Use of
Artificial Intelligence in the Learning of Newtonian Mechanics: A Didactic
Intervention in First-Year Undergraduate Students
Uso de inteligencia artificial en el aprendizaje de
la mecánica newtoniana: una intervención didáctica en estudiantes de primer año
de licenciatura
Christian Antonio Pavón
Brito*
Antonio Eduardo Felipe*
Olga Beatriz Ávila*
Introduction
The teaching of physics, especially Newtonian
mechanics, represents one of the most complex challenges in university science
education. Despite methodological and curricular advances, significant barriers
to students' conceptual learning persist. One of the main obstacles is the
presence of alternative conceptions, that is, intuitive ideas constructed from
everyday experiences that contradict the principles of formal physics (Flores-García et al., 2008; Chi, 2005). These deeply rooted ideas
are resistant to change even after formal instruction and hinder the
development of a structured and meaningful understanding of key concepts such
as force, motion, and inertia.
International studies have highlighted this problem
in multiple contexts. For example, the use of the Force Concept Inventory
(FCI), a widely validated instrument for measuring understanding of classical
mechanics, has revealed low levels of understanding even after physics courses
in higher education. In Latin America, research conducted in Mexico, Colombia,
Chile, and Argentina reports levels of conceptualization that rarely exceed
70%, placing most students at an average level of content assimilation (Artamónova, Mosquera-Mosquera
& Mosquera-Artamónov, 2017; Budini
et al., 2019). These results suggest that formal education, although necessary,
does not always succeed in bringing about profound conceptual change.
Ecuador is no stranger to this situation. Research
conducted in university settings has shown that students have levels of
conceptual assertiveness in Newtonian mechanics between 38% and 43%, with
recurring difficulties in interpreting problems and applying appropriate
formulas (Varas, Villalva
& Nieto, 2018). In particular, a study of students in the Experimental
Science Education program at the University of Guayaquil showed that, after
traditional classes, students were unable to understand projectile motion or
correctly apply the principles of vector analysis, obtaining average grades
below 4 out of 10 (Amaguaya & Castro, 2022).
These deficiencies compromise not only academic performance but also the future
teaching practice of these students.
Faced with this situation, the educational community
has explored various teaching strategies aimed at improving conceptual
understanding in physics. Among these, peer instruction has shown encouraging
results. This approach, developed by Eric Mazur (1997), is based on discussion
among students to confront ideas, justify answers, and collectively arrive at a
more solid understanding. Several studies have shown that this methodology not
only promotes active learning but also contributes significantly to conceptual
change by exposing students to cognitive conflicts in a collaborative
environment (Crouch & Mazur, 2001; Smith et al., 2009).
However, even active methodologies can have
limitations if they are not adapted to the individual characteristics of the
student. In this regard, artificial intelligence (AI) has emerged as a
promising tool in the field of education. Thanks to its natural language processing,
pattern analysis, and learning personalization capabilities, AI can offer
adaptive environments that respond in real time to the cognitive needs of each
user (Woolf, 2010; Chen, Chen & Zhijian, 2020).
These environments have the potential to transform the teaching of abstract
concepts such as those in classical mechanics, providing explanations,
examples, exercises, and feedback in an immediate and contextualized manner.
The application of AI in science education has begun
to spread worldwide. Tools such as intelligent tutors, adaptive learning
systems, and virtual assistants are being used to improve conceptual
understanding and reduce barriers to learning (VanLehn,
2011; Kasinathan, Mustapha & Medi,
2017). For example, platforms such as DreamBox and Labster have proven effective in personalizing learning in
mathematics and natural sciences, allowing students to
explore at their own pace, make mistakes, and receive personalized guidance.
In the particular case of Newtonian mechanics, these
technologies allow for the simulation of dynamic scenarios, guided questions,
and the generation of visualizations of complex physical phenomena. In this
way, students not only access explanations, but also interact with the content,
explore their hypotheses, and correct their mistakes, which
is key to the development of structured scientific thinking (Rosales
& Sulaiman, 2020).
At the regional level, Latin America has begun to
explore the use of artificial intelligence in education with initiatives that
are still disparate. Countries such as Brazil and Chile are leading projects
that incorporate intelligent tutoring systems, adaptive learning platforms, and
educational simulations at different levels of the education system (Gómez, Del
Pozo, Martínez & Martín
del Campo, 2020). In these countries, results have shown improvements in
students' conceptual understanding, particularly in areas such as mathematics
and science, and an increase in content retention thanks to more personalized
interactive environments.
In contrast, other countries such as Mexico and
Argentina are in the early stages of implementation. Although platforms
offering personalized recommendations or academic performance prediction
systems already exist, limitations remain in terms of access to technological
infrastructure and teacher training in the pedagogical use of these tools (Peñaherrera, Cunuhay, Nata &
Moreira, 2022).
The case of Ecuador reflects similar challenges, but
also opportunities. Although there are isolated efforts in some higher education
institutions, such as the University of Guayaquil and the National Polytechnic
School, the adoption of AI in classrooms is still in its infancy (Castillo,
Tapia, Placencia & Pavón,
2024). Limited connectivity in rural areas, a lack of adequate technological
devices, and poor teacher training in the use of AI-based resources hinder
systematic implementation. Despite this, the growing availability of accessible
platforms, the presence of educational digitization initiatives,
and academic interest in pedagogical innovations constitute fertile ground for
progress in this direction.
In this context, tools such as ChatGPT,
a language model developed by OpenAI, have begun to
be used as virtual assistants in the educational process. Its ability to offer
detailed explanations, answer open-ended questions, propose contextualized
exercises, and generate immediate feedback makes it a powerful resource for
promoting autonomous learning. Its potential does not lie in replacing
teachers, but in complementing the teaching process, facilitating access to
knowledge outside the classroom, and allowing for more individualized attention
to students (Moreno, 2019; Peñaherrera et al., 2022).
In this sense, the integration of AI-based tools
such as ChatGPT into physics teaching may represent a
viable solution to persistent problems of conceptual understanding, especially
in an environment where the number of students per classroom makes personalized
attention and the timely identification of conceptual errors difficult. If
appropriate AI-mediated teaching activities are designed, students can actively
interact with the content, verify their answers, receive explanations tailored
to their level, and, above all, reconstruct their misconceptions through
dialogue with a system that responds immediately.
Thus, there is a need to investigate how this type
of intervention can affect students' actual learning, especially in populations
with a history of low performance in physics. Based on the above, the objective
of this study is to evaluate the effect of guided activities mediated by
artificial intelligence on the level of conceptual knowledge of Newtonian
mechanics in first-year students of the Bachelor's Degree in Physics Education
at the University of Guayaquil during the May-August 2024 semester, comparing
the results obtained with those of a control group that received traditional
peer instruction.
Materials
and methods
This study was developed using a quantitative
approach and a quasi-experimental design with a non-equivalent control group,
which allowed us to evaluate the effects of an artificial intelligence
(AI)-mediated teaching intervention on the learning of Newtonian mechanics. The
purpose of this first phase, corresponding to the May-August 2024 semester, was
to measure the variation in students' conceptual knowledge after applying a
specific teaching sequence, comparing the performance between a control group
and an experimental group.
The quasi-experimental design was selected due to
the impossibility of randomly assigning participants to groups. Instead, two
parallel courses within the same university degree program were used, which
made it possible to reduce the effects of external variables without altering
the existing academic structure (Hernández, Fernández
& Baptista, 2014).
The population consisted of first-year students in
the Bachelor's Degree in Physics Education at the University of Guayaquil. For
this phase of the study, two parallel groups were used:
·
The control group, made up
of 33 students from the morning course, received traditional peer instruction.
·
The experimental group, made
up of 30 students from the afternoon class, participated in guided activities
using AI through a web platform based on ChatGPT.
Both groups shared similar characteristics in terms
of age (between 17 and 20 years old), previous education, and institutional
context. In addition, all participants were taking the Physics I course, which
focused on the fundamentals of classical mechanics.
Two multiple-choice written tests were used to assess conceptual knowledge: a
diagnostic test (pre-test) administered at the beginning of the semester and an
exit test (post-test) at the end of the semester. Both tests were adapted from
the Force Concept Inventory (FCI) instrument, which has demonstrated validity
and reliability in measuring the understanding of key concepts in Newtonian
mechanics (Hestenes, Wells & Swackhamer,
1992).
Each test consisted of 30 items covering the
following topics:
·
Newton's laws of motion
·
Net force and acceleration
·
Uniformly accelerated
rectilinear motion
·
Free fall
and projectile trajectory
·
Conservation of linear momentum
The answers were coded numerically and the scores
were normalized to facilitate subsequent statistical analysis. In addition, a
sociodemographic survey was administered to characterize the study population.
During the May–August 2024 semester, the
intervention was implemented in the experimental group. It consisted of a
guided teaching sequence composed of five thematic modules, each accompanied by
interactive exercises, simulations, and virtual assistance via ChatGPT. The AI tool was used as a personalized tutor,
available to answer questions, clarify concepts, and provide immediate feedback
at any time during the process.
Each module addressed a fundamental theme of
Newtonian mechanics, with a structure based on:
·
Initial diagnosis: trigger
question or preliminary simulation.
·
Theoretical exploration:
guided reading + open-ended questions directed to ChatGPT.
·
Practical application: contextualized
exercises and problem solving.
·
Automatic feedback:
immediate review of answers with justification.
·
Metacognitive reflection:
final questions to identify conceptual changes.
The control group, meanwhile, developed the same
content, but through face-to-face sessions focused on the peer instruction
model (Mazur, 1997), where students discussed their answers in small groups
under the teacher's guidance.
The study was conducted in three phases:
·
Diagnostic phase:
application of the pretest and sociodemographic survey in both groups.
·
Intervention phase:
implementation of differentiated teaching strategies for 14 weeks.
·
Evaluation phase:
application of the posttest and comparative analysis of the results.
The data were analyzed using SPSS software. Normality
tests (Shapiro-Wilk), descriptive measures (mean and standard deviation), and
t-tests for independent samples with a significance level of α = 0.05 were applied to
determine significant differences between the two groups.
Results
This section presents the findings obtained after
applying the AI-mediated teaching intervention during the May-August 2024
semester. The analysis focuses on comparing the levels of conceptual knowledge
in Newtonian mechanics between the experimental group, which worked with ChatGPT as a virtual assistant, and the control group,
which followed a traditional peer-to-peer instruction model.
Pretest results (diagnostic test)
The pretest results showed that both groups started
from similar levels of conceptual understanding. The mean score obtained by the
experimental group was 4.18/10 (SD = 1.01), while the control group achieved a
mean score of 4.10/10 (SD = 1.05). The independent samples t-test yielded a
value of t(61) = 0.30; p = 0.76, indicating that there
were no statistically significant differences between the groups at the
beginning of the training process.
These initial results corroborate the homogeneity of
the groups in terms of prior knowledge and validate the subsequent comparison
based on the effects of the intervention.
Post-test results (exit test)
At the end of the semester, the exit test was
administered to measure the impact of the differentiated teaching strategies.
The results show a clear improvement in both groups, although more pronounced
in the experimental group.
The experimental group obtained an average of
7.72/10 (SD = 0.94), representing an increase of 3.54 points compared to the
pretest.
The control group obtained an average of 5.60/10 (SD
= 1.02), increasing by 1.50 points on average.
The independent samples t-test applied to the
post-test results revealed a statistically significant difference between the
two groups (t(61) = 8.31; p < 0.001), indicating
that the use of guided activities with artificial intelligence had a
considerable positive effect on conceptual learning.
Complementary qualitative analysis
During the intervention, it was observed that
students in the experimental group used the virtual assistant to clarify
concepts and verify their answers in real time. The most frequently asked
questions were related to the interpretation of acceleration and velocity
graphs, the analysis of forces in systems with friction, and the understanding
of the principle of inertia. This continuous interaction with the AI system seemed
to favor more autonomous and reflective learning.
In addition, greater motivation and active
participation were evident in the experimental group. The availability of a
resource that responded immediately and adaptively gave students the confidence
to take on complex conceptual challenges without fear of making mistakes.
Conclusions
The results of this first phase of the
quasi-experimental study provide clear evidence of the potential of artificial
intelligence (AI) as a teaching resource for improving conceptual learning in
Newtonian mechanics in university settings. In particular, the integration of ChatGPT into a sequence of guided activities allowed
students in the experimental group not only to achieve better scores on the exit
test but also to develop a more autonomous, reflective, and meaningful learning
process.
From a pedagogical point of view, the substantial improvement observed in the
experimental group compared to the control group can be attributed to several
key factors. One of these is the personalization of learning. AI provided
immediate responses tailored to each student's level of understanding, allowing
for constant feedback that is difficult to achieve in traditional environments
with high student density (VanLehn, 2011; Woolf,
2010). This personalized support is essential in teaching content such as
classical mechanics, which often generates high rates of misunderstanding due
to the presence of persistent alternative conceptions (Chi, 2005; Flores-García et al., 2008).
In addition, AI acted as a facilitator of conceptual
change by allowing students to confront their own ideas, reformulate
hypotheses, and test their reasoning in real time. These opportunities for
interaction increased the depth of learning, as Rosales and Sulaiman
(2020) have also pointed out in similar studies on AI-assisted physics
teaching.
Another important finding was the higher level of
motivation and participation observed in the experimental group. Continuous
access to a virtual assistant gave students a sense of control over their
learning process, reducing academic anxiety and strengthening their
self-confidence in tackling new problems. This is in line with recent studies
that highlight the positive emotional impact that the use of AI can have in
education, provided that it is mediated by appropriate pedagogical strategies
(Chen, Chen & Zhijian, 2020; Moreno, 2019).
The comparison with the control group is also
revealing. Although students in this group participated in an active modality
through peer instruction—a strategy that has been shown to be effective in
various studies (Mazur, 1997; Smith et al., 2009)—learning outcomes were
significantly lower. This does not invalidate the effectiveness of the
traditional model, but rather suggests that the use of AI can complement and
enhance existing strategies, creating hybrid environments where opportunities
for conceptual interaction and metacognition are maximized.
At the methodological level, the application of
diagnostic and exit tests, together with a quasi-experimental design, allowed
for the control of external variables and ensured the validity of the results.
The similarity of initial scores between the two groups confirms that the
improvement observed in the experimental group is not due to prior differences,
but to the direct effects of the intervention.
However, this study also faces limitations that
should be considered in future research. The duration of the intervention,
restricted to a single semester, prevents the evaluation of long-term effects.
In addition, the sample was limited to a specific institution and degree
program, so the results should be generalized with caution. Dependence on
connectivity and access to technology is another critical factor, especially in
contexts such as Ecuador, where significant digital infrastructure gaps still
exist (Peñaherrera et al., 2022).
Despite this, the findings allow us to conclude that
artificial intelligence, when integrated in a planned manner into the teaching
process, can be an effective tool for improving conceptual learning in physics.
In particular, its use with first-year undergraduate students helps to overcome
initial cognitive barriers, strengthen critical thinking, and promote active
and personalized learning.
This study not only confirms previous results in
international contexts, but also provides local evidence on the feasibility of
applying AI in higher science education in Ecuador. It is therefore suggested
that teachers consider the progressive incorporation of AI-based tools as a
complement to their teaching strategies, especially in areas where conceptual
abstraction represents a persistent difficulty.
Finally, the results obtained open up new lines of research. Future studies
could explore the emotional impact of interaction with AI systems, analyze the
individual learning trajectories that these systems allow to be traced, and
design hybrid methodologies that integrate the best of traditional teaching
with the advantages of intelligent and adaptive learning.
..........................................................................................................
References
Amaguaya, M., & Castro, D.
(2022). Dificultades en la comprensión del movimiento de proyectiles en
estudiantes de Física. Universidad de Guayaquil.
Artamónova, V., Mosquera-Mosquera, C., & Mosquera-Artamónov,
D. (2017). Análisis del nivel de comprensión conceptual en mecánica clásica a
través del FCI en universidades latinoamericanas. Revista Latinoamericana
de Física Educativa, 11(2), 45–53.
Budini, P., Etchegoyen, A., Mazza, G., & Romero,
M. (2019). Estudio del rendimiento conceptual en Física I en estudiantes
universitarios. Enseñanza de las Ciencias, 37(1), 135–153.
Castillo, G., Tapia, A., Placencia, L., & Pavón, C. (2024). Implementación
incipiente de inteligencia artificial en la educación universitaria
ecuatoriana. Revista Ecuatoriana de Tecnología Educativa, 8(1), 22–39.
Chen, Y., Chen,
L., & Zhijian, L. (2020). Artificial Intelligence in Education: Applications and Challenges. International Journal of Educational Development, 77,
102249. https://doi.org/10.1016/j.ijedudev.2020.102249
Chi, M. T. H.
(2005). Commonsense conceptions of emergent processes: Why
some misconceptions are robust. Journal of the Learning Sciences,
14(2), 161–199.
Crouch, C. H., & Mazur, E. (2001). Peer
Instruction: Ten years of experience and results. American Journal of
Physics, 69(9), 970–977. https://doi.org/10.1119/1.1374249
Flores-García, R., Carrascosa-Alís, J., Dreyfus, B., Hoehn, J., Elby, A., Finkelstein, N., & Gupta, A.
(2008). Concepciones alternativas
sobre la fuerza y el movimiento en estudiantes universitarios. Revista de
Enseñanza de la Física, 21(1), 5–20.
Gómez, D., Del Pozo, E.,
Martínez, L., & Martín del Campo, J. (2020). Panorama de la inteligencia
artificial en la educación en América Latina. Revista Iberoamericana de
Educación, 84(2), 115–138.
Hernández, R., Fernández, C.,
& Baptista, P. (2014). Metodología de la investigación (6.ª ed.). McGraw-Hill.
Hestenes, D., Wells, M., & Swackhamer, G. (1992).
Force Concept Inventory. The Physics Teacher, 30(3), 141–158.
https://doi.org/10.1119/1.2343497
Kasinathan, G., Mustapha, A., & Medi, S. (2017).
Adaptive Learning Systems and Personalized Learning in Higher Education: A
Case Study. International Education Studies, 10(12), 50–59.
https://doi.org/10.5539/ies.v10n12p50
Mazur, E. (1997). Peer Instruction: A User’s
Manual. Prentice Hall.
Moreno, S. (2019).
Inteligencia artificial en la educación: personalización del aprendizaje en
física. Revista de Tecnología y Didáctica, 5(2), 33–45.
Peñaherrera, P., Cunuhay, A., Nata, Y., & Moreira, D.
(2022). El rol de la inteligencia artificial en la educación superior
ecuatoriana. Revista Científica del Ecuador, 6(1), 77–94.
Rosales, D.,
& Sulaiman, T. (2020). Effects of AI-Supported Modules in the Teaching of
Newtonian Mechanics. International Journal of Learning and Teaching,
12(4), 334–341.
Smith, M. K.,
Wood, W. B., Adams, W. K., Wieman, C., Knight, J.
K., Guild, N., & Su, T. T. (2009). Why peer
discussion improves student performance on in-class concept questions. Science,
323(5910), 122–124. https://doi.org/10.1126/science.1165919
VanLehn, K. (2011). The relative effectiveness of human
tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221.
Varas, M., Villalva, E., & Nieto, M. (2018). Diagnóstico del
conocimiento conceptual en mecánica clásica en estudiantes de nivelación y
educación superior en Ecuador. Revista de Ciencias Físicas y Aplicadas,
10(2), 123–138.
Woolf, B. P. (2010). Building Intelligent Interactive Tutors: Student-centered strategies for revolutionizing e-learning. Morgan Kaufmann.