(November 2023) Unlocking the World of AI: Session 2 of AI in Education Series

0:00
[Rick Voithofer] … to have doctor Ana-Paula Correia,

0:03
who is a professor in the Learning Technologies Program.

0:06
It’s the Department of Educational Studies,

0:09
and she’s also director of the Center on Education and Training for Employment.

0:14
We also have Sean Hickey, who’s a PhD

0:17
candidate in the Learning Technologies Program and he is the Assessment

0:22
Assessment Deputy Director for CETE.

0:26
They both have created a guide that they will be sharing with you as part of their

0:31
presentation that you can refer to later.

0:34
We’re so happy to have you.

0:36
After they they’re done presenting, we’re going to have some breakout rooms

0:41
and we’ll have more on that later.

0:42
So I’ll I’ll pass it over to to Sean and to Professor Correia.

0:48
Thanks.

0:49
[Ana-Paula Correia] Good morning everyone.

0:51
Happy to be here with you today.

0:53
I want to thank Detra, Rick, Casey and the Centre for Digital Learning

0:59
and Innovation for this invitation.

1:02
Both Sean and I are from the Centre on Education Training for Employment.

1:08
We go by CETE. At our center

1:11
we have been building capacity in using generative AI,

1:15
you know step by step small steps.

1:18
And we’ve learned that starting with Chat GPT to discuss AI in education will help

1:24
us understanding how to interact with a large language model.

1:30
So this is why we select Chat GPT to discuss briefly with you today.

1:36
And I’m going to ask Sean to get us started by explaining what Chat GPT is.

1:43
[Sean Hickey] All right.

1:44
Thank you, Ana-Paula

1:45
and thank you for having us.

1:47
We’re like like Doctor Correia said, we’re just very excited to talk about

1:51
this.

1:51
As we explore these technologies, we want to share everything we’ve learned.

1:55
So I guess to start with, you know what is ChatGPT,

1:58
I’m actually going to back up just a little bit and and review what is

2:03
artificial intelligence a little more broadly.

2:07
And that’s because a lot of people think that it’s very new and it really isn’t.

2:12
AI tools have been around for a long time.

2:16
It was in the 1980s when Microsoft started adding Autocorrect to Microsoft

2:21
Word, which is a very early form of artificial

2:25
intelligence because it is predicting what you intended to type versus what you

2:31
actually typed.

2:32
We have that technology now in our phones.

2:35
When you’re typing a text message, there’s predictive text.

2:38
And of course, for the past decade or so, we’ve had digital assistants like Siri

2:43
and Alexa and Google that have been answering our questions.

2:47
And these are all examples of artificial intelligence.

2:50
And the keyword there is not intelligence, it’s artificial.

2:54
There’s no actual person there, no consciousness,

2:58
no understanding or knowledge.

3:00
It’s just programming, and in fact, the head of programming at Open AI

3:06
described the GPT model that runs ChatGPT as a bag of mostly math.

3:13
It’s a it’s a large number of statistical models and algorithms that basically

3:19
interpret what you’re asking and generate a response,

3:23
again based on the statistical analysis of the the order in which the words are

3:29
presented.

3:31
But that’s kind of what makes it a little bit different from some of the AI tools

3:34
that we’ve had before.

3:36
So ChatGPT is based on what’s called a large language model,

3:41
and that refers to the fact that it has not just been programmed with a number of

3:49
questions and canned responses, but instead has been trained on a corpus

3:56
of data that include web pages, encyclopedia articles, news articles,

4:02
journals, books, movie scripts, TV shows.

4:06
Basically any printed resource you can imagine on the Internet has been fed into

4:13
this model and as a result it has a capability that is referred to as Natural

4:20
Language Processing, which is Chat GPT’s chat ability that it

4:25
can interpret what you say

4:28
if you just ask a plain English question and give you a response,

4:32
that sounds like a response a person might give.

4:36
And that is I think one of the the key differences between things that came

4:40
before it.

4:42
If you think about like Siri or Alexa, somebody had to program all of the

4:47
possible questions that you could ask.

4:49
And so one of the things people use Siri the most for is asking about the weather.

4:54
And so some programmer or group of coders had to come up with a list of 30 or 40

4:59
different ways you might ask what the weather is like today.

5:04
Chat GPT and the similar AI tools based on large language models are different in

5:10
that they in they do have the power to interpret what you’re saying and generate

5:16
a response.

5:18
They’re also a little bit different in the fact that they they can learn.

5:22
They not only are they trained with a huge corpus of data, like I said,

5:27
but they are trained in tiny ways.

5:30
Every time you ask it a question, every time it generates a response,

5:35
the way it has been used in, you know, search engines or at the Chat GPT’s

5:41
website, it’ll often ask you like, is this the response you’re looking for?

5:47
And those sorts of things also help train it.

5:49
And so it is constantly learning.

5:52
And the latest version of ChatGPT based on the GPT 4 model which came out earlier

5:59
this year, is actually capable of connecting to the

6:04
Internet.

6:04
And so it is up to date, potentially as recently as you know 10

6:09
minutes ago and able able to respond to information that was just posted on the

6:15
Internet.

6:15
So it’s always growing and always learning,

6:17
which makes it a little bit different from some of the tools that came before.

6:23
If you haven’t used ChatGPT yet, I definitely encourage you to do so just

6:28
to try it out.

6:30
There are also other tools.

6:32
Google has one called Bard.

6:36
Microsoft has what it’s referring to as Copilot,

6:39
but they’re all very similar tools.

6:42
Some of them are actually based on the same underlying technology,

6:47
and if you go to Chat GPT’s chat.openai. com is the website,

6:51
you just create an account.

6:54
There is a paid version, but if you just want to try it out,

6:58
you can create an account for free and then you just type in complete sentences

7:03
and it will respond like a person responding to a chat.

7:07
And that’s kind of how it got its name.

7:10
And so I want to give a quick demonstration of what looks like.

7:14
Suppose you are working on a project and you want to invite a colleague to

7:21
collaborate.

7:22
So I will kind of show you what that looks like.

7:27
One second to change windows here and oh, there it is.

7:37
All right, so this is my ChatGPT window and I have

7:42
my request down here.

7:44
I’m asking to send an e-mail to a colleague or I should say compose.

7:47
That’s a little bit better, compose an e-mail to a colleague asking

7:52
them to collaborate on a research project.

7:54
So just a basic idea of one of the things you could use ChatGPT for and it’s gonna

8:01
think for a minute, Oh, there it goes.

8:04
I was gonna panic if it didn’t start and you’ll see as it’s writing you can sort

8:10
of see it live as it’s generating the text and you’ll see some things in

8:15
brackets.

8:16
And I like to think of this sort of response as a template.

8:19
So you could copy and paste this into an your e-mail program into Outlook and then

8:26
just look for those things that are in brackets and replace them and put in like

8:33
as you can see in the closing of the e-mail, your name your or your position,

8:39
scrolling up collaboration points, the name of the project,

8:44
the colleagues area of expertise, etc.

8:48
And that this is a good start.

8:51
And I I have found personally with my work that sometimes staring at a blank

8:55
page is the hardest part of doing any sort of creative work.

8:59
And it’s a lot of people will probably agree that it’s easier to respond to

9:03
something.

9:03
So even if I change every single word, this gives me a really good opening thing

9:08
to kind of, you know, bounce ideas off of brainstorm what have

9:12
you.

9:13
But you know suppose switch back here, you wanted to get a little more specific,

9:22
It is possible to really tailor the request that you give Chat GPT to to get

9:31
the a better response.

9:34
And so it it’s all about adding details to that prompt until some of the details

9:40
you might add.

9:41
You know what is the project?

9:43
Who is the person you’re asking to join you? What is their level of expertise?

9:49
What expertise do you add?

9:51
Any details you can you can add, and you’re gonna get a more tailored

9:55
response.

9:56
And so I’ll kinda show you what that looks like here as I find my windows

10:02
again.

10:04
So this is the original e-mail but I’m just going to and I’ve I already typed it

10:08
so I’m just pasting it in here.

10:10
But so compose an e-mail of no more than 200 words to my colleague Dr.

10:15
Ana-Paula Correia who has expertise in adult learning and workplace training

10:19
asking her to collaborate on a project in which we’ll be studying the efficacy of

10:23
different on the job training models.

10:26
This is a pilot project to put the dates that it’s going to run and then you know

10:30
if it goes well we might seek some federal funding.

10:32
So this this is the information now I’m giving ChatGPT and so it’s going to

10:38
create a better response that is closer to what I’m looking for.

10:43
So you can see it already uses Doctor Correia in the introduction,

10:49
it references the dates of the study, it mentions the goal, and it’s it’s brief.

10:55
This is actually really pretty close to what I would send if I were composing

11:00
this sort of e-mail.

11:01
So the more information you can give ChatGPT,

11:05
the better type of response you’re going to get.

11:10
And we refer to that specific process as prompt engineering.

11:15
And so I just want to very briefly discuss the steps by which you can really

11:21
edit and and and engineer the prompt that you’re giving ChatGPT or these other AI

11:27
tools in order to get the best response back.

11:32
And the first is context.

11:34
You know, explain what it is you want.

11:36
I don’t just want text, I want text appropriate for an e-mail.

11:40
I am asking a colleague for support using clarity,

11:45
avoiding vague terms or broad questions.

11:50
I could just say compose an e-mail, but that’s not going to give it enough

11:53
information, right?

11:53
You want to give as much detail as possible.

11:56
When appropriate, you can include examples.

11:59
At CETE, we have used ChatGPT to create some test

12:02
multiple choice questions and so we give it an example question and then it clones

12:08
that question and creates additional test items based on what it sees in the test

12:14
question.

12:16
Relevant information for the e-mail example that I just used.

12:20
When does the project start?

12:21
What’s it about?

12:22
What’s the expertise?

12:23
Etcetera.

12:25
One important thing is to consider your language.

12:28
While ChatGPT has a really broad base of knowledge,

12:32
most of that knowledge is written for the the laypeople out there.

12:36
So it’s news articles, it’s books, it’s encyclopedia entries.

12:41
So the more highly technical information you put into the prompt,

12:45
the more likely ChatGPT is gonna stumble a little bit and not provide you exactly

12:50
what you want.

12:51
So if you can use sort of common vernacular, that’s better.

12:54
And then the response parameters.

12:56
So you want the terms, How many words do you want?

13:01
Do you want it to be a bulleted list?

13:02
Do you want it to be a paragraph?

13:04
You can include all of these instructions as you engineer your prompt,

13:08
and the response you get is going to be just that much more useful.

13:14
All right.

13:14
And with that, I will turn it over to Doctor Correia,

13:17
who’s going to talk a little bit about the scholarly and academic side of the

13:22
work you can do with ChatGPT.

13:24
[Ana-Paula Correia] Thank you,

13:24
Sean.

13:25
Can you go back for a slide?

13:26
I just want to make a note because we’ve been working and discussing and studying

13:32
about his prompt engineering during engineering for a few months and it

13:37
really works.

13:38
And the engineering of the prompt is where our intelligence goes in because

13:44
then the ChatGPT just does uses the coding to generate an output.

13:50
And I will just want to just share one thing.

13:53
We need to be careful when we craft our prompt to not include names.

14:01
Locations of real like names.

14:04
I never use CETE on my prompts.

14:07
I have an acronym Chat GPT code for CETE because we I’m I’m very concerned that

14:14
all these information we’re putting that is real and is identifiable goes and

14:19
feeds the beast which is the large that data set.

14:23
So just a little bit of warning we’re going to talk about we’re going to talk

14:29
about drawbacks and downsides of AI but since Sean was talking about the prompts

14:35
and we got the demonstration.

14:37
Thank you so much for that.

14:39
Is really good.

14:40
When we see the model actually working.

14:43
I just want to put this idea on everyone’s mind that please be aware that

14:48
whatever you put on the prompt, it’s is going somewhere that you might

14:54
not want it to go.

14:55
So use acronyms or use doctor X or I mean maybe not XY.

15:01
So just be careful with any names or real people,

15:06
real locations and the real names of centers when you engineering your prompt.

15:14
Next please, Sean.

15:16
Thank you.

15:17
So just a little bit about how we can use ChatGPT to assist us in our scholarly

15:24
work.

15:25
It can be used in different ways.

15:27
I’m just going to highlight three.

15:29
One is to help us analyse and visualize data by generating natural language

15:36
descriptions of graphs and charts.

15:40
The graphs and charts properly explain it and described in language in narratives

15:47
can also help communicate research findings.

15:51
And these are always going to be a first draft that we need to look carefully,

15:57
look at carefully and critically.

16:00
But as Sean said, he’s better having a starting point that

16:03
don’t have anything and we always struggle with that.

16:06
Probably I do so regarding scholarly writing or scholarly work ChatGPT can be

16:13
used to assist us with research design, for instance,

16:18
generating drafts of research questions, hypothesis and study design based on the

16:25
prompt

16:26
engineering by us because we have control on how we want to steer the the tool to

16:34
generate an output and again those are initial drafts and initial things we can

16:41
refine later.

16:43
But it’s a good, it might be a good starting point and

16:47
it’s definitely a starting point that’s going to push us to move forward.

16:52
Also a Chat GPT can assist us in refining our academic writing.

16:58
So Sean gave the example of the e-mail writing and but we can also use Chat GPT

17:04
for our specific writing on papers, research briefs, white papers, Sean,

17:11
next please.

17:13
And again, it’s just examples, you probably already all know about these,

17:17
but it’s just like maybe a review.

17:19
So we can use Chat GPT to refine our academic writing by making our writing

17:26
more clear and concise.

17:29
For instance, we can draft a paragraph or section of

17:33
our paper and then put in the prompt something like that,

17:39
“Rewrite the following paragraph for clarity and succinctness for an academic

17:46
audience, ” and then we paste our paragraph or

17:50
section.

17:51
Be careful to remove all the identifiers and now we can see the output is going to

17:58
be maybe in a style that is more academic or more design more tailored to the

18:05
audience that we are catering to, which is an academic audience.

18:10
We can agree, disagree, like it or not like,

18:14
but again we have another round of another iteration on rewriting our

18:19
paragraph or section of our paper.

18:22
We also can use ChatGPT as our personal assistant editor and use it as a

18:28
proofreading tool.

18:30
I also use ChatGPT as a translator tool.

18:37
My my first language is Portuguese and sometimes I’m I’m invited to write for

18:44
journals in Brazilian journals.

18:47
So the Brazilian Portuguese is a little bit different from the Portuguese from

18:52
Portugal.

18:54
So I used Chat GPT to adjust my translation into Portugal into Portuguese.

19:01
So it’s more Brazilian Portuguese, so it’s it’s better for the audience that

19:07
I’m targeting.

19:09
Well, you you can also take English and

19:11
translate in Portuguese directly.

19:13
It’s a better translator than a Google translator in my opinion regarding

19:18
English to Portuguese and vice versa.

19:22
We can also use ChatGPT to help diversify our vocabulary and also refine our

19:28
writing style.

19:30
But again, always being critical about what ChatGPT

19:33
is going to put out as an answer as an output because it might not match what

19:39
the style we are looking for in regards to the audience we are writing for.

19:45
But there’s more examples, just something for you to think about.

19:50
And of course there is downsides on using AI.

19:55
Can you move please, Sean?

19:57
One of them is never trust the, the references and the quotes and the

20:07
citations offered by Chat GPT.

20:11
Probably Chat GPT 4 is it will be better, but I’m still working on the free version

20:18
of Chat GPT 3, so I know the references.

20:22
They sound really believable, but they don’t exist.

20:27
At least you know the model is evolving so fast that is much better now than

20:34
maybe a year, tot a year, 3 weeks ago that I used to assist with my

20:40
scholar rewriting.

20:42
All the references that were given by Chat GPT are fake, they don’t exist,

20:48
but they sound very believable and they are even an APA style.

20:54
So please, one of downsides of using AI is the

20:58
limited knowledge.

21:01
Again, this model is still evolving, the tool is still getting better and

21:06
there’s still limitations as as of today.

21:10
Maybe the limitations as of today are not the same as of three weeks ago,

21:14
but we know that it’s evolving.

21:17
Therefore the the information might might be fabricated,

21:22
it might be out of date because Chat GPT answers are not against reliable sources

21:29
of information.

21:31
There’s also a term that we use that is Chat GPT hallucinations because it’s just

21:38
making up stuff.

21:40
But the the tricky part is that the the make up stuff is really good and it’s

21:47
just like oh the probability of something like that be written somewhere is is high

21:54
because it it sounds believable.

21:58
That is the one the part that really concerns me.

22:03
Another one and you’re going to see this downside everywhere about generative AI

22:10
is the bias.

22:12
Chat GPT responses are created based on patterns and relationships,

22:16
learning from the training data.

22:18
Same for other large language model other than Chat GPT.

22:24
So that that those large training data sets most probably contain bias.

22:32
Therefore because of that you will help perpetuate stereotypes and reinforcing

22:39
social inequalities.

22:41
So the bias is also a a a problem that is is permeating many of the learning

22:47
technologies right now that we have not only generative AI, but it might be more

22:53
I liked it when we work with these systems.

22:58
Another one is lack of emotional intelligence and also absence of creative

23:04
thinking.

23:05
Those are unique characteristics of human kind that are still we cannot this point.

23:14
I don’t.

23:14
I cannot see the future.

23:16
But this point we cannot expect Chat GPT

23:19
to give us, to understand and respond to emotions,

23:24
or to provide a level of empathy and emotional support on its answers that a

23:30
human can.

23:32
So Chat GPT is not a companion, and it’s just a tool that we can use to

23:37
assist us on our everyday life as well as on our work life,

23:42
and in particular as researcher as in our work as scholars.

23:48
So Chat GPT does not understand this

23:52
the the specifics of the context where a particular problem emerge,

23:58
it does not understand the people’s lives.

24:01
So when it suggests a potential solution for a problem it’s probably not very

24:08
creative because it’s not grounded on the reality and we can go on, on downsides.

24:15
I think the balance is to really understand it, to take advantage of it,

24:20
but also be very critical on the output, on the things that tool can offer for us

24:25
because at the end of the day, it’s our work that’s going to be impacted.

24:30
So we want it to be the best possible work and we don’t want to be distracted

24:37
by corrections and biases and and fabricating information from generative

24:44
AI assistant.

24:45
That’s it.

24:46
So if you want to learn more about Chat GPT, we put together,

24:51
we used this reference guide at CETE for our build internal capacity and we make

24:57
it available to anyone.

24:59
So this was this quick reference ebook is available for free,

25:05
you can access cite us and it goes over many of the things we discussed with you

25:13
today here.

25:14
So thank you.

25:15
Thank you so much for listening and I hope you have a a fruitful breakout group

25:21
discussion.

25:23
Thank you.

25:26
[Rick Voithofer] Thank you, Doctor Correia and Sean.

25:29
Doctor Correia, I know that you’ve got to get to your

25:32
next meeting, so we appreciate your presentation.

25:35
Sean, are you able to hang out for a little bit

25:37
to answer some questions?

25:39
[Sean Hickey] Yeah, I have

25:40
I have a few minutes before my next meeting.

25:41
I can answer a couple questions.

25:43
[Rick Voithofer] OK, great.

25:44
So does anybody have, also, Sean, did you put the, the,

25:48
the guide in the chat?

25:50
Is that what I see there?

25:51
[Sean Hickey] I put the slides in and then there are links in the slides to

25:56
both the Apple bookstore and PDF version of the the guide.

26:00
So you can download it to your preferred platform.

26:04
[Rick Voithofer] Perfect.

26:04
Thanks.

26:06
So we’ll open the floor up if you would like to unmute your.

26:09
[Ana-Paula Correia- I’m gonna have to say bye.

26:11
Thank you for the opportunity.

26:13
Bye, bye.

26:13
[Rick Voithofer] Thank you.

26:14
Bye, bye.

26:14
Thank you.

26:14
[Julie Johnson] Hi, I have a quick question.

26:17
We have just a couple of minutes left.

26:20
Casey, would you maybe like to share a little

26:23
bit about the AI community?

26:25
I see you have a post there to to finish us off today.

26:31
[Casey Rinehart] Sure.

26:32
Yeah, so I mentioned we do have a EHE AI

26:35
community.

26:36
If you’re not already a member, I put a link in in there for how you can

26:40
join and and learn a little bit more.

26:43
We’re using that space to share resources as we get them.

26:47
You know, post information, you know, Greg,

26:49
Greg’s in this room now and is in the community,

26:52
the community and shared a a blog he’s finding interesting right now.

26:55
Those kinds of things and conversations are happening in the community.

26:59
So please feel free to join and and share and learn along with us.

27:05
And I’ll also mention that I’m collecting all the the resources that have been

27:08
shared today in the chat and we’ve got a few others that I think haven’t been

27:12
shared.

27:13
We’ll put all of that together as a follow up and send that out to everybody

27:16
who attended today so that we can have it in one,

27:19
we can have everything in one spot.

27:28
[Rick Voithofer] Thanks, Casey.

27:29
And finally we’ll be doing, I think our next one will be on teaching,

27:34
Is that right?

27:35
OK.

27:36
And that will be at the beginning of the year.

27:38
Is that right as well?

27:40
[Casey Rinehart] Yeah, we’re anticipating.

27:41
Yeah, at the beginning of the year.

27:43
OK.

27:44
And so we may be reaching out to see if there’s anything specific that you’d like

27:51
us to cover.

27:53
You know, we’ve gotten now we’re kind of a year

27:57
into this really a big social experiment.

28:01
And so we got to the point where people have been using it for quite a while.

28:05
And so you know perhaps if any of you, you know,

28:08
maybe would like to present some of the things you’re doing in that we, you know,

28:13
we’d like to include as much of our EHE community as possible.

28:18
So thank you all for attending and you know we’ll look forward to seeing you if

28:25
not sooner at the beginning of the year.