentioned previously. And it also can ruin when you, you do this, who's actually accountable? Is it the person who's put it out? Is it the AI platform? Who is actually going to be to blame for this? But I can tell you, if you put something out that's completely inaccurate and then you blame ChatGPT, it's not going to look good on you because you're the one who's actually put it out into the public sphere.
The funny thing is with AI is in ways that it does hallucinate. I phrase it, beware of the articulate idiot. Because while it may be trained on vast amounts of data, it's so well-written that sometimes it lulls you into this false sense of security that you think that it's accurate. And what can be funny is it can give you a whole page of stuff that is accurate, and then just slip in something really that is so completely wrong that you will miss it. And that's why you need to be really vigilant in what you're doing with with AI.
It can make up citations as well. If you get it to do, say, use Deep Research to go and do a research project, or even Perplexity to go off and bring back citations and things like that, it can get them wrong, and it sometimes it makes up these citations. And being an academic, there's been instances where in my field, people have been found out when they've actually put their their papers in for review because some of the citations just do not exist. And I think there's been, I'm pretty sure it was one of the big organizations was got in trouble for the same thing recently, and there's been some government reports as well where they found out that clearly AI is being used, and the citations used in there just are non-existent. And it can also make things up.
Google AI summaries are really good, but even those can be incorrect. Now, when you go to Google and you ask a question, it gives you that summary and points you in the direction of some websites if you wanted to seek further, but that summary itself can be wrong. You again, you need to take that into account and just make sure that what you're reading and consuming and putting out is correct.
In terms of the hallucination rates for AI, here they are in layman's terms at the moment. ChatGPT is performing really well. It means that with the new model, GPT-5, it's saying that maybe one answer in a hundred will be inaccurate. However, you still need to check because you don't know which. It's not like it'll go, "Okay, you do 99, and then the hundredth one will be incorrect." It could slip it in. It could be any anywhere in that hundred. That's, it's much more accurate than a lot of them, but it still gets it wrong. The free tier, however, it means about 1 in 20 responses will be wrong. If you pay for the pay for ChatGPT, the hallucination rate is much lower than if you're using the free tier.
Be very careful if you're using free AI tools because the instances of it being inaccurate are much higher. It's about two to 3%. That's nearly as accurate as five for ChatGPT. Claude is about four to five%, starting to get up there. Grok is about 5%, so about 1 in 20 responses as well. And Meta AI, again, one in 20. These are the average ones, and a lot of these tools are free. Perplexity Pro was frightening in that one in six summaries contained inaccuracies. And what is really concerning is Microsoft Copilot, 27%. That's about one in four of its outputs. And what is alarming is that Copilot is the preferred option for many organizations. And so we really need to make sure that we're telling people about this in organizations and letting them know to make sure that they do fact-check because often we believe that, "Okay, my organization has condoned this, and this is what we're, we're using, we think that it's okay." But while the data security may be much more robust in using something like Microsoft Copilot because it stays within the walls of the organization, it doesn't mean the outputs are any more accurate. And in this case, this had the highest rate of hallucination out of all of them.
Be very careful in terms of the research that's happening in the field. Now, there's studies because this is a real problem. Hallucination is one of the biggest problems with AI tools. And if you, I don't know if you saw the launch of GPT-5. I I got up at 3:00 a.m. to see what it could do and everything. And the mention of hallucination, even though it's one of the biggest problems, was just this little tiny mention in there. And they, they, of course, they're going to talk about all the good things rather than focus on the the negatives. And so this meant that they could have talked about, in comparison to the others, how much better it is. But yeah, it's still an issue. There's research being undertaken constantly about trying to solve this problem with particularly generative AI.
Now, in terms of what some of the research has come back in terms of recommendations on how to help manage the problem, it's not going to solve it, but manage it. The clearer your prompts are, it actually reduced mistakes. And I've got some examples coming up as well. Really specific questions were the best. Also, double-checking, making sure that you are double-checking each step of the process that you're using AI to make sure that what it's giving you is correct. Giving it extra context is really important. Then it can use, say, a document or something like that has the information that you need, or you want it to be like to use as a guide. And then also breaking it down. So, if you had a big task with multiple steps, rather than giving the AI a massive prompt with all of that in there, it's better to just give it like a one task, next task, so it builds on from that.
In saying that, though, while this is all new research, the new models haven't really had this kind of research conducted on them yet. So, it's going to be interesting to see when GPT-5 has similar research on it and what the findings will be. But at this stage, these tips are really helpful. Also, training the AI, not that we can actually do that with the public available tools, but when it can be trained on particular data, it will cut it down. This could help if you are creating a custom GPT. And that's another webinar I've done. And a custom GPT, if you're unaware, it's a sort of feature of ChatGPT. They have them in Gemini too, they're called Gems, but custom GPT where you can train it to do a particular task for you. So then when you click on it, it will just do it for you. And you can give it specialized instructions and contextual documents and things like that. And that sort of helps as well because it'll only use those if you prompt it to to generate that output.
However, there are some optimistic findings coming through in terms of mitigating hallucinations. So, there's been studies using fact-checking agents. So, AI agents are there. It's AI that can actually conduct actions autonomously. You can give it a task to do, and it'll go off and do it for you. And so they're using agents to fact-check the outputs of AI before it's then given out. And so this has been pretty successful. They're finding the accuracy is increasing, but it's still, as you can see here from this study from Creps, they're still getting half-truths. It's not 100% accurate. There's still the inaccuracy in there, but it is reducing it.
Some agents have been used to try and flag dodgy references. As I said, AI can make up references, and it fixed nine out of 10 of those in a particular study. That's really helpful. If that sort of research can continue to develop, then it's going to help us in the long run. But still, even when this is getting better and much more streamlined in terms of reducing hallucinations, I still think I'm going to check it. I think I'll always have trust issues with AI in that way.
And this research overall means that again, the prompts make a big difference. AI can be more accurate than people at routine tasks, but it needs to be checked along the way, and double-checking citations, of course. And it's a helpful assistant. It's not going to do the job for you 100%. It's almost like an intern, although it is getting smarter. It really is. And so making sure that when you are generating whatever it is using an AI tool, you use AI as a first draft generator. So, it generates what you want, and then you go and you fact-check, and then you edit to make sure that it aligns with the output.
Even when you are trained on your own writing and things like that, it's still important to put your own, what I like to call, your human sparkle onto whatever it is before it goes out. I was going to say the uni universe, but out to the public, whatever you're using that piece of content for. And just make sure that who, if you're working in an organization, that you have these safeguards in place to work out if it does slip through, who's actually going to be liable for those mistakes when they happen and act to remedy that.
Now, in terms of your toolkit, this has come through my research for my book. And there's different ways that you can and different actions that you can do in the lead-up, during, and after you've generated the content using generative AI that can help to reduce those hallucinations. The first thing is reducing misinformation before it even starts. As I've said, have an AI policy in your organization, and make sure that there are processes in that policy around things like fact-checking and being transparent and ethical AI guidelines and requirements to avoid things like bias and stereotypes and things like that. Everyone is aware of that and train your team on those things. And if you don't know how to do that yourself, and you're say a sole trader or you just want to know, there's lots of different training modules and things online that can help you with that.
Staff can be unaware of the risk of AI hallucination. So, the lawyer from Western Australia who I mentioned previously, he said when he was hauled over the coals for what he did, I'm assuming it was a he. I'll just say they, sorry, they said that, "Oh, they thought they could rely on ChatGPT." There's a lot of, I guess, people are just unsure or just don't know what the risks are. So, it's really important to educate, and it's ongoing training as well. So, not just one-off and that's it. Like, it has to be a continual conversation. And so, once you have your policies and processes, then training your team in those and showing them how also images and audio can be manipulated, and just all the things that you want them to be aware of so they can avoid it, and then use really positive AI practices.
The other thing you can do in that preparation stage, so you've got your policy and everything, is like I said before, supplying contextual documents. Here's an example. I created a custom GPT for LinkedIn and LinkedIn posts. And I went and found my best-performing LinkedIn posts and did screenshots and put it in a document, and then I gave it to the AI to then get an idea of my writing style, the structure, the types of tone, all those sorts of things that I wanted. And then it could draw from that.
In reducing misinformation, just say you want something repurposed from a report or something like that, and there's not any, I guess, sensitive information in that report. You can give that to the AI and just make sure that it's only using that to draw from to to make sure that it's accurate information. Or the other thing is you can get it to create placeholders for the information, and then you put in the specific information that you need, and it can be more like a template structure. That way, you're really on top of that. Never include confidential, privacy, or sensitive information, and check for bias as well. But giving it that contextual stuff is really helpful.
And then with your prompts in this part, I always give it a role. And in that role, I always say that it's highly ethical and accurate. Does it work? I haven't done any testing. At least you've set the tone of what you're expecting. And I I always do that to make sure that it does that. And then I give it boundaries in the prompt. So, if I've given it a contextual document, say a report, I say, "Only use the documents I have provided. Don't add any additional information or external information," because often it just slips things in and adds things you might want, and you just have to be very clear to say, "I don't want that. I just want you to use what I've given you."
I also ask it for sources. It's very good now that it actually does give you sources of the, it gives you links and things to where it's getting information, but I always tend to ask as well. And at the end, I learned this from a few AI experts. They put at the end of their prompts, "The accuracy of this content is important to my career." Now, does it help? I don't know. Why not give it a go? It's probably better to have it in there than not. I just think that's good practice. That's including those sort of safeguards in your prompt.
Now, the second stage, when you're actually generating the content, and it's generated what you wanted for you, then I would include, it's given you what the output, then I prompt it and say, "Review your analysis of the documents I uploaded and identify any inaccuracies." And you'll be surprised how many it will actually come back with. Google Gemini is very good at that. The only problem here is you also need to go, once it gives you those inaccuracies, go back to that contextual document and cross-check, right? You're not relying on the AI to do the fact-checking or accuracy check, but it can actually help to highlight things that it's gotten wrong, and then you can go back and check.
And the other thing is to then identify all the facts, figures, and names, places manually, and do that cross-check. And sometimes it's good to get someone else to help you with the cross-check as well, just to make sure. Sometimes you can get be so close to things that you don't see them anymore. And working in communication, I don't know how many times when we had like a document that was going out to the public, and we'd spent months and months on it, and it still went out with something wrong in it. It's better to get another couple of pairs of eyes on it as well, particularly if it's going out publicly.
And then prompt the AI with each section. I say, "Provide the source for this information and steps you used to arrive at this result," particularly if you're using it for analyzing data or something like that. I get it to break that down. In many of the models now, it does that automatically, like through the process. But sometimes I ask it again because I've heard from some AI experts that often when it's actually detailing what it's doing, it's not actually doing that. It's almost like performative to make it look like it's being transparent. And sometimes I really ask it, "What, what are the steps?" And then it can be actually quite different. And why this actually works is because it isolates parts of the AI, and then it helps to expose those hallucinations if you're not aware. So, this was from a study from Sun Ed from last year. That's what they used to help mitigate these hallucinations.
And then final stage, in stage three, this is the post-production stage of the the process. It's traditional fact-checking. I just go and double-check if there's any facts or figures, just doing things like searching on the internet, checking the source that's given you, because it can often send you into a dead end. Just because there's a link in it doesn't mean it's actually correct. Go, if I'm using say, using it to help me to identify peer-reviewed research, I then go to Google Scholar and I copy and paste the title of the journal article that AI has given me to make sure that it is actually a proper resource. Making sure that you're doing things like that and doing that cross-checking is really important.
Also, checking the dates of things because if something's really old, it may not be very current. Checking who the author was to make sure they have the the qualifications to be saying what they're saying. Just those things that you normally do when you are fact-checking. And I do have a little anecdote to tell you. I was on Google Gemini. I do a live in our Facebook group every Friday. And I went to Google Gemini and I said, "Look, what are some highlights or headlines this week from the news that are focused on AI for strategic communication?" And it came back with these amazing stories, and I thought, "These are going to be great. I'll be able to really give a lot of information here and valuable information." And then I said to it, "Are these correct? Are these true?" And it came back and said, "Oh, no, these are just examples. Just be careful." That's why it's so important to fact-check and not make sure that what it's giving you is, you believe in it 100%.
And so, of course, you still need human oversight. I think that's the key for this. And it's what I really want to drive home too, is it's important for junior staff members not to use AI until they understand what their role is and how to generate documents and pieces of information themselves first. For example, because I work in communication, it's really important that junior staff understand, say, how to write a media release or how to write a report before just going to the AI to make sure that they can understand what something that's inaccurate looks like. If they don't know the task, then they won't be able to identify and do the fact-checking properly. And I think we need to all be the subject matter experts. We can't expect AI or ask AI to do something or a task that we don't know how to do ourselves. And it's really important just to make sure that, yeah, we're the ones in control and can differentiate between that.
Things are improving. As I said, there was a study that I didn't talk about earlier from Gossmar and Dal, and they used three AI agents to fact-check each other in terms of trying to reduce hallucinations, and they did that by 96%. And they did this 310 times just to see how how accurate or how proficient they would be. And yeah, things are happening in the space, but we're not there yet by by any means. It's not widespread.
Oh, and then finally, after you do your fact-checking, make sure that you edit, as I said, and give it that human sparkle because if we don't do that, if we don't prompt it the right way, and then we have the risk of sounding like everyone else. AI at the moment, all the models, they tend to write things in a very similar way, and that you can spot that AI is being used a mile away. I see it on LinkedIn all the time, and often I just can't read past that first line because it's just so over the top, and it's just the same stuff over and over again, and you want to make it look like you wrote it. So, the best way to use AI is to make it look like you're not using AI, and by doing that editing at the end, you're pulling out all of those sort of AI clichés and things just to make it your own. You want to make it sound like you and that it's your work. So, doing that is really important.
And you might think, if I have to do all this stuff, why wouldn't I just write it myself? Maybe, but I often find that AI can really help to speed up a task. It actually helps to fill that space between you and the empty screen. Sometimes getting started is the hardest part. And it's easier to work with something than nothing. It's really good in that respect, but making sure that you're there, adding your own feel and your own voice to it is really important.
In terms of your editing, just remember to watch for clarity, accuracy. You've got that emotional tone. AI can't do that. And as I said, that human sparkle, AI can't have the human experience. They can't really share a human story. Only you can do that. Only you can really put in that emotion in a way that's going to connect with your audience. Making sure that you're mindful of that and putting that in. Making sure that it's also aligned, if you're doing it on behalf of your business or organization, it's aligned with your organization's brand voice, values, uh, branding guidelines, those sorts of things. It's culturally sensitive. AI where it's automatically by default biased. You want to make sure that you are aware of that.
Examples, I can give you so many examples. Not even just culturally sensitive, but just avoiding that bias. If I, I remember it was this time last year, I wanted to create an image to celebrate the end of the first semester. My prompt was not very specific. So, this is what I got back. "Show a lecturer in a lecture theater celebrating the end of semester." And the image I got back was a thin, able-bodied, white, young man in the lecture theater. And it just automatically goes towards those sorts of stereotypes. It's really important to make sure it's much more inclusive.
Also, check that it relates to your audience because you need to be very specific about who you're trying to connect with in that editing stage. You can help that. It's good to include that in your prompt at the beginning too, to help that. Make sure it complies with copyright and also make sure that it's readable and it has those keywords in that are really important for when people are trying to search for you online.
In terms of strategies and safeguards, it's really about, and this is touches on what I already talked about, just making sure that in your organization you have that culture and those frameworks for responsible AI use so that what you're putting out is accurate. And it starts again with the policies and processes and that training, really important. And yeah, I think we've gone through all of this as as well. The main thing is here is you have to still be heavily involved. AI can do parts of a task, but it should never do the whole thing without being human involved in the loop.
In terms of a workflow, you can see here the human begins with it, and then the AI drafts it, human fact-checks it. Then you again, as the human, edit, and human eyes sign off before it's published and then out. You can see the AI part is very minimal, but it's still very helpful, and it can still save a lot of time.
And what to do next? Often, it's a good idea to go and just see if you don't have these policies and processes in place in your organization. See what people, how people are using it. Go and do an audit. What tools are they using? How are they using it? What are they doing? Then you know the state of play, and then you can sort of use them and bring them on board to develop your processes and your policies. Training is really important, and not just, as I said, once on your professional development day. Ongoing training is really important because the technology is constantly moving and things change. So, it's really important to keep that training up to date. And yeah, like having that that comprehensive policy, but also making sure that policy is organic and flows along with the technology as well, because you can't just have the policy and then that's it for 10 years because it'll be out of date very quickly.
And before we wrap up, I just wanted to let you know that I have a free community on Facebook called AI for Strategic Communication Community. And in there, I post AI news. It's really a community where people who are navigating how to use AI and learning about AI come together, and we share and ask questions and just share what we're doing. And I do a free, this webinar as well. I promote that in there. And I also give a really big discount for my AI workshops that I do each month for community members. And if you even just search for the name on Facebook, that will come up, or here's the QR code for that.
And there are my books. I'll just move those on. There's the one I mentioned previously. And then I have a social media book as well. And this new edition has AI through every chapter. And it's about all the stages of social media management process, from strategy development, evaluation, and the and content creation as well.
My next webinar will be on the 23rd of September, and it's going to be focused on using AI to help you with marketing your events and also planning and managing your events because we're heading into that pointy end of the year where there's going to be a lot of events relating to Christmas and end-of-year functions, etc., or even planning for next year's events. I'll show you some really great techniques of how to use AI to make that easier. And I've worked with organizations like Yumandi Markets to help their events team and their marketing team use AI to plan. And they found it so helpful. I think it was like, we planned for their Christmas night market event using AI, and we got enough work done within 20 minutes. It would have taken them a whole week to do in terms of the planning, and saved them thousands of dollars by hiring an external person to do it. This will be a really useful session for those of you who run events.
My final takeaway is, you know, your AI outputs are only as good as your quality control, and it's so important to be vigilant and protect your reputation and credibility by being that human in the loop. You, you are the last line of defense and the first line of trust with the people that you're trying to connect with. Please make sure that you, when you're using AI, that you follow those steps to try and mitigate any inaccuracies.