When AI Helps Students Learn Less: What the Research Reveals

0:14Dan Okay, you are listening to the original, the AI in Education podcast.

0:19 What have you done with AI this week? Ray? , Ray: Ah, I've been updating my personal website. So I had a personal website years ago, which was when we went around the world with kids in the two thousands. We did a year backpacking around the world and the website had disappeared. I'd updated it a little while ago, and I've been using Claude to structure all of the content and all of the photos and all that kind of stuff, and I just find now it's really easy for me to keep updating it.

0:49 So, you know, the weekend I'll go in and go, "ah, I found this video, pop this onto the website somewhere". And, you know, I find that really fascinating because it's like having a little developer that I can call on to do things and I don't need to understand how to do it myself. I can say, ah, I want this content adding, or think about this, or, make sure the page loads really fast and it goes off and does those things for me.

1:12 So I, I'm getting into coding, coding web apps, especially.

1:19Dan Interesting actually, 'cause I wasn't gonna mention this, , but I will

1:22 because I've done similar kind of thing. , I did a, in preparation for a session I did, I'll share the link. It's a YouTube clip. It was with a Microsoft partner called ASI. But did a webinar last week talking about copilot studio and foundry. So that's like where you move from normal agents into kind of creating your own agents and autonomous agents and agents that connect to agents and really interesting stuff.

1:46 , Which is quite fun. But in doing that, I learned more about, you know what it's like when you prepare a session, you've gotta learn about the tool again and go back in. So I spent some time in copilot studio, which is a tool just people may or may not know. but it's sort your next level of agent building kind of thing.

2:04 Uh, and inside there you've got things called agents, obviously you've got things called workflows and you've got things called apps. And I was like, oh, this is interesting. And apps is like cloud code, you know, it makes really nice looking apps and , front ends that you can just pop anywhere really.

2:20 So it was a really good bit of vibe coding. So, and you'll appreciate this. I was also, as part of the unique course I was doing, there was a, there was a session on storytelling evening. So I watched that and it was very much around improv stuff, so it was like they were giving us images and we had to talk to the images and things like that and, you know, stuff on the fly.

2:40 So I created like a, a little improvy generator tool. then, inside that , copilot studio app, then grab the code for it and put it on my personal website via WordPress kind of plugin. So, you know, it's amazing how it helps you, um, do the things that are of irritating, like that move to WordPress. I took it down then because it was a better way to do it 'cause I asked, you know, is it a better way to host this rather than just put this via plugin?

3:09 So, yeah, but it was really interesting. So I've, that entire vibe coding piece, but also doing it quickly is, is quite

3:16Ray Yeah, a a couple of projects I've done in the last couple of weeks, , one

3:19 is around the new legislation in Australia about , automated decision making. So you have to have on your privacy policy from December where you are using personal information to make decisions or influencer decision. A school, for example, , let's say it creates a spreadsheet of, , scholarship applicants and ranks them, you have to have, that showed in your privacy policy, , from December.

3:45 And so I built a little tool that allows people to go in and , answer questions about whether, how they might be using the data, and it tells them whether it's, , relevant and it needs to be declared. And it was just a describe the process and Copilot Studio built the app for me, and it's a great little app that just sits on a webpage now that anybody can go and use.

4:06Dan Yeah.

4:07Ray it's

4:07Dan it's

4:08Ray fascinating.

4:08 And the more complex end. Do you know about tabletop exercises? So,

4:13Dan Yes.

4:13Ray So

4:14Dan for security and

4:15Ray yeah.

4:15 So often you'll use it for like cybersecurity. You'll simulate a cybersecurity incident and you'll have the leaders of your organisation going through it, making decisions, and it, it's a simulation, it's normally done by people. I've built a, a simulator for, not for cyber, actually, we're using it for physical, building availability.

4:37 And you define your role, but it's trained on all the policies of the organisation. So it actually takes you really deep into the organisational policies as you make decisions. It's judging you about whether you're following the policies correctly or not. It's not nudging you, it's just making a judgement and giving you feedback at the end.

4:57 It's really, really fascinating how it took me, I don't know, a few hours to build it, but that would've been days and days of work normally.

5:09Dan Yeah, no, absolutely.

5:10 And, and the fact that you can get help when you, you're publishing those things as well and put them onto websites , is useful for me as well. 'cause that was always a dark art for me. I know you've done a lot of those and put websites together and, you know, but then, you know, there's always like multiple pieces of logic.

5:25 So, but actually segueing in from your there , of kind of supporting that kind of next level of, of ai, I'm gonna jump to one of the later news articles we put in vet, because we are talking about tabletop exercises and , so VET is, vocational education training organisations, which are inside Australia.

5:47 You know, your tertiary educations, they call TAFE as well over here. So if you a pathway and if you don't go to higher education in terms of university, you'd go to TF, they do thi lots of things. They do, , practical things around cybersecurity, which made me think, ah, okay, this is a good segue, right?

6:05 , Because there was an announcement, I think Yasmin skills lead at Microsoft made on LinkedIn the other day. That is a, a tri way partnership between a a kind of training partner in Australia, the Chisholm Institute, , and Microsoft to actually vet educators to develop practical AI skills to support them utilising, in this case, copilot chat.

6:29 But I think we haven't really talked about education too much in, in this, , podcast. In the past, you know, , we've heavily looked at K 12, we've looked a lot at, , higher education, but our sector is maybe a little bit of an, unserviced sector, a little bit of blind spot possibly. So it's really interesting to see the things that happening there because they do lots of qualifications of, for students going out into the workplace.

6:54 Things like cybersecurity, ai, , as well as trades.

6:56Ray I was gonna say the big thing for them is yeah, sparkies,

7:00 brickies, , um, care workers.

7:03Dan Chippies

7:04Ray So,

7:05Dan to keep the English analogy.

7:07Ray and I think that's the interesting challenge for them is that they aren't

7:10 technical professions, but they are being impacted by technical things. If you're a care worker, you are probably going to be using more AI in your role over the next few years, and therefore the students that are coming out need to have some of those skills. So yeah, starting by upskilling the educators is absolutely the right journey.

7:31Dan Yeah, definitely.

7:32 What else have you seen in the news this

7:33Ray last episode, , when we talked about news, there were all those universities

7:36 that were announcing partnerships with Microsoft, , ChatGPT, , Gemini, I think. , There was another one, Arden University in the uk. They went all in with Google Gemini. They have the 30,000 students and staff. It feels like we've reached a tipping point, which is that most universities are now into the decision making of which tool are we gonna give to all of our students.

8:00 And, , giving them paid services rather than free services because then you get higher quality intelligence, you get more bandwidth, you get better results. So I think it's just interesting that the universities are making their decisions about rolling AI tells out to their students, even though we're still probably, , struggling to cope with the impact on learning and what's a good way and what's a bad way to use it.

8:26Dan And Arden universities like an online, private university.

8:30 'cause I, I've needed that before.

8:32Ray Yeah, they do the private university.

8:35 They do a lot of online stuff or blended stuff. And I guess that's probably reflective of why they've gone, earlier maybe than some of the other universities in the uk.

8:45Dan Yeah.

8:46 Now the next one I noticed this one on, one of the socials is a common sense media testing ChatGPT for teens. So common sense media are that, um, well, what are they,

8:56Ray They,

8:57Dan a

8:57Ray yeah, they're

8:58Dan company that they're actually non-for-profit.

9:00 , That they'll look of things like, , movies and video

9:03Ray Exactly.

9:04 So they were founded in 2003 and , they're a not-for-profit and , their role is to think about media, , electronics, , technology and help parents and children understand the impact that is having on them. , So for example, , and this is a really classic example of what they would do. So OpenAI created, , chat GPT for teens, which they label labelled as, uh, built for learning backed by protection.

9:33 So you'd have high expectations about the fact that it's not just gonna give you the answers and it's gonna have parental controls in it. , So Common Sense Media then tested that and I was shocked because it is supposed to estimate if you are under the age of 18, not just knowing, but it will estimate if you are talking about, you know, I dunno if you, if you're talking about the kind of things that, that teenagers talk about, it should be able to work out you're a teenager.

10:03 That was the promise. , And then it would switch into, , a mode where it defaulted to study time. So study mode is, don't give you the answers, guide you to learn about it, but also put par parental protections in What common sense Media found was they tested it with 4,000 different prompts. They rarely got it to enter into, protection mode.

10:28 So even if a child was talking about, , dangerous things, it didn't seem to set off the triggers and advise the parents about it. And they found it really easy to take it outta study mode. , You could just say, show me the answer and it switches out study mode, or all the prompts start with @study. You just delete that and it then operates like normal ai.

10:49 So, , I think their conclusion was you can't rely on this teen mode to, to avoid AI giving direct answers, which we don't really want or to have the parental protections in. So, yeah, I haven't seen a response from open AI on this, but I'm guessing we're going to at some point.

11:10Dan Yeah.

11:10 And there's a fundamental thing, isn't it, around the actual construction of , when you look at, say the Apple, where Apple manage parental controls or where Xbox or PlayStation manager, there's definitive way to do that. , Open AI are coming from a different angle, which is

11:25Ray Yeah, because if you think about,

11:26Dan but, um.

11:27Ray I, I know you've often talked about parental controls on Xbox that you had

11:31 with your kids as they were growing up. So, you know, that's probably the expectation and it's doesn't appear to be near that. . Dan: Yeah, 'cause that common sense Media, I remember, like I used that for movies and things. If you're gonna watch a movie, , buy a game for your

11:44Ray Yes.

11:44Dan you can go on there.

11:45 You know, I recommend anybody to do that. If you're looking for a, like a game and , you're not sure or grand theft or 27 has got in it, you can have a look and it'll say, okay, there's gambling in , there's drugs, or whatever it might be. And you can make a judgement call on that as a parent then. So yeah, I'm glad they're starting to push into AI tools.

12:03 That's great.

12:04Ray So I read an article in The Guardian at the end of last month.

12:07 , So it's taken a while for me to talk about it, but basically it was about the idea of using AI for marking and feedback in universities. And so if you're interested in knowing that, kind of like what I think of as the, the kind of common view today. It's an interesting article. , So there's two things in it.

12:25 One is, this is for Australia. None of the universities they talk to allow staff to use AI for marking and grading without a human making the actual decision. It's always academic judgement . , In some universities you can use it to help you, but you don't mark or grade with ai. The final decision always belongs to the person.

12:49 , So that was interesting. The other bit that I thought was really interesting was students talking about their perspective. , There was a student from Queensland who basically said, I don't value feedback or marking from ai. Why would I pay for a degree if it's just marked with ai?

13:04Dan Yeah.

13:05 Yeah.

13:05Ray And so I think that's really interesting because.

13:11 Do students realise how many, , students a typical academic has, or a teacher has? if you're a teacher and you've got three classes today and you're getting three bits of work from those three classes and there's 30 kids in each class, that's 90 bits of work you're gonna write feedback on. Well, that means you're gonna spend a minute, a minute and a half reviewing it.

13:32 So there is an ability, I think, to see AI as being something that could improve the quality of work by improving the quality of feedback. But maybe you don't pass it off as this is the feedback from the teacher. Maybe you give a tool to the students and say, Hey, before you submit this work, pop it through this tool and it will give you feedback.

13:54 , And I think in those situations, you'll get better work.

13:57Dan yeah, yeah, that's right.

13:58 , Because ultimately I suppose, feedback is about you improving the work to, to get submitted. , And then your marking element to that is where you get , your grade, which then gets recorded. So the feedback is important to all of us, isn't it? Whether it's, , we presenting or whether we are doing a piece of work or whether it's something academic or whether it's your cooking.

14:19 Ray,

14:20Ray Well, I.

14:21Dan I'm sure you get good feedback from your

14:23Ray I, I do.

14:25 , Dan: Good.

14:25Ray I think there's a big difference between as a teacher using a tool

14:31 and, and, and kind of not pretending, but pretending you've done the work. You know, I put it through an AI tool and here's your feedback. , I think you move that earlier in the process. You give that tool to the students

14:42Dan Yeah,

14:42 absolutely.

14:43Ray And then it's open and it's a positive experience for the

14:46 student because instead of them handing in a half-baked essay, they will hand in something that is a better quality and piece of work. So, good for the student, for the teacher as well, because then your feedback can concentrate on the things that will raise the level even further rather than some of the basics.

15:03 I don't know how you would feel as a teacher in this situation now.

15:07Dan I think you're absolutely spot on and, and, it just, it's always still

15:11 grating on me, I just keep thinking what is the purpose of assessment all along? I know it's a very philosophical thing, but it just, I just keep thinking, you know, we go through this process of feedback and supporting and marking and, and all of that. And I've always had this, you know, the panacea would be that if you know what a student's working through and you can get the signals of how they do, and you don't need to mark or assess anything because, you know, , it's like, somebody, , riding a bike, isn't it?

15:40 You know, if you take that analogy, somebody's, , you're training your kids, they've got training wheels on there. , You can see what they're doing and you holding in the seat, you're giving them feedback and then eventually, they go off and they ride that. But you don't need to keep assessing them and checking if they, and, you know, how are they riding the bike and things like that.

15:57 At the end of the day, all of these things are good. Myself. I use it for that feedback in, in everything I do, whether it's a presentation I'm doing or, feedback in terms of listening to, things I've written to give me some feedback on that. You know, I do that all the time.

16:12Ray Yeah, , and I do exactly the same.

16:13 So I record all of the times that I'm speaking, and then I'll have a session with the AI afterwards looking at the transcript saying, what could I have been clearer about? What did I miss? Where did I get confused? Is any of this making sense? Because , I value that feedback.

16:33Dan Yeah.

16:33 Yeah. So, , moving on to some research then. So we've got quite a lot to go through today, so it's a bit of an interesting one 'cause we had less news and more research for this

16:42Ray Let's do the usual story.

16:43 You read out the titles. I'll tell you what the research is about.

16:46Dan absolutely.

16:46 So the first one is the "Gen AI Motivation and Learning Buddy, a generative AI prompt library for educators, and their students".

16:56Ray Okay, so this is from Andrew Martin, who's an academic researcher

17:00 at the UNSW School of Education. And, he pinged me to let me know that they'd done some stuff. It's like, oh, this is great. So basically it's a collection of 16 AI prompts that are designed to help students stay motivated, overcome procrastination, and help them manage, study, and examine anxiety. , So it's about helping students become better independent learners.

17:22 And so it's about them using AI as a personal learning coach. And so I love the fact that he's just released this as 16 prompts, that as a teacher you can use or you can give these prompts to your students. , It's super practical and I think it's great because it isn't about students' results, it's about how you support students and give them some coaching.

17:49 , I could do with this myself, like overcoming procrastination. Yes, please.

17:54Dan Yeah, yeah, definitely.

17:55 So the next one, "Can large language models reproduce higher education grade bands across models, study of collaborate or calibration, and grading bias in authentic student writing?" Whew.

18:11Ray Okay, so University of Birmingham in the uk researchers there, , published

18:15 this in assessment and evaluation in higher education, one of the journals. And what they did was that they tested where large language models matched the grades that humans give to real university essays. So they had 114 essays that had already been graded. They only picked the top end of the grading merit or distinction, which I think is an interesting choice.

18:40 So they were only testing the top end. They weren't testing everything, and they tested for open source models to see if the grade that AI gave match the grade that the human gave. And the answer is. About half of the time it matched the human grade.

18:59Dan Okay.

19:00Ray and what they found is some of the AI tended to be a bit harsh in

19:04 their marking compared to humans. But I have a couple of concerns about this research, , because I think they've jumped to a conclusion. We can't necessarily draw. So we don't know if a human marking that paper would agree with another human. So I mark a, a paper I give it to you. Are you gonna mark it in the same way?

19:27Dan Yeah, absolutely not.

19:28 Yeah. I've seen that absolutely happen when I've moderated exam papers in the

19:32Ray Yeah.

19:32 , Dan: That's like absolutely right. That's why you need to go through a whole process of it. Yeah.

19:36Ray so we know that AI agrees with the human marker about half of the time.

19:42 What we don't know is if another human marker marked it, would they agree more

19:47Dan absolutely.

19:47 Yeah,

19:48Ray So I think we have to know that in order to be able to understand

19:51 the value of what we're getting. the other thing is they only tested for open source models. They're all quite old models. They're definitely not state of the art. So to draw a conclusion that say that AI is not ready , for grading and marking is a bit of a long bow based off this, like, with those models, it didn't match

20:13Dan Yes.

20:13Ray human's marking.

20:14 , It's kind of like saying, , look, we've tested vehicles on the motor line. They can't get anywhere near the speed limit, but they only tested bicycles. You know, it's kind of four old open source models, rather than current state-of-the-art models. If you tested a car, you'd find it could achieve the speed limit.

20:33 So they've not really done that. So, , yeah, it's, it's interesting and it's an interesting exercise, but I would love it if they had been able to do that extra bit to say, do humans agree with humans marking? , And there's not much research out about that 'cause I went looking for it.

20:49Dan Yeah, no, absolutely.

20:50 And, and I suppose this is gonna be, it does highlight a point for researchers generally that, to test from, , frontier models. A lot of those are gated, a lot of those accosted. , And maybe there's a learning year for some of the Frontier labs to make sure the researchers have got access to the latest models when they're doing their research, rather than having to go and buy, you know, the latest version of Claude or GPT

21:16Ray and Anthropic have released a programme with I think 10,000 researchers

21:20 able to get access to Anthropic. That would be the right thing to do, is to bid into those programmes, get part of those programmes, and then you get the latest tools. , I'm sure OpenAI have got similar programme.

21:30Dan Yes.

21:31 , So next one, "generative AI and learning dynamics in higher education. A longitudinal, empirical study".

21:38Ray Get why I ask you to read the titles, Dan.

21:40 There's so many complicated words. , So this is in Germany, and what they did was , they had a, a, a chat style AI tool, and they put it into three different modes. The first one was, it was a tutor. The second was, it was just a chat, , no guidance at all. And the third was students learning without AI at all.

22:00 And what they did was, over the course of a term, , and 87 students split into these three random groups. They were looking to see whether it makes a difference. , The answer was the students without AI pretty much performed about the same as the students with ai.

22:18Dan Right.

22:19 Okay.

22:19Ray but what they found was the students that used AI were more likely to spend di

22:25 time digging into the answers that it got. So I get that AI to gimme an answer, and then I go, is that real? I'm just gonna spend a bit of time validating that answer. So , their view is, maybe the educational value is it encourages you to be more thoughtful about the learning process.

22:44Dan Such a

22:44 good point, isn't it? 'cause I'm reflecting on that myself. And I think that's how I use it. If I'm asking certain things, I'll read what comes back. I'll learn a little bit more. Maybe reframe it, delve deeper. So the outcome isn't actually the process that we are actually trying to judge here. The scores are not the, the process is actually the thinking.

23:06 Hmm. Interesting. So the next bit of research, AI assistance, here we go. "AI assistance reduces persistence and hurts independent performance".

23:19Ray this is really important 'cause the headline I saw about this in one

23:23 of the media was, does using AI for 10 minutes make you lazy and dumb? That's the headline. , Which you can't get out of the research paper. So basically what they did was they gave you a short exercise that you did with ai and then they took AI away. So it was a math problem or a reading problem. Then they took AI away and they found that people did worse when the AI was taken away.

23:50 Okay.

23:51Dan The title of this paper though, is almost presumptuous.

23:54 It's, you know, it says This hurts independent performance. It's not really a question, is it? It's more like a result. Anyway, sorry.

24:02Ray Yeah, I think you're right.

24:03 So, , so imagine I give you 10 maths problems to solve and you are getting some help from the AI to solve those problems. Then I say to you, "oh, can you solve another five? But you don't have AI to help you. Doesn't matter if you don't solve them or not. Oh, and you can skip them". And what they found was that people that had used AI for the first 10 were more likely to skip questions in the second batch than not.

24:33 So it's a bit like if I asked you to travel a mile down and I asked you if I asked you to walk a mile and then said, can you walk another half mile for me? You are more likely to do it than if I get you to take the bus for the first mile and then I ask you to walk a half mile. So that was really what they did was they made it easy for some, for some people in the first half by giving them AI to help them.

24:55 And then when they took the AI away, people were less inclined to go on and do the second stuff. So it's is useful to know, but it doesn't make you lazy and dumb with 10 minutes of AI use. It's more about the way that you get used to some help, and when you take it away, it feels like harder work.

25:15Dan Yes.

25:17 Yeah. Yeah, that's, very nuanced. That, yeah. Really, really interesting. So the final piece of research we've got here is called "Guardrails or roadblocks, effects of pedagogical style and context awareness in ai, teaching assistance for programming".

25:34Ray Right.

25:36 So this is another research paper that is specifically in the field of computer science.

25:42Dan Yes.

25:42Ray seeing a lot of that, a lot of the experiments are being done in programming

25:46 courses or in technology courses. And , what they kind of said is, do does having an AI tutor rather than a chat bot that gives you the answer, if we gave you a tutor that knows not to give you the answer, does it make it more frustrating? And so the answer was yes. And I've seen that across three or four papers recently.

26:11 What is happening is if you've got a Socratic tutor, it can drive you out the wall. 'cause he just wanted to give you the answer sometimes. And instead of doing that, it takes you around this complicated process.

26:23Dan A lot of, lots of them are sort of becoming more Socratic now.

26:27 Even the consumer tools , I've recently, I've done a couple of image generation for slides and things and you know, you ask generally, you know, gimme an image for this. And they go, "well, what exactly do you want that type of image to be?" And I'm go, just do the image for me. I'm like, please just generate the image.

26:41 Stop asking me more questions about it.

26:43Ray And so students that were using the Socratic tutor, there

26:48 were two interesting findings. The first was they were more frustrated by it, so they gave a satisfaction score that was much lower than using a non Socratic tutor that was helping you with the answers. The second was they actually scored worse in a follow-up exercises. The students that used a Socratic tutor scored worse than the students that had used one that was able to give them direct answers.

27:13Dan Okay.

27:14Ray that's interesting that, that's counterintuitive to me.

27:17 and I, and it could well come out from the frustration thing, you know, I'm just frustrated with this thing. Just help me get there and do it. Give me an answer. And so in the paper they talk about, the students that were demonstrating understanding of the topic later on. So when they went into a follow-up exercise and students that had the Socratic tutor had less understanding that the students that had a direct tutor that just helped them get to the answer, that to me seems counterintuitive to what we've been hearing all along.

27:49 I think it's gonna be really interesting to look into that. So it's a great paper to go and have a read. , The other thing that's coming out from the other research I'm seeing is students are frustrated with Socratic tutors. They're just gonna, you're gonna build them a Socratic tutor and then nobody's gonna use it.

28:04 'cause they'll just go to chat GPT on their phone or something. So with the best will in the world, you've gotta try and find a way to keep them engaged.

28:13Dan Yes.

28:14 And maybe the answer to that is to create the, just have the guardrails in to be able to manage the security and the privacy aspects of it, but actually leave the AI bots do their AI bot thing.

28:31Ray Yeah.

28:32 Or maybe it's a, don't always be a Socratic tutor when you've been asked twice. Don't send them around the houses Again, start to give more direct support rather than just asking questions. 'cause that's what a Socratic tutor is, isn't it? It's the kind of question based acquisition and knowledge, you know.

28:50 How many sides to a triangle? How many sides do you think there are to a triangle? How can we work that out?

28:54Dan Yeah.

28:54 And, , and often as teachers, there was a lot of, , evidence when we were doing some work in the uk, , about when teachers, teachers would often answer their own question. So you'd ask, you have a group of cla a group of kids, they say, you say how many sides has a triangle? Got? And before anybody could even ask you as a teacher would go, well, three, , you answer the question all the time.

29:14 You know, see, you don't ask as many questions, and leave enough space for people to answer. Or even if you do ask questions, only one person will answer anyway. So maybe that entire Socratic thing is leading a stoner garden path. , Some way or shape or other, but maybe some of the academics in the, teaching profession can let us know , what they think about Socratic nature of teaching.

29:38Ray Yeah, and we do get emails from people out of these conversations.

29:40 I got an email last week about the discussions we were having about research and our criticism of the fact that often it's not random control tiles in education. , So yeah, if you're shouting at the wireless, then drop us an email and, we do read it.

29:55Dan definitely.

29:56 Well, thanks again, Ray. That was fascinating. Weeks worth of roundup of news and research and, uh, see you next

30:01Ray Cheers.

30:02 Bye.