1 00:00:00,000 --> 00:00:08,000 Hi, I'm Ethan Malik, a professor at Wharton and the co-director of the Generative AI Lab 2 00:00:08,000 --> 00:00:09,000 here at Penn. 3 00:00:09,000 --> 00:00:14,680 Hi, I'm Lina Hmalik, and I'm co-director of the Generative AI Lab at Penn, and I have 4 00:00:14,680 --> 00:00:18,040 been working on interactive pedagogy for the Wharton School. 5 00:00:18,040 --> 00:00:21,719 So in this short class, we're going to be talking about how to use AI in education from 6 00:00:21,719 --> 00:00:22,959 the teacher perspective. 7 00:00:22,959 --> 00:00:26,940 And this covers all kinds of instruction, college level, but also K-12. 8 00:00:26,940 --> 00:00:32,380 And we're going to focus on open AI tools, specifically using chat GPT and various GPT 9 00:00:32,380 --> 00:00:33,740 models to solve problems. 10 00:00:33,740 --> 00:00:35,459 But there are other AI solutions out there. 11 00:00:35,459 --> 00:00:39,439 We'll be focusing on open AIs, though, in this particular course. 12 00:00:39,439 --> 00:00:43,180 It's important to realize that AI, even though it's a really advanced technology, is not 13 00:00:43,180 --> 00:00:44,740 inherently technical. 14 00:00:44,740 --> 00:00:46,580 And we don't expect you to have a technical knowledge. 15 00:00:46,580 --> 00:00:50,479 In fact, we don't think technology helps you very much outside of a very tiny technical 16 00:00:50,479 --> 00:00:53,020 discussion about how AI works. 17 00:00:53,020 --> 00:00:57,540 So we'll cover key concepts in generative AI, benefits and risks of using it in education, 18 00:00:57,540 --> 00:01:01,180 ethical concerns, and talk about some policies you should use in this class. 19 00:01:01,180 --> 00:01:04,779 Then in future classes, we'll be covering how to do prompting and how to actually deploy 20 00:01:04,779 --> 00:01:06,860 AI assignments. 21 00:01:06,860 --> 00:01:11,099 So for a little bit of context, AI is already impacting education. 22 00:01:11,099 --> 00:01:15,419 We know that a lot of you have already had some experience with AI or are actively using 23 00:01:15,419 --> 00:01:16,419 AI. 24 00:01:16,419 --> 00:01:22,300 In a recent Wharton Family Foundation survey, half of the teachers survey say they use tools 25 00:01:22,300 --> 00:01:24,940 like ChatGPT at least once a week. 26 00:01:24,940 --> 00:01:30,180 And most of those educators who use ChatGPT have a favorable opinion of the tool, as do 27 00:01:30,180 --> 00:01:32,139 your students. 28 00:01:32,139 --> 00:01:37,059 Educators who use the tool say that it saves them hours of work a week. 29 00:01:37,059 --> 00:01:41,620 But although it's a powerful tool, there is just no instruction manual. 30 00:01:41,620 --> 00:01:45,980 And the way that you use it, the way that it's helpful for you, might not be helpful 31 00:01:45,980 --> 00:01:47,379 for someone else. 32 00:01:47,379 --> 00:01:51,819 And so what all these educators have in common, what you might have in common with them, is 33 00:01:51,819 --> 00:01:53,819 that there is a need for more guidance. 34 00:01:53,819 --> 00:01:58,379 In this course, we want to try to help give you some guidance, give you tools and use 35 00:01:58,379 --> 00:02:03,180 cases to show how these systems operate and what you can do to help you teach and help 36 00:02:03,180 --> 00:02:04,900 your students learn. 37 00:02:04,900 --> 00:02:08,460 So in order to talk about AI in education, we first need to talk about AI. 38 00:02:08,460 --> 00:02:13,179 And it's a very imprecise term that might mean anything from science fictional robots 39 00:02:13,179 --> 00:02:15,820 to the large language models we use today. 40 00:02:15,820 --> 00:02:20,119 For a long time, when people talked about AI, what they talked about was the idea that 41 00:02:20,119 --> 00:02:25,679 you could train machine learning systems to predict sequences of numbers. 42 00:02:25,679 --> 00:02:27,240 And the predictor is used for all sorts of things. 43 00:02:27,240 --> 00:02:32,199 So when you watch a Netflix movie, the algorithm is predicting what other movie you might want 44 00:02:32,199 --> 00:02:33,199 to watch. 45 00:02:33,199 --> 00:02:35,919 When you're scrolling through TikTok, the machine learning algorithm, based on all the 46 00:02:35,919 --> 00:02:39,839 things that you've liked before and other people have liked, is guessing what you might 47 00:02:39,839 --> 00:02:41,160 want to see next. 48 00:02:41,160 --> 00:02:44,000 And so these systems are really good at making those kinds of predictions. 49 00:02:44,000 --> 00:02:46,199 And AI has been booming for that reason. 50 00:02:46,199 --> 00:02:49,220 But these forms of AI were not really good at human language. 51 00:02:49,220 --> 00:02:50,580 Because human language isn't as predictable. 52 00:02:50,580 --> 00:02:54,419 If I end a sentence with the word filed, am I filing my taxes or filing my nails? 53 00:02:54,419 --> 00:02:55,740 The AI didn't know. 54 00:02:55,740 --> 00:03:00,619 So a new kind of AI was invented in the late 2010s called the large language model based 55 00:03:00,619 --> 00:03:06,220 on the transformer mechanism with attention that allowed the AI to pay attention to not 56 00:03:06,220 --> 00:03:10,100 just the last word in the sentence, but the entire sentence, the entire context around 57 00:03:10,100 --> 00:03:11,580 it, and so on. 58 00:03:11,580 --> 00:03:15,899 And these new approaches to AI we call large language models. 59 00:03:15,899 --> 00:03:18,059 And what they are are word prediction machines. 60 00:03:18,059 --> 00:03:21,339 So just like those other forms of AI were about predicting the next number in a sequence, 61 00:03:21,339 --> 00:03:24,139 what the AI does is predict the next word or token in a sequence. 62 00:03:24,139 --> 00:03:29,539 So if I start a sentence with the best pet for anyone to have is a, the AI would use 63 00:03:29,539 --> 00:03:33,699 all of the knowledge it's learned from training on all of human language to predict the next 64 00:03:33,699 --> 00:03:39,899 word in the sentence, which might be 60% chance that word's dog, 30% chance it's cat, 10% 65 00:03:39,899 --> 00:03:41,220 it's parrot. 66 00:03:41,220 --> 00:03:43,100 And it will most likely pick the most likely word. 67 00:03:43,100 --> 00:03:45,139 Sometimes it picks the second most likely or the third most likely. 68 00:03:45,139 --> 00:03:47,500 So there's a little bit of randomness built into this. 69 00:03:47,940 --> 00:03:50,940 But the AI is just very fancy autocomplete. 70 00:03:50,940 --> 00:03:54,779 That doesn't actually help us understand how to use it, but it's useful to know what these 71 00:03:54,779 --> 00:03:58,899 things are doing because it also helps tell us some of the downsides of AI. 72 00:03:58,899 --> 00:04:02,820 One of those is the fact that AIs hallucinate. 73 00:04:02,820 --> 00:04:04,940 I told you that AIs predict the next word in the sentence. 74 00:04:04,940 --> 00:04:07,419 They're not looking at a database of every sentence ever written. 75 00:04:07,419 --> 00:04:10,800 They've learned statistical patterns in language and they're reproducing it. 76 00:04:10,800 --> 00:04:15,580 That means that what the AI produces seems realistic and real, but it may not always 77 00:04:15,580 --> 00:04:16,579 be. 78 00:04:16,660 --> 00:04:21,260 So when what the AI tells you a sentence is, is different than what the reality is, we 79 00:04:21,260 --> 00:04:22,899 call that a hallucination. 80 00:04:22,899 --> 00:04:24,100 And all AIs do is hallucinate. 81 00:04:24,100 --> 00:04:25,779 They always are making up the next word in the sentence. 82 00:04:25,779 --> 00:04:28,579 The fact that they're accurate so often is actually somewhat surprising. 83 00:04:28,579 --> 00:04:29,579 They are accurate a lot. 84 00:04:29,579 --> 00:04:34,220 We find they outperform humans in almost every sort of standardized test for what that's 85 00:04:34,220 --> 00:04:35,220 worth. 86 00:04:35,220 --> 00:04:38,019 They're quite accurate, but they still make things up that are very plausible. 87 00:04:38,019 --> 00:04:39,019 And we call this hallucinations. 88 00:04:39,019 --> 00:04:40,500 It's a way of getting misled with AI. 89 00:04:40,500 --> 00:04:42,579 We'll talk a little bit more about that later in the course. 90 00:04:42,579 --> 00:04:47,399 But it's not the only risk that you face as an educator using these systems. 91 00:04:47,399 --> 00:04:52,540 So we know that students are using these systems and because they have the power and the capacity 92 00:04:52,540 --> 00:04:58,140 to write essays and help solve problems and talk to students through solutions and in 93 00:04:58,140 --> 00:05:03,820 fact give students the answers, these systems, it's important to remember, don't teach 94 00:05:03,820 --> 00:05:04,820 out of the box. 95 00:05:04,820 --> 00:05:06,380 They are not a teacher. 96 00:05:06,380 --> 00:05:07,920 They just provide information. 97 00:05:08,000 --> 00:05:13,239 And we know that just reading information is not, in fact, learning. 98 00:05:13,239 --> 00:05:18,700 And we have some early evidence, in fact, that students can use AI as a crutch when 99 00:05:18,700 --> 00:05:22,339 given access to the tool and feel like they're learning something. 100 00:05:22,339 --> 00:05:26,399 So they have the illusion of understanding, but when confronted with a test, they actually 101 00:05:26,399 --> 00:05:29,160 have not learned that topic or subject. 102 00:05:29,160 --> 00:05:31,600 And AI cheating is really hard to stop. 103 00:05:31,600 --> 00:05:37,160 You can't easily detect AI writing and you can't use AI detectors when you grade. 104 00:05:37,160 --> 00:05:41,299 Not only do AI detectors not work, they're really easy to defeat. 105 00:05:41,299 --> 00:05:46,739 And they also have a high false positive rate against non-native English speakers. 106 00:05:46,739 --> 00:05:48,679 So they really can't be used. 107 00:05:48,679 --> 00:05:52,440 Cheating is not the only concern you might want to think about when using AI systems. 108 00:05:52,440 --> 00:05:53,440 Another is bias. 109 00:05:53,440 --> 00:05:58,760 So AI systems have been trained on human writing and human writing contains lots of biases. 110 00:05:58,760 --> 00:06:02,160 The types of writing that people select to use to train the models might have biases 111 00:06:02,160 --> 00:06:03,320 themselves. 112 00:06:03,320 --> 00:06:05,559 And the AI picks up this information as it goes along. 113 00:06:05,559 --> 00:06:09,320 We know, for example, that there are subtle biases in AI work. 114 00:06:09,320 --> 00:06:12,839 So if you ask the AI to write a letter of recommendation for someone and the person 115 00:06:12,839 --> 00:06:17,799 that it's writing for is female, we know that some AI systems are likely to be biased and 116 00:06:17,799 --> 00:06:22,040 talk about that woman being warm as opposed to if it's asked to write a letter of recommendation 117 00:06:22,040 --> 00:06:25,720 for a man, in which case it's more likely to say they're competent, a fact we see in 118 00:06:25,720 --> 00:06:29,000 real life letters of recommendation too that the AI repeats. 119 00:06:29,000 --> 00:06:32,140 And so you want to be aware of these sorts of biases and risks. 120 00:06:32,140 --> 00:06:34,559 There are some areas where AI is more biased than others. 121 00:06:34,559 --> 00:06:38,160 In a lot of cases it's mitigated, but we're really just still learning about that. 122 00:06:38,160 --> 00:06:39,799 So it's something you want to keep in mind. 123 00:06:39,799 --> 00:06:42,160 I know people are also concerned about things like privacy issues. 124 00:06:42,160 --> 00:06:47,000 And while those are a valid concern and you need to look at the legal and regulatory situation 125 00:06:47,000 --> 00:06:51,519 that you're in as an instructor, there are now a lot of ways in which to ensure privacy 126 00:06:51,519 --> 00:06:54,279 in these systems that we won't go into detail here. 127 00:06:54,279 --> 00:06:59,559 But open AI offers a lot of options to create private chats and the other AI systems do 128 00:06:59,559 --> 00:07:00,559 the same. 129 00:07:00,559 --> 00:07:03,119 So that may not be as much of a concern. 130 00:07:03,119 --> 00:07:08,359 And another thing that we think that people should be aware of is issues of access and 131 00:07:08,359 --> 00:07:10,100 equity of access. 132 00:07:10,100 --> 00:07:14,920 So the nice thing right now when we're recording this is that open AI has made available to 133 00:07:14,920 --> 00:07:19,679 the world access to its most advanced models, which means that anyone with internet access 134 00:07:19,679 --> 00:07:23,760 can get a limited number of interactions with those models for free. 135 00:07:23,760 --> 00:07:27,839 So even if your students are not paying for chat GPT access, they should be able to access 136 00:07:27,839 --> 00:07:30,160 the same advanced models that everybody else does. 137 00:07:30,160 --> 00:07:32,600 Where that continues in the future, we can't tell you one way or another. 138 00:07:32,600 --> 00:07:36,200 It kind of depends on the time, but we hope to see that continued devotion towards free 139 00:07:36,200 --> 00:07:37,799 access to everybody. 140 00:07:37,799 --> 00:07:40,359 So what does this mean for you in your classroom? 141 00:07:40,359 --> 00:07:42,679 You can ban AI in your classroom. 142 00:07:42,679 --> 00:07:46,559 Now banning AI in your classroom does not mean that students won't use it outside of 143 00:07:46,559 --> 00:07:51,760 class, but it does mean that you'll be preventing students from using it inside of class. 144 00:07:51,760 --> 00:07:53,480 That is a very legitimate option. 145 00:07:53,480 --> 00:07:57,279 We know that students don't come to class as experts and that gaining expertise in any 146 00:07:57,279 --> 00:08:02,320 topic takes a lot of work, a lot of friction, a lot of practice. 147 00:08:02,320 --> 00:08:08,359 And in banning AI, we are really doubling down on what we know pedagogically works, 148 00:08:08,359 --> 00:08:14,799 including low stakes testing, flipped classrooms, blue book exams, separating assessments from 149 00:08:14,799 --> 00:08:15,799 homework. 150 00:08:15,799 --> 00:08:20,799 So we can certainly, that is one option that you can certainly take on. 151 00:08:20,799 --> 00:08:24,600 The second option is actually one we don't recommend anymore, which is taking a sort 152 00:08:24,600 --> 00:08:28,480 of laissez-faire integration of AI into your class situation. 153 00:08:28,480 --> 00:08:34,039 So originally when the first AI models came out that were widely used, like the CHAT GPT 154 00:08:34,039 --> 00:08:38,200 3.5 that came out a couple of years ago, the systems were pretty good, but not amazing. 155 00:08:38,200 --> 00:08:39,880 They made lots of mistakes. 156 00:08:39,880 --> 00:08:43,280 And so it was relatively easy to tell your students, hey, just use the AI, acknowledge 157 00:08:43,280 --> 00:08:47,159 that you're using it, and you're responsible for the errors in your assignment. 158 00:08:47,159 --> 00:08:52,159 The problem is that GPT 4.0 and successor models are so good that they make far fewer 159 00:08:52,159 --> 00:08:53,880 errors than your students do. 160 00:08:53,880 --> 00:08:57,440 So you can't just simply say use AI and it'll work out, because then you're setting up yourself 161 00:08:57,440 --> 00:09:02,400 a recipe for people to cheat either accidentally or on purpose, and you're in risk for learning 162 00:09:02,400 --> 00:09:03,400 loss. 163 00:09:03,400 --> 00:09:07,159 So I think we need to move away from a model of just casually using this and start to think 164 00:09:07,159 --> 00:09:11,919 more seriously about how to integrate AI into classrooms in a transformative way. 165 00:09:11,919 --> 00:09:17,059 So while it's certainly OK to, as we just discussed, ban usage and think about other 166 00:09:17,059 --> 00:09:22,179 approaches to working with AI in class, we're going to talk a bit for the rest of this session 167 00:09:22,179 --> 00:09:26,219 about transformation, about how you can use AI to create assignments, help you do work, 168 00:09:26,219 --> 00:09:30,820 and what the future of education might look like, and how you can help shape that future. 169 00:09:30,820 --> 00:09:34,340 We've been recording, as you probably guessed, this entire presentation so far. 170 00:09:34,340 --> 00:09:38,219 So I'm actually going to give what we recorded so far to the AI and get some feedback on 171 00:09:38,219 --> 00:09:40,580 what we might do better in our next video, which is coming up. 172 00:09:40,580 --> 00:09:52,539 So given this script of our class so far, what should we do better in the second video? 173 00:09:52,539 --> 00:09:53,539 OK. 174 00:09:53,539 --> 00:09:54,539 Great. 175 00:09:54,539 --> 00:09:55,539 OK. 176 00:09:55,539 --> 00:10:01,179 So it's pretty complimentary so far, but its three main suggestions is that we show more 177 00:10:01,179 --> 00:10:06,340 examples, which we will absolutely do in the next video, that we make sure to include links 178 00:10:06,340 --> 00:10:10,979 and additional resources, and those will be available to you on the learning platform. 179 00:10:10,979 --> 00:10:13,900 So that's excellent advice, I think, and we'll take that for our next video. 180 00:10:13,900 --> 00:10:19,099 In the next video, we'll be sharing how you can use AI to help you in your teaching and 181 00:10:19,099 --> 00:10:21,380 how to craft AI assignments for students.