ChatGPT is not a pocket calculator -- Problems of AI-chatbots for teaching Geography

Abstract

The recent success of large language models and AI chatbots such as ChatGPT in various knowledge domains has a severe impact on teaching and learning Geography and GIScience. The underlying revolution is often compared to the introduction of pocket calculators, suggesting analogous adaptations that prioritize higher-level skills over other learning content. However, using ChatGPT can be fraudulent because it threatens the validity of assessments. The success of such a strategy therefore rests on the assumption that lower-level learning goals are substitutable by AI, and supervision and assessments can be refocused on higher-level goals. Based on a preliminary survey on ChatGPT's quality in answering questions in Geography and GIScience, we demonstrate that this assumption might be fairly naive, and effective control in assessments and supervision is required.

Full text

ChatGPT is not a pocket calculator -
Problems of AI-chatbots for teaching Geography
Simon Scheider
Utrecht University, Department of Human Geography and Spatial Planning, The Netherlands
s.scheider@uu.nl
Harm Bartholomeus
Wageningen University, Laboratory of Geo-Information Science and Remote Sensing, The
Netherlands
harm.bartholomeus@wur.nl
Judith Verstegen
Utrecht University, Department of Human Geography and Spatial Planning, The Netherlands
j.a.verstegen@uu.nl
Abstract
The recent success of large language models and AI chatbots such as ChatGPT in various knowledge
domains has a severe impact on teaching and learning Geography and GIScience. The underlying
revolution is often compared to the introduction of pocket calculators, suggesting analogous adapta-
tions that prioritize higher-level skills over other learning content. However, using ChatGPT can be
fraudulent because it threatens the validity of assessments. The success of such a strategy therefore
rests on the assumption that lower-level learning goals are substitutable by AI, and supervision and
assessments can be refocused on higher-level goals. Based on a preliminary survey on ChatGPT’s
quality in answering questions in Geography and GIScience, we demonstrate that this assumption
might be fairly naive, and effective control in assessments and supervision is required.
2012 ACM Subject Classification Social and professional topics; Computing methodologies
→
Artificial intelligence
Keywords and phrases validity, assessments, AI chatbots, geography exams
Acknowledgements We want to thank all our survey participants. The work was supported by
the European Research Council (ERC) under the European Union’s Horizon 2020 research and
innovation programme (grant agreement No 803498).
1 Introduction
The introduction of transformer-based large language models, such as GPT-3, and related
chatbot services such as ChatGPT, has hit teachers hard on all levels of the educational
system, from primary schools to universities. In the subsequent debate, voices in academia
and practice range from alarm calls to appeals to remain calm and accept the unavoidable.
In the latter camp, the argument goes that teachers should embrace this development and its
supposed benefits for teaching
1
rather than fight against the threats. If allowing ChatGPT in
assignments, so the story goes, students and teachers can benefit from collaborative learning,
personalized teaching material and a reduction of the workload for teachers [5, 11].
Sam Altman, the CEO of ChatGPT’s maker Open AI, used similar arguments when
comparing the current situation to the introduction of pocket calculators2:
1https://www.computingatschool.org.uk/resource-library/2023/january/
chatgpt-for- teachers-a-guide-by-evan-dunne
2https://www.yahoo.com/news/ceo-chatgpt- maker-responds-schools-174705479.html
arXiv:2307.03196v1 [cs.CY] 3 Jul 2023

2 ChatGPT is not a pocket calculator
Generative text is something we all need to adapt to [...] We adapted to calculators and
changed what we tested for in math class, I imagine. This is a more extreme version of that,
no doubt, but also the benefits of it are more extreme, as well.
However, though there is no doubt that benefits exist, we believe the threats should be
carefully considered before assuming that the former outweigh the latter, or that adaptation
is the right strategy. In particular, general benefits of language models and related chatbot
services for society as a whole do not automatically translate to benefits for teaching. The
role of a teacher is to 1) teach students knowledge and skills, and, for this purpose, to 2)
assess to what extent this has succeeded. If chatbot services can mimic these skills, then
what is at stake in education is nothing less than the validity of assessments, i.e., the way
teachers make sure students have gained essential knowledge and skills.
The severeness of the threat, therefore, is a function of the quality of answers of the
chatbot to questions in essential areas of knowledge and levels of skills. For this reason, we
argue that finding out the real magnitude of the problem for education requires investigating
the quality of ChatGPT for specific areas of knowledge and skills first. In this paper, we report
on a preliminary assessment on ChatGPT’s quality in teaching Geography and GIScience at
university level.
2 The problem in a nutshell
2.1 GPT-3 and ChatGPT
The recent breakthrough in AI-based chatbots is largely due to the development of a particular
generation of deep neural networks, called transformers which are pretrained on huge amounts
of text. GPT-3 (Generative Pre-trained Transformer) is a transformer model that can produce
texts of a human-like quality. Generative pre-trained means that it has been trained in an
unsupervised and unspecific way on vast amounts of texts to continue texts (words, code or
data) given in some input prompt [
3
]. GPT-3 was released in 2020 and was trained on 45
TB of text data from Wikipedia, book libraries and crawled web texts. The model has a
size of roughly 800 GB and consists of 175 billion network parameters [
3
]. Based on this
model, the chatbot ChatGPT was fine-tuned in several stages, using human feedback as well
as a reward model for reinforcement learning [
17
]. The latter was mainly used to form an
”ethical shell” to avoid that ChatGPT exhibits unethical behaviour [2].
As a result of these developments, the capacity of GPT-3 and ChatGPT to automatically
answer questions, follow instructions, summarize and translate texts, correct grammar and
generate new texts of excellent quality is unprecedented [
6
]. Furthermore, similar transformer
models can not only be used to write prose and poetry, but also to automatically synthesize
computer programs [14, 3], and for other modalities than text.
2.2 Is the use of ChatGPT in assessments fraud?
Students can use AI chatbots to generate suggestions as input for their learning, and teachers
can profit by generating learning or testing material
3
. So if ChatGPT is the next step after
the pocket calculator revolution, eventually, it will become a fundamental cultural technique.
This begs the question as to whether its use should be considered fraud. Academic
dishonesty normally includes plagiarism, fabrication, falsification, misrepresentation, and
3https://theconversation.com/chatgpt-students- could-use-ai-to-cheat-but-its-a-chance-to-rethink-assessment-altogether-198019

S. Scheider, H. Bartholomeus and J. Verstegen 3
misbehaviour [
4
]. ChatGPT does not clearly fall under any of these categories. Yet, it still
leads to inaccurate measurements of skills, thereby cheapening the degrees held by alumni,
misleading employers, and diminishing moral integrity by being unfair to students who do
not use it. To understand why, consider the following definition of fraud in the examination
regulations of Utrecht University:
Fraud and plagiarism are defined as an action or omission on the part of the students which
produces an incorrect representation of their performance as regards their knowledge, skills
and understanding, which may result in an examiner no longer being able to assess the
knowledge or ability of a student properly and fairly [...] (Paragraph 5 of the Education and
Examination Regulations (EER) at Utrecht University, The Netherlands)
While ChatGPT of course can play a constructive role in teaching
4
,its use in student
assessments inevitably falls under this definition, as soon as an examiner cannot distinguish
anymore between student skills and ChatGPT skills, and thus cannot validly assess student
skills according to the learning goals. This is independent of whether we assess learning
goals in open- or closed-book exams or in home writing assignments [
15
]. It applies even
when asking students to be creative, e.g., to criticize ChatGPT answers (as suggested in
5
),
since it can criticize its own answers, too [
17
]. According to the fraud triangle in academia
[
4
], students have an incentive for fraud due to work pressure, see opportunities for fraud
when lacking institutional supervision, and tend to rationalize fraud a-posteriori by denying
their responsibility. All this directly applies to ChatGPT: It releases students from work
pressure (incentive), is difficult to control due to lack of supervision, and fraud can be easily
rationalized, in particular, by the reasons for questioning fraud given above.
2.3 Why learning requires supervision
If using AI chatbots is fraudulent because it threatens the supervision of learners, then, one
might argue, why not give up supervision in teaching? Yet, on closer inspection, this turns
out to be a spurious argument. Despite the importance of autonomy in learning cognitive
constructions [
18
,
7
], learning fundamentally requires guidance [
1
]. Thus to acquire particular
skills, students need instructions and explanations as well as feedback on the quality of their
results [
15
]. Teachers have an essential role to play, not only by providing guidance but also
by interrogating students6.
Because of this, the learning process cannot afford giving up on assessing essential skills
in supervision. One can argue that some skills might lose their relevance when substituted
by AI, while others become more important as cultural techniques. Yet, as a matter of fact,
some skills always need to be learned, in particular, distinguishing valid from invalid answers.
And this is why the pocket calculator analogy breaks down: Mental arithmetic is a skill that
is easily substitutable, yet choosing formulas to solve mathematical problems is not. Mental
arithmetic is subordinate to mathematical problem solving since the latter is needed to judge
whether an arithmetic formula can yield a valid answer in the first place. Thus, giving up on
learning this skill just means to give up on the very concept of validity [
10
]. The question of
substitutability boils down to whether the skills simulated by ChatGPT are (1) essential or
rather subordinate to skills required by our culture to judge about a state of affairs, and (2)
4For example, in order to practice critical analysis of AI generated texts.
5https://www.computingatschool.org.uk/resource-library/2023/january/
chatgpt-for- teachers-a-guide-by-evan-dunne
6
”Since the knowledge gained comes primarily through interrogation of and by others, education is
relational, depending on personal interaction between teacher and student.”[9]

4 ChatGPT is not a pocket calculator
whether its quality is sufficient for substituting these skills. To answer these questions, let us
take a closer look at the concrete skills required in Geography and GIScience.
3 ChatGPT quality in geographic exams and assignments
An OpenAI technical report [
16
] evaluated the performance of GPT-3.5 in publicly available
US exams, mainly Advanced Placement (AP) tests; GPT-3.5 scored worst (0th-7th percentile
of test-takers) in the AP Calculus BC and the best (91th-100th percentile of test-takers) in
the AP Environmental Science. So, at least at college-level ChatGPT may be relatively good
at answering geography questions. We have designed a survey among university teachers in
Geography and GIScience to obtain their opinion on the quality of ChatGPT’s answers to
university exam and assignment questions.
3.1 ChatGPT quality survey
After an introduction to the study, an informed consent and sample-evaluation questions on
age and gender, the participant was asked to select an exam question or assignment they would
use for one of their courses and to pose this question/assignment to ChatGPT. We asked
the participant to fill in the question and the answer, the name of the corresponding course,
the educational level of the course (Bachelor 1, 2, 3, Master, PhD, or other). The answer
of ChatGPT was then to be rated by the participant based on correctness, completeness,
conciseness, and clarity, corresponding to question quality dimensions suggested in [
19
]. The
survey was implemented in Qualtrics XM7.
The participants were university teachers recruited across three departments at two
universities: the Departments of Human Geography and Spatial Planning, and Physical
Geography at Utrecht University, The Netherlands, and the Laboratory of Geoinformation
Science and Remote Sensing at Wageningen University, The Netherlands. The survey was
conducted between January 30th 2023, and April 21st 2023.
To analyse the results of the survey, we first filtered out invalid entries (e.g., missing or
incomplete questions or answers). Next, we classified each question according to Bloom’s
(revised) taxonomy; from low to high level: remember, understand, apply, analyze, evaluate,
create [
12
]. This was done based on the verb in the question. Finally, we analyzed differences
in quality scores (correctness, completeness, conciseness, and clarity) per geography domain,
education level, and Bloom’s taxonomy level.
3.2 Survey results
We had 41 valid survey responses, by 15 females, 26 males, and 0 other genders, with a median
age of 45 years. Questions were distributed as following among the different knowledge
domains (can be multiple per question): GIS and/or Remote Sensing (14), Urban Geography
(9), Physical Geography (7), International Development Studies (6), Ecomonic Geography
(3), Hydrology (3), Spatial Planning (2) and Other (4). In the majority (80%), the answers
of ChatGPT have correctness scores that would allow it to pass the exam or assignment
(score equal to or larger than 5.5 in Dutch exams), see Fig. 1. A bit less,
≈
74% of the
answers are complete within the boundaries of the same threshold, and the same applies to
clarity. The conciseness scores, in contrast, are much lower (47%); many survey participants
7https://www.qualtrics.com/

S. Scheider, H. Bartholomeus and J. Verstegen 5
commented that the answers were overly long, beyond the point, and they would become
tired if all students answered in this way.
In contrast to what one may expect, the quality of the answers does not decrease with
education level (BSc 1, 2, 3, MSc). No clear relation between education level of the question
and quality of the answer appears from our survey.
Figure 1 Distribution of correctness and completeness scores for answers. Red lines show means
and std deviations.
In the levels of Bloom’s taxonomy, a difference in quality appears between the lowest
three and highest three Bloom skill levels (Fig. 2). A corresponding Wilcoxon rank sum
test was significant for both correctness and completeness with p = .005, having similar
distribution shapes and sample sizes but a significant shift in the median between upper
and lower levels. Questions that require students to evaluate,analyze and create tend to be
much less correct as well as complete. Yet, still, the correctness of corresponding answers
tends to range above 5.5, and thus they may pass a test. Correspondingly, some survey
participants commented that they were particularly worried about the capacity of ChatGPT
to write essays and other types of assignments; this falls in the ’evaluate’ or ’create’ category.
Assuming that the higher the skill level, the more essential the skills for learning, these
results raise the serious concern that chatbots may pose a considerable threat to the validity
of assessments in Geography exams and assignments.
Figure 2 Quality scores of answers to questions grouped by Bloom’s taxonomy level.

6 ChatGPT is not a pocket calculator
4 Discussion and conclusion
Our preliminary results indicate that ChatGPT will likely pass Geography and GIScience
exams and assignments in its present form; this is independent of the skill (Bloom’s taxonomy)
as well as the educational level. This means that even assessing only (true) academic skills
does not help, as ChatGPT masters these too. In terms of correctness, the answers to
Geography questions score satisfactorily (able to pass the exam) over different levels of
education as well as different skill levels. The answers score slightly worse in terms of
completeness and clarity, and much worse in terms of conciseness. Herein, note that we rated
only the first answer, while ChatGPT can make an answer more concise if requested (but in
our experience correctness then goes down). Higher Bloom levels are harder to match, but
are on average still sufficiently correct for passing grades given the tested questions. Note
that some questions at these levels could not be tested because they involve interpreting a
provided graph or map, which cannot be pasted in the chat; therefore, our results may be
biased towards higher grades.
Since the skills underlying these higher levels are not substitutable when learning and
thus require supervision, the introduction of ChatGPT is not comparable to the introduction
of the pocket calculator. This means that we are forced to change Geography and GIScience
assessments in one way or another. We can either:
1.
Adapt the learning goals, and thus the skills being assessed, following the way of
thinking ”students do not need to learn the things that ChatGPT can do for them”.
Since especially higher-level skills are essential for learning itself, we argue that this is
not really an option for such skills. Be aware that the line of reasoning to ”embrace the
development and allowing ChatGPT in assignments” means, perhaps unintentionally,
disregarding essential learning goals. What is and is not essential also depends on what
society demands.
2.
Control the assessment environment, for example by preventing access to the
internet during the assessment. This would be a good option for the lower skill levels
like ”remember”, ”understand” and ”apply”. For higher skill levels (especially ”create”)
preventing access to the internet may not a feasible option, as students typically need
access to information for such tasks. Furthermore, another control option is to use
assessment media that are still inaccessible for AI chatbots. For example, asking students
to analyse local geodata, interpret maps or statistical figures may be comparably safe,
because AI models may not have been trained on such tasks and data8.
3.
Control and assess the learning process instead of the obtained skills. For
example, closely supervise the way students arrive at a result and rate their learning
based on observing this process. This might be an option to rate higher skills levels
(”create”, ”evaluate” and ”analyze”).
4.
Forbid the use of chatbots, check compliance with these rules a-posteriori (there is
software providing a likelihood that a text is written by a chatbot
9
), and punish in case
of non-compliance. This applies to all levels but may also be the least effective measure;
detection software is currently not capable of detecting a large part of the AI-generated
text.
Our analysis specifically focused on exams and assignments in Geography teaching at
8
However, this statement clearly has a limited expiry date, given that multi-modal AI models are being
developed that can easily bridge the gap between images, speech and text [8].
9e.g., https://openai-openai- detector.hf.space/

S. Scheider, H. Bartholomeus and J. Verstegen 7
university level; it does not imply that chatbots cannot be beneficial for teaching at all.
ChatGPT offers great assistance in e.g. scripting tasks and may replace browsing through help-
functions or personal assistance, but only when the right question is asked. A student needs
to be able to evaluate the quality of the outcome which requires a thorough understanding of
the content and background knowledge. Thus, there is a place for AI-chatbots in education,
but including it in examination introduces a major risk for the validity of assessments.
The discussion around the use of ChatGPT in university teaching adds a new dimension to
ethical behaviour in AI, science, and education in general, which is also part of the university
learning outcomes. Geography teachers specifically have a responsibility to teach ethics in
combination with spatial thinking in the changing global environment, since failing to do
so would ”deprive incoming generations of leaders and bystanders of capabilities essential
for navigating uncertain environmental futures” [
13
]. Guidelines, such as given in Figure
1 of UNESCO’s Quick start guide [
20
], which is a flow diagram asking the reader whether
quality of the output matters, they have expertise on the topic, and are willing to take full
responsibility of inaccuracies, may help teachers decide which of the actions above is most
applicable to the course or assignment at hand, and may at the same time raise awareness
with students what they may be sacrificing when using a chatbot. In the end, ”ideas inform
ethics, ethics inform policies, and policies make places” [
13
], and this flow of action requires
people to be able to create, evaluate, and analyze independently.
References
1
Gert Biesta. Risking ourselves in education: Qualification, socialization, and subjectification
revisited. Educational Theory, 70(1):89–104, 2020.
2
Jan Broersen. Hoe dom is kunstmatige intelligentie? is chatgpt een gevaar voor onze
samenleving? (english: How stupid is artificial intelligence? is chatgpt a threat for our
society?), 2023. talk in Stududium Generale, 22 Feb 2023, Utrecht, The Netherlands.
3
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla
Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language
models are few-shot learners. Advances in neural information processing systems, 33:1877–1901,
2020.
4
Debra D Burke and Kenneth J Sanney. Applying the fraud triangle to higher education:
Ethical implications. J. Legal Stud. Educ., 35:5, 2018.
5
Ramon Dijkstra, Zülküf Genç, Subhradeep Kayal, and Jaap Kamps. Reading comprehension
quiz generation using generative pre-trained transformers. 2022.
6
Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences.
Minds and Machines, 30:681–694, 2020.
7
E. v. Glasersfeld. A constructivist approach to teaching. In L. P. Steffe and J. Gale, editors,
Constructivism in education, pages 3–15. 1995.
8
Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu,
Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for
dialogue with humans. arXiv preprint arXiv:2305.04790, 2023.
9
Paul W Gooch. Course Correction: A Map for the Distracted University. University of Toronto
Press, 2019.
10
Peter Janich. Logisch-pragmatische Propädeutik: ein Grundkurs im philosophischen Re-
flektieren. Velbrück Wiss., 2001.
11
Enkelejda Kasneci, Kathrin Seßler, Stefan Küchemann, Maria Bannert, Daryna Dementieva,
Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, et al. Chatgpt
for good? on opportunities and challenges of large language models for education. Learning
and Individual Differences, 103:102274, 2023.

8 ChatGPT is not a pocket calculator
12
David R. Krathwohl. A revision of bloom’s taxonomy: An overview. Theory Into Practice,
41(4):212–218, 2002. doi:10.1207/s15430421tip4104\_2.
13
Thomas Barclay Larsen and John Harrington Jr. Implications of the anthropocene for
professional ethics in american geography education. In Geography Teacher Education and
Professionalization, pages 245–261. Springer, 2022.
14
Kevin Lu, Aditya Grover, Pieter Abbeel, and Igor Mordatch. Pretrained transformers as
universal computation engines. arXiv preprint arXiv:2103.05247, 1, 2021.
15
Wilbert McKeachie and Marilla Svinicki. McKeachie’s teaching tips. Cengage Learning, 2013.
16
OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774v2. URL:
https://cdn.
openai.com/papers/gpt-4.pdf.
17
Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin,
Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models
to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022.
18
Virginia Richardson. Constructivist pedagogy. Teachers college record, 105(9):1623–1640,
2003.
19
Chirag Shah and Jefferey Pomerantz. Evaluating and predicting answer quality in com-
munity qa. In Proceedings of the 33rd international ACM SIGIR conference on Research and
development in information retrieval, pages 411–418, 2010.
20
UNESCO. Chatgpt and artificial intelligence in higher education - quick start guide.
Technical report, 2023. URL:
https://www.iesalc.unesco.org/wp-content/uploads/2023/
04/ChatGPT-and-Artificial-Intelligence-in-higher-education-Quick-Start-guide_
EN_FINAL.pdf.