What is an LLM—and why do they matter for teachers and students?
In this AI Foundations video from Ed3, you’ll learn what an LLM (Large Language Model) is, how it works, and why it can feel surprisingly “human” even though it isn’t. If you’ve ever wondered how tools like chatbots generate fluent responses, lesson ideas, summaries, or feedback—this is the core concept behind it.
An LLM is a deep learning model trained on massive amounts of text (books, articles, websites, and public documents). It doesn’t “know” facts the way people do. Instead, it predicts the next most likely word one token at a time, based on learned patterns. That predictive power can look like thinking—especially because language contains so much knowledge and reasoning.
This video covers:
A common misconception is that LLMs are truth machines. They’re not. They’re pattern generators, and the quality of the output depends on the prompt, the context, and the user’s ability to verify and think critically.
This video is part of the AI Foundations series by Ed3, supporting educators worldwide in making informed, ethical, and human-centered decisions about AI in classrooms.
👉 Learn more about Ed3: https://www.ed3global.org
👉 Explore professional learning, courses, and events designed for educators navigating AI responsibly.
👉 Join our community of practice: https://community.ed3global.org
00:33 What an LLM is
00:47 “Large” datasets + parameters
01:01 “Language” tasks LLMs can do
01:12 “Model” = next-word prediction
01:42 Tokens explained
02:01 RLHF and why chatbots can feel personal
02:26 Why LLMs can seem like they’re thinking
02:29 Real-world uses (and why that matters in schools)
02:45 Shared risks with generative AI
04:22 LLM-specific concerns: bias, misinformation, hidden training data
04:44 The danger of AI companions and blurred roles
05:05 How educators can use LLMs well (and safely)
05:37 Seven tips for thoughtful classroom use
06:22 Helpers, not friends or authority figures
06:45 Separating hype from reality
1
00:00:04,866 --> 00:00:05,200
Yeah.
2
00:00:05,200 --> 00:00:06,500
You were like “rah rah rah rah”.
3
00:00:06,500 --> 00:00:08,266
Oh my gosh. Don't bring that up.
4
00:00:08,266 --> 00:00:08,866
I don't know
5
00:00:08,866 --> 00:00:09,766
do I stay blonde
6
00:00:09,766 --> 00:00:12,366
do I spice it up like go green, blue, pink?
7
00:00:12,366 --> 00:00:13,000
I don't know.
8
00:00:13,000 --> 00:00:14,400
Well, I think you look good
9
00:00:14,400 --> 00:00:15,666
no matter what,
10
00:00:15,666 --> 00:00:16,766
but a bright color
11
00:00:16,766 --> 00:00:19,366
will tend to clash with your outfits.
12
00:00:19,366 --> 00:00:22,133
Blonde goes with anything.
13
00:00:22,133 --> 00:00:23,966
See Chili dog has recently learned
14
00:00:23,966 --> 00:00:25,300
how to talk,
15
00:00:25,300 --> 00:00:25,733
and she's
16
00:00:25,733 --> 00:00:28,200
actually really good at giving advice,
17
00:00:28,200 --> 00:00:28,833
but might be
18
00:00:28,833 --> 00:00:29,633
because I trained her
19
00:00:29,633 --> 00:00:32,633
on a large language model.
20
00:00:33,000 --> 00:00:35,966
LLM stands for Large Language Model.
21
00:00:35,966 --> 00:00:37,466
It's a type of deep learning model
22
00:00:37,466 --> 00:00:39,566
trained on massive amounts of text.
23
00:00:39,566 --> 00:00:40,500
Everything from books
24
00:00:40,500 --> 00:00:41,666
and scientific articles
25
00:00:41,666 --> 00:00:43,466
to websites, social media posts
26
00:00:43,466 --> 00:00:45,033
and public documents.
27
00:00:45,033 --> 00:00:47,033
Let's break down what it means.
28
00:00:47,033 --> 00:00:47,966
“Large” means it's
29
00:00:47,966 --> 00:00:50,100
trained on huge datasets,
30
00:00:50,100 --> 00:00:51,833
billions of words from books,
31
00:00:51,833 --> 00:00:54,500
websites, articles and conversations.
32
00:00:54,500 --> 00:00:56,400
It also uses billions of parameters
33
00:00:56,400 --> 00:00:57,500
that are like knobs
34
00:00:57,500 --> 00:00:58,900
that it uses to adjust
35
00:00:58,900 --> 00:01:00,733
while learning patterns.
36
00:01:00,733 --> 00:01:01,500
“Language”
37
00:01:01,500 --> 00:01:03,866
means it works with human language.
38
00:01:03,866 --> 00:01:05,800
Reading, writing, translating,
39
00:01:05,800 --> 00:01:07,766
summarizing, answering questions,
40
00:01:07,766 --> 00:01:09,433
and even chatting.
41
00:01:09,433 --> 00:01:12,400
And “Model” means it's a predictive system.
42
00:01:12,400 --> 00:01:13,600
Given some text,
43
00:01:13,600 --> 00:01:14,166
it predicts
44
00:01:14,166 --> 00:01:16,666
what word or phrase should come next
45
00:01:16,666 --> 00:01:19,000
based on patterns it has learned.
46
00:01:19,000 --> 00:01:21,233
This scale is what makes LLMs
47
00:01:21,233 --> 00:01:22,700
so flexible.
48
00:01:22,700 --> 00:01:25,000
They've seen examples of formal writing,
49
00:01:25,000 --> 00:01:26,100
casual conversation,
50
00:01:26,100 --> 00:01:29,100
technical instructions, and even poetry.
51
00:01:29,166 --> 00:01:30,000
That variety
52
00:01:30,000 --> 00:01:32,033
allows them to mimic styles and adapt
53
00:01:32,033 --> 00:01:34,800
responses to different contexts.
54
00:01:34,800 --> 00:01:37,300
An LLM doesn't know facts.
55
00:01:37,300 --> 00:01:38,400
Instead, it predicts
56
00:01:38,400 --> 00:01:40,800
the next most likely word in a sequence,
57
00:01:40,800 --> 00:01:42,633
one token at a time.
58
00:01:42,633 --> 00:01:44,666
And tokens are simply chunks of text
59
00:01:44,666 --> 00:01:47,466
like words, syllables, or characters.
60
00:01:47,466 --> 00:01:48,633
During training,
61
00:01:48,633 --> 00:01:51,300
the model sees countless examples of text
62
00:01:51,300 --> 00:01:52,700
and adjusts its parameters
63
00:01:52,700 --> 00:01:55,100
to reduce mistakes in prediction.
64
00:01:55,100 --> 00:01:55,800
Over time,
65
00:01:55,800 --> 00:01:58,133
it gets very good at producing fluent,
66
00:01:58,133 --> 00:01:59,966
coherent text.
67
00:01:59,966 --> 00:02:01,500
Some LLMs are refined
68
00:02:01,500 --> 00:02:02,866
with additional methods
69
00:02:02,866 --> 00:02:03,900
like Reinforcement
70
00:02:03,900 --> 00:02:05,500
Learning with Human Feedback,
71
00:02:05,500 --> 00:02:07,666
which aligns outputs with human values
72
00:02:07,666 --> 00:02:09,366
or desired behaviors.
73
00:02:09,366 --> 00:02:10,633
This is how they can generate
74
00:02:10,633 --> 00:02:12,000
human-like conversations
75
00:02:12,000 --> 00:02:14,666
so they can seem personal, empathetic,
76
00:02:14,666 --> 00:02:16,333
or even friendly.
77
00:02:16,333 --> 00:02:17,633
Because language encodes
78
00:02:17,633 --> 00:02:18,600
so much knowledge
79
00:02:18,600 --> 00:02:19,466
from history,
80
00:02:19,466 --> 00:02:22,466
science, culture, and everyday reasoning,
81
00:02:22,533 --> 00:02:23,466
predicting words
82
00:02:23,466 --> 00:02:26,133
well can make it look like it's thinking
83
00:02:26,133 --> 00:02:26,866
Today,
84
00:02:26,866 --> 00:02:29,133
LLMs are used for everything
85
00:02:29,133 --> 00:02:30,233
wedding speeches,
86
00:02:30,233 --> 00:02:32,366
brainstorming, presentations,
87
00:02:32,366 --> 00:02:34,200
breakup texts, lesson writing,
88
00:02:34,200 --> 00:02:35,166
quiz banks, book
89
00:02:35,166 --> 00:02:36,133
authorship articles,
90
00:02:36,133 --> 00:02:36,933
song lyrics, trip
91
00:02:36,933 --> 00:02:39,933
itineraries, legal briefs, you name it.
92
00:02:40,533 --> 00:02:43,400
Since LLMs are used for generative AI,
93
00:02:43,400 --> 00:02:45,766
we have to watch out for the same issues.
94
00:02:45,766 --> 00:02:48,766
Let's cut to our primer on generative AI.
95
00:02:48,900 --> 00:02:50,700
Generative AI models are trained on
96
00:02:50,700 --> 00:02:52,233
massive datasets,
97
00:02:52,233 --> 00:02:53,633
much of which comes from books,
98
00:02:53,633 --> 00:02:55,300
articles, art, and media
99
00:02:55,300 --> 00:02:56,700
created by real people
100
00:02:56,700 --> 00:02:58,533
available on the internet.
101
00:02:58,533 --> 00:02:59,533
That means that the outputs
102
00:02:59,533 --> 00:03:00,300
might remix
103
00:03:00,300 --> 00:03:01,966
or resemble copyrighted works
104
00:03:01,966 --> 00:03:03,766
without proper attribution.
105
00:03:03,766 --> 00:03:05,633
We need to help students understand
106
00:03:05,633 --> 00:03:07,633
just because AI generates content
107
00:03:07,633 --> 00:03:09,000
doesn't mean they own it
108
00:03:09,000 --> 00:03:10,466
or can use it freely.
109
00:03:10,466 --> 00:03:11,833
This raises important lessons
110
00:03:11,833 --> 00:03:13,166
about digital citizenship
111
00:03:13,166 --> 00:03:14,733
and respecting creators.
112
00:03:15,800 --> 00:03:16,933
Students can use AI
113
00:03:16,933 --> 00:03:19,366
to generate essays, homework, or artwork
114
00:03:19,366 --> 00:03:21,033
that looks like their own work,
115
00:03:21,033 --> 00:03:22,200
but it isn't.
116
00:03:22,200 --> 00:03:23,066
This challenges
117
00:03:23,066 --> 00:03:24,633
how we measure learning.
118
00:03:24,633 --> 00:03:25,366
Since tools
119
00:03:25,366 --> 00:03:26,400
detecting generative
120
00:03:26,400 --> 00:03:28,666
AI plagiarism are unreliable,
121
00:03:28,666 --> 00:03:30,266
and because AI is everywhere
122
00:03:30,266 --> 00:03:31,866
in the workforce as well,
123
00:03:31,866 --> 00:03:34,866
the answer isn't banning AI tools.
124
00:03:34,900 --> 00:03:35,666
The first answer
125
00:03:35,666 --> 00:03:36,733
is having a conversation
126
00:03:36,733 --> 00:03:39,366
with students about how we can use AI.
127
00:03:39,366 --> 00:03:40,966
When is AI support allowed?
128
00:03:40,966 --> 00:03:43,566
Like brainstorming, drafting, outlining?
129
00:03:43,566 --> 00:03:45,466
And when is it crossing into plagiarism?
130
00:03:45,466 --> 00:03:46,733
Like copy pasting
131
00:03:46,733 --> 00:03:48,633
without engagement or learning.
132
00:03:48,633 --> 00:03:49,500
And secondly,
133
00:03:49,500 --> 00:03:51,366
we need to redesign our assignments
134
00:03:51,366 --> 00:03:52,366
to allow for thoughtful
135
00:03:52,366 --> 00:03:53,666
attribution of generative
136
00:03:53,666 --> 00:03:55,200
AI that allows students
137
00:03:55,200 --> 00:03:56,266
to demonstrate their own
138
00:03:56,266 --> 00:03:58,500
skills and understanding.
139
00:03:58,500 --> 00:04:00,166
Since generative AI creates
140
00:04:00,166 --> 00:04:02,266
content by remixing patterns,
141
00:04:02,266 --> 00:04:04,966
it doesn't produce truly original ideas.
142
00:04:04,966 --> 00:04:06,566
It simulates them.
143
00:04:06,566 --> 00:04:07,466
That means if students
144
00:04:07,466 --> 00:04:09,000
lean too heavily on AI.
145
00:04:09,000 --> 00:04:10,466
Their own voices, perspectives,
146
00:04:10,466 --> 00:04:12,700
and creativity can get lost.
147
00:04:12,700 --> 00:04:13,733
We play a key role
148
00:04:13,733 --> 00:04:15,233
in encouraging balance,
149
00:04:15,233 --> 00:04:17,733
using AI as a partner to spark ideas
150
00:04:17,733 --> 00:04:19,466
and not as a replacement for students
151
00:04:19,466 --> 00:04:21,066
authentic contributions.
152
00:04:22,200 --> 00:04:23,466
We also have to consider
153
00:04:23,466 --> 00:04:25,500
LLM specific issues.
154
00:04:25,500 --> 00:04:28,500
For example, the data isn't perfect.
155
00:04:28,600 --> 00:04:29,300
It can contain
156
00:04:29,300 --> 00:04:30,133
bias,
157
00:04:30,133 --> 00:04:33,133
outdated information, or misinformation.
158
00:04:33,166 --> 00:04:34,033
And because companies
159
00:04:34,033 --> 00:04:35,366
don't always disclose
160
00:04:35,366 --> 00:04:36,900
the exact data sets,
161
00:04:36,900 --> 00:04:38,200
we need to stay critical
162
00:04:38,200 --> 00:04:40,233
of what these models produce.
163
00:04:40,233 --> 00:04:41,400
Additionally,
164
00:04:41,400 --> 00:04:44,166
because an LLM can seem human,
165
00:04:44,166 --> 00:04:45,300
people of all ages
166
00:04:45,300 --> 00:04:46,800
can sometimes forget who
167
00:04:46,800 --> 00:04:48,666
or what they're talking to.
168
00:04:48,666 --> 00:04:50,966
Much like me talking to my dog,
169
00:04:50,966 --> 00:04:52,066
it can create a false sense
170
00:04:52,066 --> 00:04:53,100
of understanding,
171
00:04:53,100 --> 00:04:55,300
cause over-reliance on the tool,
172
00:04:55,300 --> 00:04:57,400
and blur human roles.
173
00:04:57,400 --> 00:04:58,266
The dangers of these
174
00:04:58,266 --> 00:04:59,233
AI companions
175
00:04:59,233 --> 00:05:01,600
are not to be underestimated.
176
00:05:01,600 --> 00:05:02,900
For educators,
177
00:05:02,900 --> 00:05:05,600
LLMs are the backbone of many text-based
178
00:05:05,600 --> 00:05:06,733
AI tools.
179
00:05:06,733 --> 00:05:08,233
They're a powerful thought partner
180
00:05:08,233 --> 00:05:09,000
and teammate
181
00:05:09,000 --> 00:05:11,666
for improving teaching and learning.
182
00:05:11,666 --> 00:05:13,066
They can differentiate content
183
00:05:13,066 --> 00:05:14,600
for diverse reading levels,
184
00:05:14,600 --> 00:05:16,100
provide multiple examples
185
00:05:16,100 --> 00:05:18,100
or analogies for the same concept,
186
00:05:18,100 --> 00:05:19,933
support multilingual classrooms
187
00:05:19,933 --> 00:05:21,200
through instant translation
188
00:05:21,200 --> 00:05:22,933
and so much more.
189
00:05:22,933 --> 00:05:24,333
On the flip side,
190
00:05:24,333 --> 00:05:25,433
the power of these tools
191
00:05:25,433 --> 00:05:27,733
becomes a concern for safety.
192
00:05:27,733 --> 00:05:29,266
LLMs can convince people,
193
00:05:29,266 --> 00:05:30,500
especially children,
194
00:05:30,500 --> 00:05:33,066
that they are real humans.
195
00:05:33,066 --> 00:05:33,900
Here are some tips
196
00:05:33,900 --> 00:05:34,633
that will encourage
197
00:05:34,633 --> 00:05:37,200
more thoughtful use of LLMs.
198
00:05:37,200 --> 00:05:38,100
One.
199
00:05:38,100 --> 00:05:40,066
Treat LLM outputs as drafts,
200
00:05:40,066 --> 00:05:41,733
not final products.
201
00:05:41,733 --> 00:05:42,433
Two.
202
00:05:42,433 --> 00:05:44,033
Always cross-check information
203
00:05:44,033 --> 00:05:46,133
against trusted sources.
204
00:05:46,133 --> 00:05:47,000
Three.
205
00:05:47,000 --> 00:05:48,166
Teach kids to ask
206
00:05:48,166 --> 00:05:49,833
better questions of the LLM
207
00:05:49,833 --> 00:05:51,933
and model critical reading.
208
00:05:51,933 --> 00:05:54,366
Four, design process-oriented
209
00:05:54,366 --> 00:05:56,366
assignments and require attribution
210
00:05:56,366 --> 00:05:57,733
of the LLM.
211
00:05:57,733 --> 00:05:58,566
Five.
212
00:05:58,566 --> 00:06:01,033
Discuss authorship and integrity.
213
00:06:01,033 --> 00:06:02,900
When is an LLM helpful
214
00:06:02,900 --> 00:06:04,600
and when is it damaging?
215
00:06:04,600 --> 00:06:05,533
Six.
216
00:06:05,533 --> 00:06:07,966
Discuss overreliance on AI.
217
00:06:07,966 --> 00:06:08,633
Because it can be
218
00:06:08,633 --> 00:06:10,200
a convincing thought partner,
219
00:06:10,200 --> 00:06:11,333
it's important to discuss
220
00:06:11,333 --> 00:06:12,100
habits of mind
221
00:06:12,100 --> 00:06:13,500
that help students retain
222
00:06:13,500 --> 00:06:15,166
their self-governance.
223
00:06:15,166 --> 00:06:16,300
And most importantly,
224
00:06:16,300 --> 00:06:17,233
Seven.
225
00:06:17,233 --> 00:06:18,433
Engage in constant
226
00:06:18,433 --> 00:06:19,466
dialog with students
227
00:06:19,466 --> 00:06:22,500
about how LLMs and AI are not human.
228
00:06:22,533 --> 00:06:24,600
LLMs can be positioned as helpers,
229
00:06:24,600 --> 00:06:26,400
but should not be seen as friends,
230
00:06:26,400 --> 00:06:28,866
peers, or authority figures.
231
00:06:28,866 --> 00:06:29,933
Large Language Models
232
00:06:29,933 --> 00:06:32,066
are powerful pattern generators,
233
00:06:32,066 --> 00:06:33,133
but their outputs reflect
234
00:06:33,133 --> 00:06:34,800
the limits of their data.
235
00:06:34,800 --> 00:06:37,300
The opportunity isn't just to use them.
236
00:06:37,300 --> 00:06:39,600
It's to model critical engagement,
237
00:06:39,600 --> 00:06:40,533
ensuring students
238
00:06:40,533 --> 00:06:42,400
see the utility of LLMs
239
00:06:42,400 --> 00:06:43,733
without over-relying on them
240
00:06:45,133 --> 00:06:46,266
As educators,
241
00:06:46,266 --> 00:06:47,466
knowing how LLMs
242
00:06:47,466 --> 00:06:48,466
work helps us separate
243
00:06:48,466 --> 00:06:50,566
the hype from the reality.
244
00:06:50,566 --> 00:06:51,633
So we can make wiser
245
00:06:51,633 --> 00:06:53,233
choices in our classrooms.
246
00:06:53,966 --> 00:06:54,666
Shhh.
