Essential questions in earth and geosciences according to large language models

István Gábor Hatvani; Manfred Mudelsee; Zoltán Kern

doi:10.1515/geo-2022-0677

Article Open Access

Essential questions in earth and geosciences according to large language models

István Gábor Hatvani , Manfred Mudelsee and Zoltán Kern

Published/Copyright: August 16, 2024

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Open Geosciences Volume 16 Issue 1

Abstract

Can large language models (LLMs) capture a trustworthy global snapshot of the current issues and top-priority questions facing Earth and Geoscience? This article discusses the 100 most important questions facing Earth and Geosciences in the twenty-first century according to the largest of the LLMs. The study uncovered a discrepancy in responses using the synonymous terms earth sciences and geosciences; as such, users of publicly accessible LLMs must account for this bias and exercise caution in interpretation. Eight fundamental topics irrespective of the query terms earth sciences/geosciences were found: while two of them originate from fundamental research in extraterrestrial geoscience and Earth’s interior, the remaining six address geoscientific questions, important due to the associated societal challenges and environmental issues. The outlined eight fundamental topics strongly resonate with eight Sustainable Development Goals of the United Nations, in certain cases with not just one.

Graphical abstract

Keywords: generative pre-trained transformer; Earth and Geoscience inquiries; LLM; text analysis

1 Introduction

At the overture of the twentieth century, Hilbert presented the top priority problems for mathematics [1], inspiring many scientists to embark on groundbreaking research for more than a century; some of the questions remain unsolved [2]. Defining top-priority questions in a scientific discipline can act as a compass, (i) guiding researchers toward the most critical issues within a scientific discipline and (ii) pinpointing the current state of the discipline and where its challenges lie. This ensures that research efforts are focused on topics that have the potential to make significant contributions to the field. When the scientific community focuses on key questions, it encourages collective efforts to address these challenges, leading to a more cohesive and impactful research community. This task of seeking out the crucial questions of a discipline can be achieved through (i) traditional means, where an established and renowned professor [1,3,4] or group of scholars conduct comprehensive literature reviews [5]; (ii) employing large-scale questionnaires [6]; or (iii) utilizing the contemporary approach of online surveys; or a desired combination of the three [7,8].

Accurate, thorough literature review is a complex and time-consuming exercise [9,10], which – despite the best intentions and experience – can be prone to bias. Providing a comprehensive, balanced, and engaging overview is therefore difficult [10]. The large-scale availability of large language models (LLMs) can simplify and speed up such tasks [11,12,13,14], but at what cost?

The number of research fields impacted by artificial intelligence (AI) is ever increasing [15,16,17,18], raising concern in science in general [17,19], including earth- and geoscientists [20,21] and decision makers alike [22]. One of the AI tools with the most widespread impact are LLMs, probably due to their accessibility and the fact that anyone using the internet speaks at least one language, while specific AI tools require the knowledge of scripting languages. The first openly available, most widely adopted and user-friendly LLM was the OpenAI’s Generative Pre-trained Transformer 3 (GPT-3) released in June 2020, followed by Bard powered by the PaLM 2 model from Google AI, made available to the public in over 180 countries and territories in May 2023.

The usefulness of such tools available to the public is no question to a fair extent [17]; however, numerous concerning issues have been raised in the short time of their global presence already [15,22,23], such as bias in the answers given, privacy and security through reinforcement learning input from users, transparency in its development (training data used), plainly wrong answers given [24], and academic integrity concerns [25]. “It is no longer possible to accurately distinguish text written by a human mind from that generated by a highly parallelizable artificial neural network with substantially fewer neural connections” [16].

These issues are, of course, not black-or-white and almost all of these are subject of debate [26]. For example, the use of LLMs may level the playing field regarding the presentation of a scientific paper between native and non-native scientists [27,28]. Many disciplines have addressed the practical benefits, ethical issues, and potential risks [14,28,29,30]. We assume that general concerns and issues are presented sufficiently in literature for the moment and the next step is to actually see how these turn out in the short/mid-term.

As shown above, asking top-priority questions for a given discipline is labor intensive, which can be eased by exploiting the potential in LLMs [12,14,30], but can it really? This article tests how LLMs can be used to expedite the process of finding the paramount questions by generating 100 questions important for research in Earth and Geosciences and assessing the validity of the outcome. In the absence of AI-driven tools, achieving such a set of answers (100 questions) would require gathering perspectives from hundreds of individuals worldwide and several months of effort to attain the intended goal [30]. It is no question that one can expect to gather the “100 most important” questions in Earth and Geosciences with LLMs fast and easy, but it is definitely necessary to critically assess how repetitive, vague, and meaningless the answers are, even considering if “earth” or “geosciences” are the search term.

2 Materials and methods

2.1 The derivation of 100 most important questions for Earth and Geosciences using LLMs

The definition of Earth Sciences [31] and Geosciences [32] point strictly in the same direction and in many cases are used interchangeably [31,33]. Geoscience is a multidisciplinary field that encompasses a broad spectrum of subjects and deals with the mechanisms responsible for molding the Earth’s surface, investigates the utilization of natural resources, and explores the intricate interplay between water and ecosystems [33].

Two of the most popular LLMs (ChatGPT [GPT-3.5, version 25 September, accessed on 27 September 2023] and Bard) were asked the following questions on 27 September 2023:

Can you please give me the 100 most important questions geosciences face in the 21st century?

Can you please give me the 100 most important questions earth sciences face in the 21st century?

While ChatGPT provided the 100 questions for both Earth and Geosciences, Bard gave 35 and 34 answers, respectively. When asked, Please give me no less than a list of a 100 most important… it irritatingly still provided only 40 points. Therefore, in this research, the list of questions provided by ChatGPT was in the focus, which, of course, does not mean that the 100 most important questions given by ChatGPT are accurate (Table S1).

2.2 Methodology

Next step was to investigate if Bard or ChatGPT is applicable in assisting in the text mining tasks of the study. First, the task was to tokenize the text with the two LLMs. Bard was asked to complete the task with the list of questions for Geosciences (G-list) (Table S1a) provided by ChatGPT, due to reasons discussed previously. Unfortunately, when asked Please remove the numbers, punctuation and extra white spaces from the list above and the common English stopwords according to the “stopwords-iso” list only a fraction of the 100 answers were processed (67) by ChatGPT and stopwords such as what, can, do, how, we, etc., were left in the text, while others were removed. Thus, it understood the task, but did not complete it. On the other hand, Bard was only able to accept 84 of the questions due to input limitations and left stopwords such as on, and, of, etc., in the text. Moreover, when asked Please list the stopwords you removed from the original list both LLMs listed stopwords as being removed although (i) these were not removed or (ii) were not even included in the original text, rendering the use of ChatGPT and Bard unreliable in the text analysis process. Similar patterns were seen for the list of answers (Table S1). In the next trial, neither Bard nor ChatGPT was able to count the proper occurrence of words in the G-list. Still, it should be noted that specific tools are (being) developed to exploit the opportunities in LLMs for text mining, for example, BioBERT, which is a domain-specific language representation model pre-trained on large-scale biomedical corpora [34].

Besides the empirical analysis of the answers provided by ChatGPT (Table S1) basic text analysis was done in R [35] with the following steps. With the tm_map() function of the tm package [36], both lists were converted to a corpus object, with lowercase characters, numbers, punctuation, and extra white spaces removed as well as the common English stopwords according to the stopwords-iso list, which is a comprehensive collection of stopwords for multiple languages following the ISO 639-1 language code [37]. Afterwards, wordclouds were produced using the wordcloud() function of the wordcloud2 package [38]; subsequently, it was stemmed using Porter’s stemming algorithm [39] for further analysis. Ultimately, the stemmed text was analyzed for the most frequent 15 words and presented on barplots.

3 Results and discussion

3.1 Earth Sciences vs Geosciences: Does the query term matter?

We empirically and objectively assessed (i) the set of answers given depending on the topic of interest being Earth or Geosciences to determine if there are fundamental differences in the questions with regard to scientific content, self-repetition, although Earth and Geosciences are interchangeable terms. Then, (ii) the two sets of answers were compared to see if the same questions occur regardless of the chosen term (earth sciences or geosciences) used to formulate the query.

Empirically analyzing the two sets of questions in the G-list and the ES-list separately, different numbers of paraphrased or topically repetitive questions were found. In the G-list, there were 54 questions that were not repeated in any form in the set, while in the case of the ES-list there were 59 ones. Examples of repetitive questions: question G61 was extended by a single word bluffs in question G79:

G61. What are the potential impacts of geological processes on the stability and behavior of coastal cliffs?

G79. What are the potential impacts of geological processes on the stability and behavior of coastal bluffs and cliffs? in the ES-list.

In the G-list, question ES76 is an amalgamation of questions ES7 and ES100; moreover, these refer to the same topic as discussed in question no. ES37. This is an example of both paraphrasing and topical repetition. Specifically:

ES7. How can we accurately model and predict changes in Earth’s atmospheric composition and its impact on climate and human health?

ES37. What are the interactions between climate change, air quality, and human health, and how can we develop policies to mitigate risks?

ES76. How can we enhance the understanding of aerosols and their role in climate, air quality, and human health?

ES100. How can we improve our understanding of the role of natural and anthropogenic aerosols in climate change and atmospheric processes?

It is also worth mentioning that there was a fundamental difference regarding the scientific language, word use, and general complexity of the questions. On the one hand, in the G-list, a high number of termini technici occur such as fault zone, subduction zone, etc. (Table S1a). This is reflected in the word cloud as well, with geological and processes being the most frequent words (count above 50; Figure 1a and b). On the other hand, the ES-list (Table S1b) seems more holistic considering environmental issues related to the atmosphere (climate), biosphere (ecosystems), hydrosphere and, of course, the lithosphere.

Figure 1

Word clouds illustrating the occurrence of words appearing in ChatGPT-generated 100 important questions Geosciences (a) and Earth Sciences (c) are facing (Table S1) with the barplot illustrating the word frequency of the 15 most abundant stem words for Geosciences (b) and Earth Sciences (d).

The numerical assessment of the wordclouds again highlights the differences between the set of answers (Figure 1). Obviously, geological and processes are associated together and used as a subject very frequently in the composed G-list (Figure S1). If these two most recurrent words are omitted, four common words are found with a similarly decreasing frequency: impact, potenti*, understand, and mitig* (Figure 1b and d) without any specific disciplinary relevance.

Despite Earth Sciences and Geosciences referring to the same discipline, a large portion of the frequent words in each list reveals a noticeable bias. While in the G-list, general terms such as influenc*, distribut*, occur*, etc., dominate, in the ES-list, words such as climat*, ecosystem, sustain*, etc., dominate (Figure 1), concurring with the impressions gained from the list of questions (Table S1). This bias stems from the terminological differences found in ChatGPT’s training corpus mainly obtained from Common Crawl data up to 2021. Over the past few decades, we have observed a significant integration of the terms Earth- and environmental sciences evident in journal titles, the establishment of relevant university departments, and the emergence of specialized graduate programs, etc. Illustrating the interconnectivity of the two words, environment and Earth Sciences, in the case of the ES-list, the word environment and its stems occur almost three times more than in the case of the G-list (11 vs 4). Additionally, the five most frequent co-occurrences in the ES-list suggest a more holistic view of Earth Sciences with emphasis on pressing environmental issues as climate change (Figure S1c) rather than the G-list’s co-occurrences (Figure S1b).

3.2 Top-priority questions for Earth and Geosciences: A “machine” viewpoint

The 14 pairs of associated questions that appeared in both lists cover eight fundamental topics important within the explored discipline, irrespective of the query terms earth sciences/geosciences. Notably, two of these topics originate from fundamental research: deeper understanding of Earth’s interior (Q1 and Q2) and extraterrestrial geoscience (Q3) (Table 1). The majority of the question pairs pertain to aspects that address the functioning of the environment and society from a human perspective. One question pair primarily stems from societal concerns, focusing on the assessment of the potential consequences related to urbanization and population growth on the environment, infrastructure, natural resources, and landscapes (Q4). A fair portion of the shared topics relate to the sustainable exploitation of natural resources, namely, subsurface water resources (Q5 and Q6), geothermal energy (Q7), and mineral and metallic resources (Q8 and Q9; Table 1). Both lists exhibit strong representation in the areas of predicting and mitigating natural and geological hazards (Q10–Q12) and addressing global change (Q13 and Q14; Table 1). These global change-related questions are the only ones that resonate with the ten priority questions for geosciences coined in the early 2000s [3]. On the one hand, this is surprising, as one might expect a larger overlap; on the other hand, the problems are rather climate-related − as seen two decades later [40] − lacking the holistic view offered in the ES-list. The global change-related challenges (Table 1 and Table S1) have become a central focus within Earth Sciences and Geosciences alike, where the boundaries between “hard-core geoscientific questions” and pressing environmental issues blur as they increasingly intersect in the context of the twenty-first century [41].

Table 1

Associated question pairs obtained through querying an LLM using the query terms earth sciences/geosciences

No.	Geosciences	Earth Sciences	Topic
Q1	What is the origin and evolution of Earth’s magnetic field, and how does it influence our planet’s dynamics?	What are the fundamental processes driving Earth’s magnetic field and its interactions with the Earth’s interior?	Better understand Earth’s structure
Q2	How can we improve our understanding of the deep Earth’s structure and composition?	How can we improve our understanding of deep Earth processes, including mantle dynamics and core behavior?	Better understand Earth’s structure
Q3	How can we improve our understanding of the geology and processes on other planets and moons in our solar system?	How can we enhance our understanding of planetary geology and processes to inform potential colonization of other celestial bodies?	Extraterrestrial geoscience
Q4	What are the potential consequences of rapid urbanization and infrastructure development on geological systems and landscapes?	What are the potential impacts of rapid urbanization and population growth on the environment, infrastructure, and natural resources?	Urbanization
Q5	How can we accurately predict and mitigate the impacts of geological processes on groundwater quality and contamination?	How can we accurately model and predict the behavior of groundwater systems to ensure sustainable management and availability?	Groundwater
Q6	How can we develop sustainable strategies to manage groundwater resources and prevent over-extraction?	How can we develop effective strategies to manage and protect groundwater resources from contamination and depletion?	Groundwater
Q7	How can we develop sustainable strategies for managing and utilizing geothermal energy?	How can we develop sustainable strategies to manage and utilize geothermal energy for power generation and heating?	Geothermal energy
Q8	How can we develop sustainable strategies for the responsible mining and extraction of minerals?	How can we develop sustainable strategies for managing natural resources, including minerals, metals, and fossil fuels?	Sustainable resource extraction
Q9	What are the geological factors influencing the occurrence and distribution of rare earth elements and their importance in technology?	What are the potential impacts of mining and extracting rare earth elements on the environment and society?	Sustainable resource extraction
Q10	How can we predict and mitigate the impacts of earthquakes and tsunamis on populated areas?	How can we predict and mitigate the impacts of natural disasters such as earthquakes, tsunamis, hurricanes, and volcanic eruptions?	Predict and mitigate the impacts of geological hazards/natural disasters
Q11	How can we accurately predict and mitigate landslides and slope failures to protect communities and infrastructure?	How can we accurately predict and mitigate the impacts of landslides and slope failures in various geological settings?
Q12	What are the potential impacts of geological hazards on critical infrastructure, and how can we enhance their resilience?	How do interactions between human activities and natural processes affect the vulnerability and resilience of communities to geological hazards?
Q13	What are the mechanisms and potential impacts of ocean acidification on marine ecosystems?	What are the potential consequences of ocean warming and acidification on marine biodiversity and ecosystems?	Global change
Q14	How can we accurately date and understand past climatic and environmental changes to inform future predictions?	How can we accurately date and understand past climate changes to improve predictions for future climate scenarios?	Global change

The United Nations’ Sustainable Development Goals (SDGs) [42] encompass areas such as poverty alleviation, eradication of unsustainable consumption patterns, sustained and inclusive growth, social development, and environmental protection to create a more equitable and sustainable world, all of which can be related to different geoscience sub-disciplines [43]. Most of the eight fundamental topics (Table 1) outlined in the present study, irrespective of the query terms earth sciences/geosciences were related to eight out of the SDGs. Three topics are related to two goals each: Urbanization is strongly related to SDG 11: “Sustainable Cities and Communities” and 15: “Life on Land,” while the topic: Predict and mitigate the impacts of geological hazards/natural disasters is also related to SDG11, but most importantly to SDG 9: “Industry, Innovation, and Infrastructure.” The topic themed Groundwater (Table 1) is clearly mirroring SDG 6: “Clean water and Sanitation,” and SDG 12: “Responsible Consumption and Production.” Latter (SDG 12) uniquely resonates with the Sustainable resource extraction topic (Table 1). Lastly, the questions under the topic Global change are in very close relation to the “Climate Action” SDG (No. 13) and Q13 (Table 1) is also related to SDG 14: “Life Below Water.”

About half of the SDGs are very well represented in the eight fundamental topics of highly important questions for Earth and Geosciences for the twenty-first century, underlining how well established and important these geoscientific goals are for the sustainability of human society and the protection of the planet and its ecosystem.

4 Conclusions and outlook

Although the terms Earth Sciences and Geosciences are synonyms and are considered interchangeable in academic English, still, when used in queries with LLMs fundamentally varying outputs are obtained. The analysis returned only 14 shared questions between the two lists encompassing eight broader topics: (i) extraterrestrial geoscience and (ii) Earth’s interior encloses of fundamental research origin, while the others address geoscientific questions, which became important in the twenty-first century due to the associated societal challenges and environmental issues: (iii) potential consequences related to urbanization and population growth, (iv) subsurface water resource management, (v) sustainable geothermal energy exploitation, (vi) sustainable mineral and metallic resources extraction, (vii) the prediction and mitigation of natural and geological hazards, and (viii) addressing global change. These eight fundamental topics align with eight of the SDGs of the United Nations.

Following the revision stage of this paper, on 13 May 2024, ChatGPT-4o [44] was released. Although a detailed assessment was not possible, a quick test indicated similar impressions to those discussed in the paper. Depending on the query term used ‘Earth Sciences’ or ‘Geosciences’ the output indicated characteristically similar differences as described above. The query term ‘Earth Sciences’ tended to provide a list of questions addressing more holistic environmental issues, while ‘Geosciences’ query term gave a list of questions dealing more frequently with the Earth’s surface and interior.

Since LLMs are primarily trained on the internet, they inherit the biases embedded in their training data. Thus, this discrepancy reflects a growing trend associating environmental sciences with Earth Sciences. Consequently, it is essential for users of publicly accessible LLMs to recognize and account for this bias when using these tools for tasks and exercise caution in their interpretations. Moreover, because of the ever-growing and unsupervised use of LLMs, the enduring influence of this observed bias is likely, unless specific measures are taken. This may involve (i) training LLMs to distinguish between Earth Sciences, Geosciences, and general environmental science-related queries and/or (ii) refining their ability to process domain-specific data. A question arises, should a distinction be made (however “small”) between Earth Sciences and Geosciences? Publicly available information seems to suggest that these terms indeed reference diverse sets of phenomena.

GPTs specifically trained on discipline-oriented questions are also in the pipeline, even for Geosciences. GeoGPT was made available for testing in early 2024 and raised great concern among geoscientists [21]. It inherited the drawbacks of general LLMs – discussed in this paper above – together with other issues such as state censorship, non-disclosure of the data (texts) used, and copyright infringement [20]. In addition, we suggest Deep-time Digital Earth (the developer) to pursue the question raised in the present study prior to making GeoGPT available for the public. Our view is, field-specific GPTs may be even more intriguing to be used for researchers of a given discipline, potentially increasing their confidence in the technology. However, this increased trust may paradoxically lower their vigilance, leading to a reduced scrutiny of common issues associated with LLM outputs.

In the realm of academic publishing, varying stances exist regarding the use of LLMs. While some journals, such as Nature, do not prohibit the use of LLMs to support the improvement of shoddily written text [16,45], others, such as Science, prohibit their use, since they cannot be accounted for the answers given, thus cannot be authors [46]. Researchers are advised to exercise vigilance when relying on information from ChatGPT and other LLMs and to cross-verify responses to ensure their accuracy [14].

Acknowledgements

This work was supported the European Union’s Horizon Europe Research and Innovation Programme, GreenSCENT project (Grant Agreement No 101036480) (MM).

Author contributions: István Gábor Hatvani: conceptualization; investigation; methodology; visualization; writing – original draft and review & editing. Zoltán Kern: conceptualization; investigation; visualization; writing – original draft and review & editing Manfred Mudelsee: conceptualization; writing – original draft and review & editing, funding acquisition. The FLAE approach was employed in the sequence of authors; see 10.1371/journal.pbio.0050018.
Conflict of interest: Authors state no conflict of interest.
Declaration of generative AI and AI-assisted technologies in the writing process: During the preparation of this work, the author(s) have not used generative AI tools for writing the text.

References

[1] Hilbert D. Mathematische Probleme. Nachrichten von der Königlichen Gesellschaft der Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse. 1900;3:253–97.Search in Google Scholar

[2] Yandell B. The honors class: Hilbert’s problems and their solvers. New York, USA: CRC Press; 2001.10.1201/9781439864227Search in Google Scholar

[3] Ghil M. Hilbert problems for the geosciences in the 21st century. Nonlinear Process Geophys. 2001;8:211.10.5194/npg-8-211-2001Search in Google Scholar

[4] May R. Unanswered questions in ecology. Philos Trans R Soc Lond Ser B: Biol Sci. 1999;354:1951–9.10.1098/rstb.1999.0534Search in Google Scholar PubMed PubMed Central

[5] Bousquet J, Fokkens W, Burney P, Durham SR, Bachert C, Akdis CA, et al. Important research questions in allergy and related diseases: nonallergic rhinitis: a GA2LEN paper. Allergy. 2008;63:842–53.10.1111/j.1398-9995.2008.01715.xSearch in Google Scholar PubMed

[6] Sutherland WJ, Freckleton RP, Godfray HCJ, Beissinger SR, Benton T, Cameron DD, et al. Identification of 100 fundamental ecological questions. J Ecol. 2013;101:58–67.10.1111/1365-2745.12025Search in Google Scholar

[7] Devenish AJM, Schmitter P, Jellason NP, Esmail N, Abdi NM, Adanu SK, et al. One hundred priority questions for the development of sustainable food systems in Sub-Saharan Africa. Land. 2023;12:1879.10.3390/land12101879Search in Google Scholar

[8] Blöschl G, Bierkens MFP, Chambel A, Cudennec C, Destouni G, Fiori A, et al. Twenty-three unsolved problems in hydrology (UPH) – a community perspective. Hydrol Sci J. 2019;64:1141–58.10.1080/02626667.2019.1620507Search in Google Scholar

[9] Chan WCW. Writing excellent review articles. ACS Nano. 2023;17:1723–4.10.1021/acsnano.3c00497Search in Google Scholar PubMed

[10] Dhillon P. How to write a good scientific review article. FEBS J. 2022;289:3592–602.10.1111/febs.16565Search in Google Scholar PubMed

[11] Lavrič F, Škraba A. Brainstorming will never be the same again - A human group supported by artificial intelligence. Mach Learn Knowl Extr. 2023;5:1282–301.10.3390/make5040065Search in Google Scholar

[12] Agathokleous E, Rillig MC, Peñuelas J, Yu Z. One hundred important questions facing plant science derived using a large language model. Trends Plant Sci. 2023;29(2):210–8. 10.1016/j.tplants.2023.06.008.Search in Google Scholar PubMed

[13] Zheng Y, Koh HY, Ju J, Nguyen ATN, May LT, Webb GI, et al. Large language models for scientific synthesis, inference and explanation. arXiv preprint arXiv:231007984; 2023. Accessed on 14.12.2023.Search in Google Scholar

[14] Zhu J-J, Jiang J, Yang M, Ren ZJ. ChatGPT and environmental research. Environ Sci & Technol. 2023;57(46):17667–70. 10.1021/acs.est.3c01818.Search in Google Scholar PubMed PubMed Central

[15] Stokel-Walker C, Van Noorden R. What ChatGPT and generative AI mean for science. Nature. 2023;614:214–6.10.1038/d41586-023-00340-6Search in Google Scholar PubMed

[16] Nature Editorial. AI will transform science — now researchers must tame it. Nature. 2023;621:658.10.1038/d41586-023-02988-6Search in Google Scholar PubMed

[17] Van Noorden R, Perkel JM. AI and science: what 1,600 researchers think. Nature. 2023;621:672–5.10.1038/d41586-023-02980-0Search in Google Scholar PubMed

[18] Yang ZG, Laki LJ, Váradi T, Prószéky G. Mono- and Multilingual GPT-3 Models for Hungarian. In: Text, Speech, and Dialogue: 26th International Conference, TSD 2023, Pilsen, Czech Republic, September 4–6, 2023, Proceedings. Lecture Notes in Computer Science, 14102. Cham: Springer Nature Switzerland; 2023. p. 94–104.10.1007/978-3-031-40498-6_9Search in Google Scholar

[19] Stokel-Walker C. ChatGPT listed as author on research papers: Many scientists disapprove. Nature. 2023;613:620–1.10.1038/d41586-023-00107-zSearch in Google Scholar PubMed

[20] Cleverley PH. Geoscience AI in crisis? GEOSCIENTIST 2024. https://geoscientist.online/sections/viewpoint/geoscience-ai-in-crisis/ (accessed on 17.06.2024).Search in Google Scholar

[21] Voosen P. Chinese AI stirs panic at European geoscience society. Science. 2024. https://www.science.org/content/article/chinese-ai-stirs-panic-european-geoscience-society (accessed on 17.06.2024).Search in Google Scholar

[22] Zhou J, Müller H, Holzinger A, Chen F. Ethical ChatGPT: Concerns, challenges, and commandments. arXiv 2023; preprint arXiv:150804409. 10.48550/arXiv.2305.10646. Accessed on 14.12.2023.Search in Google Scholar

[23] Liebrenz M, Schleifer R, Buadze A, Bhugra D, Smith A. Generating scholarly content with ChatGPT: ethical challenges for medical publishing. Lancet Digital Health. 2023;5:e105–6.10.1016/S2589-7500(23)00019-5Search in Google Scholar PubMed

[24] StackOverflow, 2023. “Temporary policy: ChatGPT is banned,” https://stackoverflow.com/help/ai-policy. Accessed on 14.12.2023.Search in Google Scholar

[25] Crawford J, Cowling M, Ashton-Hay S, Kelder J-A, Middleton R, Wilson GS. Artificial intelligence and authorship editor policy: ChatGPT, Bard Bing AI, and beyond. J Univ Teach Learn Pract. 2023;20:1.10.53761/1.20.5.01Search in Google Scholar

[26] Rillig MC, Ågerstrand M, Bi M, Gould KA, Sauerland U. Risks and benefits of large language models for the environment. Environ Sci Technol. 2023;57:3464–6.10.1021/acs.est.3c01106Search in Google Scholar PubMed

[27] Berdejo-Espinola V, Amano T. AI tools can improve equity in science. Science. 2023;379:991.10.1126/science.adg9714Search in Google Scholar PubMed

[28] Foroumandi E, Moradkhani H, Sanchez-Vila X, Singha K, Castelletti A, Destouni G. ChatGPT in hydrology and earth sciences: Opportunities, prospects, and concerns. Water Resour Res. 2023;59:e2023WR036288.10.1029/2023WR036288Search in Google Scholar

[29] Zheng Z, Zhang O, Borgs C, Chayes JT, Yaghi OM. ChatGPT chemistry assistant for text mining and the prediction of MOF synthesis. J Am Chem Soc. 2023;145:18048–62.10.1021/jacs.3c05819Search in Google Scholar PubMed PubMed Central

[30] Armstrong EM, Larson ER, Harper H, Webb CR, Dohleman F, Araya Y, et al. One hundred important questions facing plant science: an international perspective. N Phytol. 2023;238:470–81.10.1111/nph.18771Search in Google Scholar

[31] Orion N, Julie CL. Earth science education. In: Norman GL, Dana LZ, Judith SL, editors. Handbook of research on science education. New York, NY, USA: Taylor & Francis, Routledge; 2023. p. 692–716.10.4324/9780367855758-26Search in Google Scholar

[32] National Research Council (NRC). Discipline-based education research: Understanding and improving learning in undergraduate science and engineering. Washington DC, USA: The National Academies Press; 2012.Search in Google Scholar

[33] USGS, 2023. What is geoscience? United States Geological Survey. https://www.usgs.gov/youth-and-education-in-science/what-geoscience. Acessed on 14.12.2023Search in Google Scholar

[34] Lee J, Yoon W, Kim S, Kim D, Kim S, So CH, et al. BioBERT: A pre-trained biomedical language representation model for biomedical text mining. Bioinformatics. 2019;36:1234–420.10.1093/bioinformatics/btz682Search in Google Scholar PubMed PubMed Central

[35] R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing; 2023.Search in Google Scholar

[36] Feinerer I, Hornik K, Meyer D. Text mining infrastructure in R. J Stat Softw. 2008;25:1–54.10.18637/jss.v025.i05Search in Google Scholar

[37] Diaz G, Suriyawongkul A, 2016. Stopwords iso. https://github. com/stopwords-iso. (accessed on 14.12.2023).Search in Google Scholar

[38] Lang D wordcloud2: Create Word Cloud by html Widget. R package version 0.2.2. 2023.Search in Google Scholar

[39] Porter Martin F. An algorithm for suffix stripping. Program. 1980;14:130–7.10.1108/eb046814Search in Google Scholar

[40] Ghil M. Review article: Hilbert problems for the climate sciences in the 21st century – 20 years later. Nonlin Process Geophys. 2020;27:429–51.10.5194/npg-27-429-2020Search in Google Scholar

[41] van Wyk de Vries B, Byrne PK, Delcamp A, Einarson P, Göğüş O, Guilbaud M-N, et al. A global framework for the Earth: putting geological sciences in context. Glob Planet Change. 2017;171:293–321.10.1016/j.gloplacha.2017.12.019Search in Google Scholar

[42] United Nations. Transforming our world: the 2030 agenda for sustainable development. Geneva, Switzerland 2015. https://sdgs.un.org/publications/transforming-our-world-2030-agenda-sustainable-development-17981. Acessed on 14.12.2023.Search in Google Scholar

[43] Gill JC. Geology and the sustainable development goals. Episodes. 2017;40:70–6.10.18814/epiiugs/2017/v40i1/017010Search in Google Scholar

[44] Achiam J, Adler S, Agarwal S, Ahmad L, Akkaya I, Aleman FL, et al. Gpt-4 technical report. arXiv 2023:2303.08774.Search in Google Scholar

[45] Nature Editorial. Adapting to AI. Nature Geosci. 2024;17;481.10.1038/s41561-024-01475-5Search in Google Scholar

[46] Thorp HH. ChatGPT is fun, but not an author. Science. 2023;379:313.10.1126/science.adg7879Search in Google Scholar PubMed

Received: 2024-03-27

Revised: 2024-05-29

Accepted: 2024-07-16

Published Online: 2024-08-16

This work is licensed under the Creative Commons Attribution 4.0 International License.

Supplementary material

Articles in the same Issue

https://doi.org/10.1515/geo-2022-0677

Keywords for this article

generative pre-trained transformer; Earth and Geoscience inquiries; LLM; text analysis

Creative Commons

BY 4.0

Essential questions in earth and geosciences according to large language models

Article

Abstract

Graphical abstract

1 Introduction

2 Materials and methods

2.1 The derivation of 100 most important questions for Earth and Geosciences using LLMs

2.2 Methodology

3 Results and discussion

3.1 Earth Sciences vs Geosciences: Does the query term matter?

3.2 Top-priority questions for Earth and Geosciences: A “machine” viewpoint

4 Conclusions and outlook

Acknowledgements

References

Supplementary Material

Articles in the same Issue

Articles in the same Issue

Articles in the same Issue