WEBVTT

00:00:01.125 --> 00:00:01.685
Meet Ada.

00:00:01.745 --> 00:00:02.845
Her homework is finished.

00:00:03.045 --> 00:00:04.645
The answer looks perfectly clear.

00:00:04.785 --> 00:00:06.805
But close the screen, and something changes.

00:00:08.750 --> 00:00:10.990
An excellent answer is one kind of success.

00:00:11.210 --> 00:00:14.010
Being able to explain and
adapt it yourself is another.

00:00:14.230 --> 00:00:16.890
They can grow together, but one
does not guarantee the other.

00:00:18.750 --> 00:00:21.590
Does artificial intelligence
help you learn, or just finish?

00:00:21.890 --> 00:00:24.290
The difference shows up
when the help goes away.

00:00:26.250 --> 00:00:27.950
Finishing with help is one test.

00:00:28.070 --> 00:00:29.490
Doing it alone is another.

00:00:29.750 --> 00:00:32.970
Remembering later, and solving a
different problem, are two more.

00:00:34.750 --> 00:00:37.450
With the tool open, you are
measuring a partnership.

00:00:37.730 --> 00:00:39.250
The result can be impressive.

00:00:39.590 --> 00:00:42.550
It still leaves open how much of
the method belongs to the learner.

00:00:44.250 --> 00:00:46.590
Remove the assistance and try a fresh problem.

00:00:46.910 --> 00:00:50.520
Can you choose a first step, explain
your choices, and notice when the result

00:00:50.530 --> 00:00:51.270
makes no sense?

00:00:53.250 --> 00:00:56.150
A later test asks what
remains after time passes.

00:00:56.410 --> 00:01:00.190
A changed situation asks whether you
recognize where the method applies.

00:01:00.410 --> 00:01:02.350
Those questions deserve separate evidence.

00:01:04.125 --> 00:01:07.475
In a randomized study of nearly a thousand
high school mathematics students in

00:01:07.485 --> 00:01:10.735
Turkey, an unrestricted chatbot
improved practice grades by forty eight

00:01:10.745 --> 00:01:13.245
percent, compared with the control group.

00:01:14.750 --> 00:01:17.870
That figure is relative to the
control group's practice grades.

00:01:18.070 --> 00:01:21.520
It does not mean forty eight extra
percentage points, or the same improvement

00:01:21.530 --> 00:01:22.550
for every student.

00:01:24.250 --> 00:01:27.890
But on the exam without the tool, that
group scored seventeen percent lower.

00:01:28.090 --> 00:01:31.190
Better homework performance had hidden
weaker independent performance.

00:01:33.250 --> 00:01:37.040
Random assignment provides stronger
evidence than comparing people who chose a

00:01:37.050 --> 00:01:38.670
chatbot with those who did not.

00:01:38.810 --> 00:01:41.770
But the result still belongs
to this particular setting.

00:01:44.250 --> 00:01:48.710
A tutor designed around teacher supplied
hints largely removed that penalty.

00:01:48.870 --> 00:01:52.600
It did not produce a statistically
significant exam improvement over the

00:01:52.610 --> 00:01:53.350
control group.

00:01:55.750 --> 00:01:58.110
The guarded tutor used teacher designed hints.

00:01:58.250 --> 00:02:01.410
Its result does not mean any
friendly chatbot will teach well.

00:02:01.570 --> 00:02:03.430
The design of the help matters.

00:02:05.250 --> 00:02:09.180
In a separate college physics experiment,
a carefully designed tutor led to

00:02:09.190 --> 00:02:12.270
higher immediate test scores
than an active classroom lesson.

00:02:13.750 --> 00:02:17.470
This comparison involved a structured
tutor and an active classroom lesson.

00:02:17.570 --> 00:02:21.210
It compares those teaching arrangements,
not all machines with all teachers.

00:02:23.250 --> 00:02:25.170
These are different designs
and different settings.

00:02:25.470 --> 00:02:28.190
Neither result tells us what every
learner will remember months later.

00:02:30.250 --> 00:02:34.150
Together, these studies suggest asking
which designs preserve the learner's

00:02:34.190 --> 00:02:34.580
thinking.

00:02:34.590 --> 00:02:38.180
They do not establish one universal
percentage for how much artificial

00:02:38.190 --> 00:02:40.070
intelligence helps or harms learning.

00:02:42.125 --> 00:02:45.545
Read a clear explanation twice
and its steps can feel obvious.

00:02:45.745 --> 00:02:48.815
That feeling does not prove you could
reconstruct the explanation without

00:02:48.825 --> 00:02:49.375
looking.

00:02:51.750 --> 00:02:54.730
When the page disappears,
its prompts disappear too.

00:02:54.950 --> 00:02:58.210
You must select the idea, retrieve
the steps, and organize them.

00:02:58.350 --> 00:03:00.470
A blank sheet can expose hidden gaps.

00:03:02.250 --> 00:03:06.760
In a classic prose passage experiment,
repeated study looked better on the

00:03:06.770 --> 00:03:07.690
immediate test.

00:03:07.830 --> 00:03:11.510
Earlier recall tests helped more when
people were tested after a delay.

00:03:13.750 --> 00:03:16.510
Repeated studying also increased confidence.

00:03:16.730 --> 00:03:19.710
Yet the delayed advantage
favored retrieval practice.

00:03:19.930 --> 00:03:23.140
Feeling fluent and remembering later
were different outcomes in these

00:03:23.150 --> 00:03:23.730
experiments.

00:03:25.750 --> 00:03:28.870
Try recalling before reopening
the answer, then check your work.

00:03:28.990 --> 00:03:30.830
Correct the missing or mistaken parts.

00:03:31.010 --> 00:03:34.340
An unsuccessful attempt is information,
not a reason to keep practising

00:03:34.350 --> 00:03:34.900
errors.

00:03:36.750 --> 00:03:40.220
In a mathematics experiment by Rohrer
and Taylor, practice spread across

00:03:40.230 --> 00:03:43.940
sessions produced better performance a
week later than practice concentrated

00:03:43.950 --> 00:03:44.890
in one session.

00:03:46.750 --> 00:03:50.460
Their second experiment favored mixing
problem types over blocking them on a

00:03:50.470 --> 00:03:51.210
later test.

00:03:51.710 --> 00:03:55.150
Mixing asks you to identify the
problem before selecting a procedure.

00:03:57.250 --> 00:04:00.050
These studies support independent
practice and later checks.

00:04:00.090 --> 00:04:03.840
They do not prove that one particular
chatbot prompt or practice schedule will

00:04:03.850 --> 00:04:04.950
work for every person.

00:04:07.125 --> 00:04:09.845
So, when learning is the goal,
give the tool a different job.

00:04:09.985 --> 00:04:12.065
Let it support the thinking
you need to practise.

00:04:15.250 --> 00:04:15.740
Step one.

00:04:15.750 --> 00:04:16.630
Make an attempt.

00:04:16.770 --> 00:04:19.530
If you know enough to start, put
your own first step on the page.

00:04:22.250 --> 00:04:25.310
For a completely new topic, study
a clear worked example first.

00:04:25.530 --> 00:04:28.170
A beginner needs guidance,
not a guessing contest.

00:04:30.250 --> 00:04:33.910
If the notation is unfamiliar,
struggling alone can be unhelpful.

00:04:33.950 --> 00:04:36.490
A worked example provides a path to inspect.

00:04:36.770 --> 00:04:38.410
Can the learner explain that path?

00:04:40.250 --> 00:04:42.170
Both sides of Ada's equation are equal.

00:04:42.350 --> 00:04:44.470
Each operation must preserve that equality.

00:04:44.570 --> 00:04:47.610
She is keeping the balance true,
not moving symbols by magic.

00:04:49.750 --> 00:04:52.650
Now ask for a hint about the
exact step you cannot explain.

00:04:52.930 --> 00:04:55.850
Ask it to check your reasoning
before showing the whole solution.

00:04:57.750 --> 00:05:00.690
Ada needs to solve two times a
number, plus three, equals eleven.

00:05:00.830 --> 00:05:03.150
Her hint is simple: undo the addition first.

00:05:04.750 --> 00:05:07.070
Subtracting three undoes
the addition on the left.

00:05:07.350 --> 00:05:08.990
Doing the same on the right leaves eight.

00:05:09.170 --> 00:05:11.950
Two times the unknown must
therefore equal eight.

00:05:13.750 --> 00:05:15.110
Divide both sides by two.

00:05:15.270 --> 00:05:17.330
One copy of the unknown equals four.

00:05:17.430 --> 00:05:20.590
That reasoning still applies when the
next equation uses different numbers.

00:05:22.750 --> 00:05:25.190
Explain why three is
subtracted from both sides.

00:05:25.330 --> 00:05:28.680
If your only reason is that the chatbot
said so, return to the meaning of

00:05:28.690 --> 00:05:29.240
equality.

00:05:34.750 --> 00:05:35.370
Step two.

00:05:35.390 --> 00:05:36.270
Hide the answer.

00:05:36.330 --> 00:05:37.270
Close the screen.

00:05:37.590 --> 00:05:39.890
Take a fresh sheet, and explain
the method from memory.

00:05:42.750 --> 00:05:44.650
Ada subtracts three from both sides.

00:05:44.690 --> 00:05:46.590
Then she divides both sides by two.

00:05:46.670 --> 00:05:47.610
The number is four.

00:05:47.730 --> 00:05:48.930
Each step has a reason.

00:05:50.750 --> 00:05:52.670
Suppose Ada cannot explain the division.

00:05:52.850 --> 00:05:56.270
She can reopen that step,
study its reason, and retry.

00:05:56.310 --> 00:05:59.590
She need not replace her entire
attempt with a copied solution.

00:06:01.750 --> 00:06:05.840
If she subtracts three on the left but
forgets the right, the work can still

00:06:05.850 --> 00:06:06.690
look orderly.

00:06:07.170 --> 00:06:11.080
Substituting her proposed answer into
the original equation exposes the

00:06:11.090 --> 00:06:11.640
mistake.

00:06:13.250 --> 00:06:13.950
Check the work.

00:06:14.030 --> 00:06:16.310
Put four back into the original equation.

00:06:16.390 --> 00:06:18.670
Two times four, plus three, is eleven.

00:06:18.810 --> 00:06:20.310
If something fails, correct it.

00:06:21.750 --> 00:06:25.630
Useful feedback names the error: apply
the same operation to both sides.

00:06:25.950 --> 00:06:28.290
Ada corrects that step and
tries another equation.

00:06:28.310 --> 00:06:29.670
A green check tells her less.

00:06:32.125 --> 00:06:32.785
Step three.

00:06:32.865 --> 00:06:33.885
Change the problem.

00:06:33.945 --> 00:06:35.045
New numbers are a start.

00:06:35.105 --> 00:06:37.845
A new situation checks whether you
can choose the method yourself.

00:06:40.750 --> 00:06:43.590
Two notebooks and a three dollar
delivery fee cost fifteen dollars.

00:06:43.970 --> 00:06:45.570
What does one notebook cost?

00:06:45.830 --> 00:06:48.010
Ada has to build the
equation before solving it.

00:06:49.750 --> 00:06:52.250
Let the unknown name one notebook's price.

00:06:52.330 --> 00:06:54.890
Two identical notebooks
cost twice that amount.

00:06:55.070 --> 00:06:58.150
Delivery is charged once, adding
three dollars to the total.

00:06:59.750 --> 00:07:02.010
Remove the three dollar delivery
fee from fifteen dollars.

00:07:02.390 --> 00:07:04.270
Twelve dollars remain for two notebooks.

00:07:04.350 --> 00:07:05.090
Divide by two.

00:07:05.210 --> 00:07:06.850
The price is six dollars per notebook.

00:07:06.970 --> 00:07:09.290
This calculation follows the
same equation structure.

00:07:10.750 --> 00:07:12.890
Two notebooks at six dollars cost twelve.

00:07:12.970 --> 00:07:14.870
Add the three dollar fee: fifteen.

00:07:15.190 --> 00:07:18.310
The check connects the calculation
back to the situation in the question.

00:07:19.750 --> 00:07:22.790
If delivery is charged for each
notebook, the equation changes.

00:07:23.090 --> 00:07:25.490
Similar words can hide
different relationships.

00:07:25.630 --> 00:07:28.590
Explain which costs repeat
before choosing a calculation.

00:07:30.250 --> 00:07:32.580
Come back later for another
fresh problem, without the tool.

00:07:32.590 --> 00:07:35.670
That checks retention separately
from what feels easy right now.

00:07:38.250 --> 00:07:40.430
Use a fresh question for the later attempt.

00:07:40.490 --> 00:07:41.750
Put the old solution away.

00:07:41.910 --> 00:07:45.410
Record what you can do independently,
then decide what to revisit.

00:07:48.250 --> 00:07:50.670
Before checking, predict what you can explain.

00:07:50.830 --> 00:07:53.290
Compare that prediction with your actual work.

00:07:53.410 --> 00:07:57.510
Use the gap to choose your next
practice, rather than relying on ease.

00:07:59.625 --> 00:08:02.465
For writing, a chatbot's rewrite
may improve the document.

00:08:02.665 --> 00:08:06.715
To practise argument, Ada can first
state her own claim and the evidence she

00:08:06.725 --> 00:08:07.685
thinks supports it.

00:08:09.250 --> 00:08:13.170
Ask which sentence is unsupported, or
what a skeptical reader would question.

00:08:13.450 --> 00:08:16.890
Then revise the paragraph
yourself and explain the change.

00:08:16.990 --> 00:08:18.590
Keep that decision with the writer.

00:08:20.250 --> 00:08:23.110
Later, write another paragraph
without those suggestions.

00:08:23.430 --> 00:08:27.360
This adapts the routine to writing;
it is not evidence that these exact

00:08:27.370 --> 00:08:29.590
prompts have been experimentally validated.

00:08:32.750 --> 00:08:34.970
For programming, explain
what each line changes.

00:08:35.050 --> 00:08:38.190
Predict the result for a specific
input, then run the function.

00:08:38.270 --> 00:08:40.130
Compare the execution with your prediction.

00:08:43.250 --> 00:08:45.700
This function returns the
first item in a list.

00:08:45.710 --> 00:08:47.330
It works for a nonempty list.

00:08:47.410 --> 00:08:48.930
What happens with an empty one?

00:08:49.050 --> 00:08:51.390
Test the assumption behind that first success.

00:08:53.250 --> 00:08:55.650
She adds an explicit check for an empty list.

00:08:55.690 --> 00:08:57.690
She explains what each branch returns.

00:08:57.830 --> 00:09:01.280
Then she tests a new requirement on
her own, instead of simply accepting a

00:09:01.290 --> 00:09:01.840
patch.

00:09:03.750 --> 00:09:05.350
A clear explanation can be wrong.

00:09:05.470 --> 00:09:09.540
Check equations by substitution, code
with suitable tests, and factual claims

00:09:09.550 --> 00:09:10.790
against reliable sources.

00:09:11.010 --> 00:09:13.370
The system's confidence is
not independent verification.

00:09:15.250 --> 00:09:19.170
Language or accessibility support can
free attention for the target skill.

00:09:19.410 --> 00:09:23.540
Practise that skill independently where
appropriate; do not demand that every

00:09:23.550 --> 00:09:24.850
useful support disappear.

00:09:26.750 --> 00:09:28.710
Sometimes a finished product is the goal.

00:09:29.030 --> 00:09:30.370
That is a different choice.

00:09:30.570 --> 00:09:34.840
Decide what to delegate, what to verify,
and which capability you want to

00:09:34.850 --> 00:09:35.400
retain.

00:09:37.125 --> 00:09:38.485
Choose one skill to retain.

00:09:38.625 --> 00:09:41.205
Pick a suitable question and
the help needed to begin.

00:09:41.405 --> 00:09:44.305
Reserve a fresh question for
a later independent attempt.

00:09:46.750 --> 00:09:48.140
Attempt with suitable guidance.

00:09:48.150 --> 00:09:49.710
Hide, explain, and check.

00:09:49.830 --> 00:09:51.630
Change the problem, then return later.

00:09:51.810 --> 00:09:53.890
Keep the next independent attempt in view.

00:09:55.250 --> 00:09:58.640
Record your attempt, the help used, the
error corrected, and what you could do

00:09:58.650 --> 00:09:59.180
later.

00:09:59.190 --> 00:10:02.470
Look for progress in your own work,
not the number of chatbot messages.

00:10:04.750 --> 00:10:07.890
This is a practical check, not
a universally proven recipe.

00:10:08.110 --> 00:10:10.810
The right amount of help depends
on the learner and the task.

00:10:12.750 --> 00:10:14.190
A finished page is useful.

00:10:14.350 --> 00:10:16.570
A skill you can use again is something more.

00:10:16.650 --> 00:10:18.490
What can you do when the answer is gone?