What are AI hallucinations—and why do they matter in schools?
In this AI Foundations video from Ed3, we explain AI hallucinations: moments when an AI tool gives an answer that sounds correct, confident, and polished—but is actually false, misleading, or completely invented.
AI hallucinations happen because most AI systems (especially LLMs) don’t “know” facts. They generate responses by predicting the next most likely word based on patterns in training data. When the model lacks good grounding sources—or the data is incomplete, contradictory, or outdated—it may still fill in the blanks with something convincing… and wrong.
This video covers:
A common misconception is that hallucinations are rare “glitches.” In reality, they’re a predictable outcome of how many AI systems generate text—especially when users ask for exact facts without providing reliable source material.
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
00:14 A confident wrong answer
00:31 What an AI hallucination is
00:40 Why AI doesn’t “know” facts
00:58 When hallucinations are more likely
01:20 Why AI answers anyway
01:36 A real-world legal example
01:48 Why this matters for students and teachers
02:09 Verify before you trust
02:18 Three strategies to handle hallucinations
02:55 Turn it into media literacy instruction
1
00:00:14,700 --> 00:00:15,866
Candy,
2
00:00:15,866 --> 00:00:17,600
you know that's not right,
3
00:00:17,600 --> 00:00:18,833
yeah?
4
00:00:18,833 --> 00:00:19,433
Neil Armstrong
5
00:00:19,433 --> 00:00:21,100
was the first person on the moon,
6
00:00:21,100 --> 00:00:22,733
but Christa McAuliffe
7
00:00:22,733 --> 00:00:24,900
was the first teacher in space.
8
00:00:24,900 --> 00:00:26,633
There's a name for this
9
00:00:26,633 --> 00:00:28,200
kind of made up answer
10
00:00:28,200 --> 00:00:31,200
it’s called an AI hallucination.
11
00:00:31,300 --> 00:00:33,000
An AI hallucination
12
00:00:33,000 --> 00:00:34,966
is when the AI generates information
13
00:00:34,966 --> 00:00:36,466
that sounds correct
14
00:00:36,466 --> 00:00:38,633
but is actually false, misleading,
15
00:00:38,633 --> 00:00:40,633
or completely invented.
16
00:00:40,633 --> 00:00:42,533
AI doesn't know facts.
17
00:00:42,533 --> 00:00:43,066
It predicts
18
00:00:43,066 --> 00:00:44,600
the next most likely word
19
00:00:44,600 --> 00:00:47,200
based on patterns in its training data.
20
00:00:47,200 --> 00:00:47,866
If it doesn't have
21
00:00:47,866 --> 00:00:49,100
the right grounding sources,
22
00:00:49,100 --> 00:00:50,133
or if the training data
23
00:00:50,133 --> 00:00:51,666
has gaps, contradictions,
24
00:00:51,666 --> 00:00:53,533
or outdated information,
25
00:00:53,533 --> 00:00:55,333
it might still fill in the blank
26
00:00:55,333 --> 00:00:57,533
with something that sounds convincing,
27
00:00:57,533 --> 00:00:58,533
but it isn't.
28
00:00:58,533 --> 00:00:59,633
This is especially true
29
00:00:59,633 --> 00:01:01,566
for niche or new topics,
30
00:01:01,566 --> 00:01:03,033
long or complex prompts,
31
00:01:03,033 --> 00:01:05,533
and for precise facts.
32
00:01:05,533 --> 00:01:08,300
In this example, the data only consists
33
00:01:08,300 --> 00:01:11,300
of green, red, and orange things.
34
00:01:11,400 --> 00:01:13,733
So when I asked for blue things,
35
00:01:13,733 --> 00:01:15,033
it presented green, red
36
00:01:15,033 --> 00:01:17,766
and orange things, but in a blue tint.
37
00:01:17,766 --> 00:01:20,600
This is a hallucination.
38
00:01:20,600 --> 00:01:21,800
Now you're probably thinking,
39
00:01:21,800 --> 00:01:23,366
why doesn't it just tell you
40
00:01:23,366 --> 00:01:24,666
that it doesn't know?
41
00:01:24,666 --> 00:01:24,966
Well, it's
42
00:01:24,966 --> 00:01:26,633
because AI has been programed
43
00:01:26,633 --> 00:01:27,833
to give you an answer
44
00:01:27,833 --> 00:01:29,200
no matter what.
45
00:01:29,200 --> 00:01:31,266
Its algorithm would rather be over
46
00:01:31,266 --> 00:01:32,500
confidently incorrect
47
00:01:32,500 --> 00:01:34,733
with its very best educated guess,
48
00:01:34,733 --> 00:01:36,833
than completely silent.
49
00:01:36,833 --> 00:01:39,033
In the Mata v Avianca case,
50
00:01:39,033 --> 00:01:39,900
ChatGPT
51
00:01:39,900 --> 00:01:41,366
gave this lawyer several
52
00:01:41,366 --> 00:01:42,900
non-existent federal
53
00:01:42,900 --> 00:01:45,366
and state court decisions as precedent
54
00:01:45,366 --> 00:01:46,633
to support his argument.
55
00:01:48,766 --> 00:01:50,900
When students and teachers use AI,
56
00:01:50,900 --> 00:01:51,766
it can be dangerous
57
00:01:51,766 --> 00:01:53,933
to consider it a source of truth.
58
00:01:53,933 --> 00:01:54,866
For example,
59
00:01:54,866 --> 00:01:57,533
if a student asks AI for historical dates
60
00:01:57,533 --> 00:01:59,266
and it provides the wrong years,
61
00:01:59,266 --> 00:01:59,800
the students
62
00:01:59,800 --> 00:02:01,833
learning will be compromised.
63
00:02:01,833 --> 00:02:02,900
And for teachers,
64
00:02:02,900 --> 00:02:04,300
these types of hallucinations
65
00:02:04,300 --> 00:02:05,666
can challenge reliability
66
00:02:05,666 --> 00:02:06,533
of the resources
67
00:02:06,533 --> 00:02:07,733
they produce, from quizzes
68
00:02:07,733 --> 00:02:09,600
to lesson plans.
69
00:02:09,600 --> 00:02:11,700
Hallucinations can sneak into all kinds
70
00:02:11,700 --> 00:02:13,733
of resources and research.
71
00:02:13,733 --> 00:02:14,033
That's why
72
00:02:14,033 --> 00:02:15,666
we need to verify the information
73
00:02:15,666 --> 00:02:18,033
before we trust it.
74
00:02:18,033 --> 00:02:18,866
There are three things
75
00:02:18,866 --> 00:02:21,633
we can do to deal with hallucinations.
76
00:02:21,633 --> 00:02:22,500
One.
77
00:02:22,500 --> 00:02:24,600
Make our prompts better.
78
00:02:24,600 --> 00:02:25,666
Ground your query
79
00:02:25,666 --> 00:02:27,366
in sources by providing the text
80
00:02:27,366 --> 00:02:29,466
it needs to derive the answer.
81
00:02:29,466 --> 00:02:31,600
When AI produces a response,
82
00:02:31,600 --> 00:02:33,600
ask it to cite the source.
83
00:02:33,600 --> 00:02:35,233
You can also constrain the task
84
00:02:35,233 --> 00:02:36,300
by breaking the problem
85
00:02:36,300 --> 00:02:38,333
or query into steps.
86
00:02:38,333 --> 00:02:40,166
Ask it to list any uncertainties
87
00:02:40,166 --> 00:02:42,333
it has about the query.
88
00:02:42,333 --> 00:02:45,333
Two. Judge before you trust.
89
00:02:45,333 --> 00:02:45,966
Double check
90
00:02:45,966 --> 00:02:47,733
AI answers with a trusted source
91
00:02:47,733 --> 00:02:48,933
and use your own judgment
92
00:02:48,933 --> 00:02:49,966
to carefully read
93
00:02:49,966 --> 00:02:51,833
and understand the response.
94
00:02:51,833 --> 00:02:54,833
If it sounds fishy, it likely is.
95
00:02:55,033 --> 00:02:57,066
And finally. Three.
96
00:02:57,066 --> 00:02:58,166
Use hallucinations
97
00:02:58,166 --> 00:02:59,300
as a teachable moment
98
00:02:59,300 --> 00:03:02,300
about fact checking and media literacy.
99
00:03:02,466 --> 00:03:04,066
AI is powerful,
100
00:03:04,066 --> 00:03:06,000
but it can still make things up.
101
00:03:06,000 --> 00:03:07,466
It's important to be skeptical
102
00:03:07,466 --> 00:03:09,133
and check the sources.
103
00:03:09,133 --> 00:03:10,133
As educators,
104
00:03:10,133 --> 00:03:11,933
knowing how AI hallucinations
105
00:03:11,933 --> 00:03:13,100
work helps us separate
106
00:03:13,100 --> 00:03:14,966
the hype from the reality
107
00:03:14,966 --> 00:03:16,066
so we can make wise
108
00:03:16,066 --> 00:03:17,366
choices for our classrooms.
