Impact Statement

With the recent release of ChatGPT and similar tools based on large language models, there is considerable enthusiasm and substantial concern over how these tools should be used. We pose a series of questions aimed at unpacking important considerations in the responsible use of large language models within environmental data science.

2. How do LLMs work?

LLMs are a family of machine learning models designed to intake text prompts and generate contextual text outputs. They are extremely large deep neural networks with upwards of hundreds of billions of parameters, also known as weights. LLMs are trained in a “self-supervised” manner, sometimes referred to as “autoregressive,” using an architecture called a transformer that relies on a surprisingly simple process called attention (Vaswani et al., Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017). These models are fed text with some words omitted and are trained to essentially “impute” the missing text. Training text comes from an incredibly large corpora of text often scraped from the internet (i.e., Reddit, Wikipedia, and the Internet Archive). A training example might be the sentence “I’m going to walk the.….” The model would predict the end of the sentence with the goal of maximizing the likelihood that the characters it predicts are the characters omitted in the original text. Assuming the original sentence was “I’m going to walk the dog,” the model’s parameters would be updated to maximize the likelihood that the three characters it predicted were indeed “dog.” By repeating this omission-based training process billions of times, models like ChatGPT “learn” to predict responses to a wide range of inputs ranging from patents to historical summaries (Brown et al., Reference Brown, Meier, Morris, Pastor-Guzman, Bai, Lerebourg, Gobron, Lanconelli, Clerici and Dash2020).

In principle, this approach can be used on any text-based data format (e.g., DNA sequences) (Gankin et al., Reference Gankin, Karollus, Grosshauser, Klemon, Hingerl and Gagneur2023), but human prose text is the most common. Of specific interest to the EDS community, these models have shown surprising capability in translating between computer code and natural language or between multiple coding languages (Merow et al., Reference Merow, Serra-Diaz, Enquist and Wilson2023).

The primary difference between previous versions of these models, such as GPT-2, and the current generation of GPT models—which as of writing was GPT-4—is that newer models have vastly more parameters and are trained on substantially more data. Nearly everything about the training paradigm, pipeline, and validation is the same as previous iterations, with the caveat that the ChatGPT implementations of these models are fine-tuned on human feedback data, referred to as reinforcement learning from human feedback (RLHF), to better reflect human dialog rather than longform text. RLHF enables LLMs to align text with the complex values of users and reject questions that are inappropriate or outside of the scope of the model’s knowledge, and seems to improve LLM’s ability to carry on a dialog (Bai et al., Reference Bai, Jones, Ndousse, Askell, Chen, DasSarma, Drain, Fort, Ganguli, Henighan, Joseph, Kadavath, Kernion, Conerly, El-Showk, Elhage, Hatfield-Dodds, Hernandez, Hume, Johnston, Kravec, Lovitt, Nanda, Olsson, Amodei, Brown, Clark, McCandlish, Olah, Mann and Kaplan2022). Fine-tuning with RLHF has ostensibly improved the dialog performance of GPT models and driven the rapid adoption of ChatGPT, and indeed much of the attention granted to LLMs hinges on outputs that are lengthy, varied, flexible, and coherent, as selected for by the RLHF process (Clark et al., Reference Clark, August, Serrano, Haduong, Gururangan and Smith2021).

3. What are the current concerns with LLMs?

Although outputs generated by LLMs can sound human-like, coherence is often incorrectly conflated with understanding. It is critical to bear in mind that the process by which LLMs generate output is not a reliable source of facts (Bender et al., Reference Bender, Gebru, McMillan-Major and Shmitchell2021; Weidinger et al., Reference Weidinger, Mellor, Rauh, Griffin, Uesato, Huang, Cheng, Glaese, Balle, Kasirzadeh, Kenton, Brown, Hawkins, Stepleton, Biles, Birhane, Haas, Rimell, Hendricks, Isaac, Legassick, Irving and Gabriel2021; World Economic Forum [Internet], 2023). Just as human-generated prose or code can contain factual or formatting errors or reflect certain world views, LLM-generated text can also contain mistakes or recapitulate the dominant views. In general, the reliability of the outputs of ChatGPT varies substantially across domains; in particular, it tends to underperform on science-related prompts (Shen et al., Reference Shen, Chen, Backes and Zhang2023). LLMs may be prone to demographic, cultural, linguistic, ideological, and political biases, (Bolukbasi et al., Reference Bolukbasi, Chang, Zou, Saligrama and Kalai2016; Caliskan et al., Reference Caliskan, Bryson and Narayanan2017; Buolamwini and Gebru, Reference Buolamwini and Gebru2018; Bender et al., Reference Bender, Gebru, McMillan-Major and Shmitchell2021; Kirk et al., Reference Kirk, Jun, Volpin, Iqbal, Benussi, Dreyer, Shtedritski and Asano2021; Ferrara, Reference Ferrara2023a). Biases in the form of systematic misrepresentations, attribution errors, or factual inaccuracies can arise from decisions made in the development and implementation of LLMs, including the selection of training data, decisions about algorithm architecture, details of product design, and policies that control model behavior (Ferrara, Reference Ferrara2023a). Also troubling is the fact that models can, with ostensibly equal confidence, create false but believable “facts,” known as hallucinations (Weidinger et al., Reference Weidinger, Mellor, Rauh, Griffin, Uesato, Huang, Cheng, Glaese, Balle, Kasirzadeh, Kenton, Brown, Hawkins, Stepleton, Biles, Birhane, Haas, Rimell, Hendricks, Isaac, Legassick, Irving and Gabriel2021; Liu et al., Reference Liu, Zhang and Liang2023).

One important yet underappreciated aspect of LLM performance is the curation of the training data corpus—the steps used to prepare cleaned training samples from raw text. This process often includes the removal of data considered harmful, irrelevant, or incorrect. The decisions about what falls into these categories are often subjective, and consequently the outcome of such decisions may vary depending on the data science workers building the dataset or the LLM’s target audience (Chung, Reference Chung2019; Miceli et al., Reference Miceli, Posada and Yang2022). While intended to remove objectionable content such as violence or hate speech, data curation can have unintended consequences and create temporal homogenization of viewpoints (Bender et al., Reference Bender, Gebru, McMillan-Major and Shmitchell2021). Further, confirmation biases may arise from training data that was unintentionally curated to reflect individuals’ viewpoints (Bolukbasi et al., Reference Bolukbasi, Chang, Zou, Saligrama and Kalai2016; Caliskan et al., Reference Caliskan, Bryson and Narayanan2017; Ferrara, Reference Ferrara2023a). Unlike humans, who naturally censor new and emerging social taboos from their work, LLMs may cast a long shadow of outdated norms. Further, due to different amounts of training data available for different languages, ChatGPT performs more effectively with (e.g.) European languages that have lots of online training content available than with languages that lack this body of data (Jiao et al., Reference Jiao, Wang, tse, Wang, Shi and Tu2023).

To return to the earlier example, when asked to predict what comes after the phrase “I’m going to walk the…,” it is reasonable to think that most people would guess “dog,” as this is a very common phrase in American English. However, it is certainly not true that the sentence must conclude with the word “dog.” For example, perhaps the person writing has a pet cat that they take on daily walks! This example may feel forced or trivial, but when you begin to consider the sum total of text on the internet, there are many viewpoints, opinions, and people that are not represented and innumerable subtle representation biases in this large body of text. The probability of text generation is commensurate with its availability on the internet, and when that text is disproportionately generated by a small slice of the world’s population, then that text will disproportionately reflect their world views. For example, Wikipedia contributors are primarily white, cisgender men, leading to subtle but important systemic biases in topic contribution and tone (Shaw and Hargittai, Reference Shaw and Hargittai2018; MIT Technology Review [Internet], 2023). When the text an LLM is trained on is the product of a limited worldview, the model output propagates a specific knowledge base and worldview.

An especially insidious way that this emerges is in errors of omission. For example, when asked who discovered the global warming potential of carbon dioxide, ChatGPT responded by making reference to publications from the 1970s by John Sawyer and Wallace Broecker (Sawyer, Reference Sawyer1972; Broecker, Reference Broecker1975) as well as the Intergovernmental Panel on Climate Change (Box 1). However, this response ignores the contributions of Eunice Newton Foote who correctly theorized the connection between increased carbon dioxide and planetary warming based on experimentation over a hundred years before (Huddleston, Reference Huddleston2023). The erasure of her story from history is perpetuated when we use models trained on the consequences of this erasure.

Box 1. Propagating bias through omission. When asked about the early recognition of carbon dioxide as a greenhouse gas, ChatGPT perpetuates a common erasure of contributions by Eunice Newton Foote. Conversation with ChatGPT 3.5 on May 2, 2023.

Troublingly, the training data used in current LLMs are so voluminous that documenting and understanding what these models are based on is nearly impossible (Bender et al., Reference Bender, Gebru, McMillan-Major and Shmitchell2021). Further, because of the size and complexities of model architecture and training corpus, even small decisions can have large reverberations in outcomes through the Butterfly Effect (Ferrara, Reference Ferrara2023b). In addition, data curation and training processes are usually opaque (ZDNET, 2023), since many models are trained by large corporations that are not bound to open science data practices or peer-review processes and as such rarely follow best practices laid out by the AI ethics research community (Mitchell et al., Reference Mitchell, Wu, Zaldivar, Barnes, Vasserman, Hutchinson, Spitzer, Raji and Gebru2019; Pushkarna et al., Reference Pushkarna, Zaldivar and Kjartansson2022). Financial barriers further limit the number of organizations with the capacity to train and host such models. This has left only a select few organizations such as Google, Meta, and OpenAI / Microsoft as competitive in this space. Open-source examples of LLMs, such as the BLOOM project, are even rarer. Open-source models generally have lower performance (Chen, Reference Chen2023), potentially due to differences in the all-important data curation step, which is usually opaque to all but the modelers. However, the model weights for Meta’s LLaMA (Large Language Model Meta AI) were leaked shortly after the model’s February 2023 release, which has allowed several open-source and local implementations of LLMs to emerge (e.g., Alpaca [Stanford], Koala [Berkeley], and Vicuna [multiple institutions]). Despite the general trend toward the use of closed-source models, there are also increasing pushes for open LLM research, such as the AI Alliance, which could help bring more transparency to model development and deployment.

4. How might LLMs affect our research practices?

Given the capabilities of LLMs like ChatGPT, we may need to rethink how our community engages with our work and one another. We explore a hypothetical example of how ChatGPT’s core functionalities (text generation, translation, and data analysis/visualization) could be leveraged in the development and execution of an environmental data science project (Figure 1; Supplementary material). In this case, a researcher is inspired by a recent publication that identified pervasive biases in biodiversity records in the United States of America due to historical residential redlining practices. The researcher is interested in further exploring how racist policies impact biodiversity data collection and turns to ChatGPT to help brainstorm research questions and approaches as well as generate code to perform relevant analyses.

Figure 1. Collaborating with ChatGPT on an environmental data science project. In this hypothetical example, we explore how ChatGPT’s (3.5) core functionalities may be used for environmental data science. We have annotated to highlight notable components of responses. Responses have been truncated for brevity. A full transcript can be found in the Supplementary material.

In response to a prompt to brainstorm a research project on the topic, ChatGPT quickly generated a lengthy research project proposal. While much of the text appears to be a useful starting place for a grant or dissertation proposal, many of the details are quite vague and seem to recapitulate obvious statements. For example, under a section entitled “Quantitative Analysis” the output suggests the user to “evaluate” and “assess” the impact of racist policies on biodiversity data collection, closely mirroring the original prompt without suggesting specific methodologies. ChatGPT is unable to provide more specific instructions on how to break down a complex research question into an actionable analytical framework without further prompting. Therefore, identifying the appropriate methods to use still requires expert knowledge. Interestingly, the output includes recommendations to meaningfully engage with Indigenous communities and center their knowledge and perspectives, which remains rare in much of environmental data science.

In response to prompts to provide code for potentially relevant analyses, ChatGPT returned code that handled simple tasks well, but it struggled with more complex and ambiguous prompting. For example, the outputted code correctly handled example datasets as spatial objects. However, minor issues arose when it suggested reassigning a dataset’s coordinate reference system instead of using a transformation. This is a relatively small error that would be straightforward for an experienced user to catch, but less obvious to a more novice user of spatial data. Additionally, the response to a subsequent, more ambiguous prompt for assistance in data visualization was largely incorrect. Despite creating example datasets to visualize as spatial objects, the suggested code for joining these datasets to one another does not treat them as such, in this case it suggests a standard merge instead of a spatial merge. This incoherence demonstrates a fundamental aspect and limitation of ChatGPT’s output generation process – it is simply predicting the most likely set of text, and therefore does not build a logically cohesive response.

While this is merely one example of the ways that ChatGPT may be used (or misused), in Table 1 we provide a more general overview of the potential applications, benefits, and pitfalls of using LLMs in EDS research. One of the most powerful uses of AI will be the automation of routine tasks, which would ostensibly allow researchers to spend more time on interpretation and deep thinking. LLMs may also aid interpretation by identifying patterns and relationships in complex data and have the potential to improve reproducibility and consistency within EDS by lowering the barriers to writing code when reproducing other’s results. ChatGPT has been shown capable of creating code for ecological tasks. It performs best when generating shortcode blocks based on highly specific prompts, although almost all suggested code requires debugging (Merow et al., Reference Merow, Serra-Diaz, Enquist and Wilson2023). These tools also have the ability to translate code to text, which could facilitate consistent methodological descriptions across published literature and help students and other researchers interpret and use a new codebase. Further, ChatGPT may be able to translate between coding languages. As a highly interdisciplinary field with many communities of practice, more seamless translation between coding languages could help bolster collaboration. Consequently, LLMs could facilitate collaboration across disciplines, from biologists to managers to statisticians.

Table 1. General overview of the key functionalities, applications, and potential benefits and pitfalls of ChatGPT

Note. This summary is based on considerations with environmental data science, although they may apply more broadly.

Despite these potential benefits, LLMs do not reduce the need for human-based validation and assessment. Using LLMs to perform data analysis can create situations where EDS practitioners who may have limited understanding of the underlying techniques might apply them inappropriately when recommended by ChatGPT, as users tend to overly trust AI-generated content (Perry et al., Reference Perry, Srivastava, Kumar and Boneh2022). These considerations will be particularly important as we train the next generation of environmental data scientists, who will begin their learning in the time of LLMs. Since LLMs produce output based on their inputs and training data, their ability to construct novel approaches to scientific or data analysis problems is necessarily limited. Additionally, writing code is a skill that not only enables the interpretation of complex environmental data but also allows scientists to think creatively and critically about their data. Removing the scientist from methodological development could be detrimental to both the development of the scientist and the correct interpretation of nuances in the data and/or analysis. It is incumbent upon users to consider potential collateral damages inherent with a more “efficient” approach and likely best to think of AI tools as supporting, rather than replacing, existing duties.

Just as the outputs of LLMs are based on our scientific knowledge base, its biases are also our own. As environmental data scientists, we are responsible for ensuring that we do not propagate and amplify embedded biases in the outputs of LLMs. However, it is not just the outputs that should be used responsibly. We also have a responsibility to design studies that, to the extent possible, contribute unbiased information back into the training corpus that will be used for future LLM training. Recognizing and completely eliminating bias can be challenging; thus, one potential use of LLMs could involve flagging elements of our prose or data analyses that introduce biases, with the caveat that these tools still may overly homogenize what is considered biased (Chung, Reference Chung2019; Miceli et al., Reference Miceli, Posada and Yang2022). Ultimately, scientists may one day outsource the coding itself, but will still need to be trained in how to prompt AI tools appropriately (Zamfirescu-Pereira et al., Reference Zamfirescu-Pereira, Wong, Hartmann and Yang2023), how to assess the validity of their outputs (Passi and Vorvoreanu, Reference Passi and Vorvoreanun.d.; Zombies in the Loop? 2023), and to consider the societal implications and applications of these outputs (Tomašev et al., Reference Tomašev, Cornebise, Hutter, Mohamed, Picciariello, Connelly, Belgrave, Ezer, van der Haert, Mugisha, Abila, Arai, Almiraat, Proskurnia, Snyder, Otake-Matsuura, Othman, Glasmachers, de Wever, Teh, Khan, De Winne, Schaul and Clopath2020; Krügel et al., Reference Krügel, Ostermaier and Uhl2023).

Looking forward, generative AI could accelerate the broader application and integration of insights from the fields of applied science and environmental justice, helping to link scientific findings with ethical considerations and management decisions across local, regional, and global scales. When deciding whether to use LLMs as a tool, environmental data scientists should reflect on what biases LLMs might introduce in the specific context of their work, and when biases or factual errors exist, how they will identify them. Researchers must have enough comfort with their field’s approaches that they can anticipate areas where LLMs will fall short (e.g., if a commonly used package has been replaced since the LLM training data were updated), and researchers should also be aware of any policies that might restrict their use of LLMs in proposing projects (e.g., grant writing), analyzing data (e.g., data confidentiality), and in communicating their findings (e.g., article drafting) (Box 2).

5. How might LLMs affect our teaching practices?

The widespread availability of LLMs will shape how we train the next generation of environmental data scientists, though the advantages and disadvantages of using systems like ChatGPT in the classroom are still being explored (Baidoo-Anu and Owusu Ansah, Reference Baidoo-Anu and Owusu Ansah2023; Meyer et al., Reference Meyer, Urbanowicz, Martin, O’Connor, Li, Peng, Bright, Tatonetti, Won, Gonzalez-Hernandez and Moore2023; Milano et al., Reference Milano, McGrane and Leonelli2023). Students are already using ChatGPT in primary school (Khan, Reference Khan2023) and higher education settings (Cu and Hochman, Reference Cu and Hochman2023) with many institutions struggling to keep pace. At one extreme, the New York City Department of Education initially banned ChatGPT on all devices and networks within the city’s school system (NBC News, 2023). Alternately, some school districts have opted to partner with Khan Academy and OpenAI to introduce automated tutoring with GPT-4, with tens of thousands of students using these AI tutors to date (Khan, Reference Khan2023). The decision on whether or not LLMs are more helpful or harmful for students’ education remains to be seen, with many institutions opting for a “wait and see” approach on whether to adopt or to ban these tools in the classroom.

Proponents argue that these resources should remain free and easily accessible (Mills et al., Reference Mills, Bali and Eaton2023), as they could help to reduce barriers to learning, improve learner confidence, and support diversity and inclusion in environmental data science (Samuel et al., Reference Samuel, George and Samuel2020). LLMs also have the potential to help instructors assist students in large enrollment courses (Popenici and Kerr, Reference Popenici and Kerr2017) and to foster a self-directed learning environment (Wilcox, Reference Wilcox1996). As novices, students greatly benefit from practicing with feedback (Keuning et al., Reference Keuning, Jeuring and Heeren2019), and for data science skills like coding, a responsive LLM that understands code could provide personalized practice for students struggling to learn coding languages. Additionally, by using an LLM for coding, students may feel more empowered to apply their knowledge and skills to new environmental problems and attain agency over their own learning, an established equitable teaching practice (Madkins et al., Reference Madkins, Howard and Freed2020). However, learning how to code without LLMs helps students to build troubleshooting skills and perseverance (Calışkan, Reference Calışkan2023); thus, there are concerns that ChatGPT would reduce the need for students to develop these skills.

When approaching how we teach EDS in the world of LLMs, it may be helpful to consider an asset-based approach (Alim et al., Reference Alim, Paris and Wong2020). This entails teaching practices and instructional materials that promote a variety of approaches to knowing and doing. While LLMs can suggest new approaches, a risk with using LLMs in this context is that (as mentioned earlier) training data can be biased (Wich et al., Reference Wich, Bauer and Groh2020), which may lead to a more privileged set of students’ needs being met. Furthermore, such tools can be expensive and potentially further economic inequity. Even non-profit systems such as Khan Academy’s GPT-4-powered AI tutor are so expensive to run that individual students wishing to access the system are required to donate $20 to help offset the cost (Khan, Reference Khan2023).

We urge a transparent and cautionary approach for instructors using LLMs in classroom settings (Eager and Brunton, Reference Eager and Brunton2023; Meyer et al., Reference Meyer, Urbanowicz, Martin, O’Connor, Li, Peng, Bright, Tatonetti, Won, Gonzalez-Hernandez and Moore2023). In deciding how and when to use LLMs in the classroom, instructors should consider the technical background of their students and especially ensure that their students understand the limitations of LLMs. They should also consider their learning goals for the course and use these goals to decide which assessments will allow use of LLMs. Once they have made these decisions, instructors should be explicit with their students about their expectations regarding the use of LLMs and should provide some basic training on the use of these tools (e.g., specificity when prompting for a code example) so that students have equal access to them (Box 3).

6. How will LLMs impact how we engage with one another?

The issues raised above pose larger questions of how we engage with each other as a community of practice. As LLMs are increasingly used to support analysis, interpretation, and decision-making, whom do we hold accountable when things go awry? How far upstream do we hold individuals responsible? What is our responsibility in shaping how AI uses the information that we generate? Given that these technologies do not credit the sources of the knowledge they repurpose, there is a potential threat to open science principles and practices. For example, if EDS practitioners decide to withhold their data or insights in order to preclude their use in training LLMs, then the use and reuse of those data and results would likely be restricted for use by human researchers as well.

In terms of accountability, the January 2023 policy for Nature family journals states that “… no LLM tool will be accepted as a credited author on a research paper. That is because any attribution of authorship carries with it accountability for the work, and AI tools cannot take such responsibility” (No authors, 2023). Presumably this type of approach also applies to code generated for the purposes of scientific analysis. LLMs may produce outputs based on inputs that were not intended to be re-used, compared to resources like StackOverflow, where contributions are explicitly volunteered for reuse. More broadly, citation and attribution is an important way of understanding the origin and credibility of statements in research. The inability of LLMs like ChatGPT (and AI-based coding assistants like GitHub’s CoPilot) to properly cite their sources suggests a need for more specialized tools that can provide proper attribution. Attribution tools are important for scientists not only for assessing the veracity of provided outputs, but also to maintain legal compliance with copy-left licenses like GPL and CC-BY-SA. To that note, some LLMs like Google’s Bard or Ought’s Elicit can already properly cite sources (web URLs and scientific articles, respectively), although these tools are still in early development.

Another potential path forward is the integration of LLMs with more mechanistic or symbolic systems like Wolfram Alpha. OpenAI, the developer of ChatGPT, has recently added the ability for ChatGPT to interact with “plugins,” including one for Wolfram Alpha, that query other services and return results, but these plugins are still in early development. These plugins have already started to lead to new applications of LLMs in novel areas, like task automation with auto-GPT, but these tools have the potential to fundamentally alter how we communicate as scientists if everything from calendar invites to emails to even Slack messages may be generated by ChatGPT rather than by our colleagues. While these integrated LLM tools are nascent, their profound implications warrant proactive discussion around community norms concerning communication in an age where any piece of text we read in any context could theoretically have been generated by an AI model.

7. What are the costs of LLMs?

While we have yet to fully grasp the opportunities enabled by ChatGPT and other LLMs in the environmental domain, if these tools deliver on the promise to rapidly synthesize concepts and generate complex code, they stand to transform society’s capacity to address some of our most pressing and complex planetary crises by allowing more people to engage on these interdisciplinary problems (Biswas, Reference Biswas2023). While the power of these tools may at times seem a panacea, it is critical to bear in mind that they exist within the physical and social realities of the world. In using these tools, we must consider the ethical implications of the costs of their development, which reflect and are likely to exacerbate existing class and racial injustices.

For example, substantial human labor is needed to counteract the dissemination of the darkest aspects of the internet within LLMs (Rome, Reference Rome2023; Time, 2023). Because ChatGPT primarily relies on text scraped from the internet, OpenAI leverages RLHF and a tool developed to detect violent, sexist, or racist language and preclude its inclusion in the training corpus. To learn to detect this language, the tool required many examples of violence, sexual abuse, and hate speech that had been coded as such by human workers. OpenAI outsourced labor for this task to workers in Kenya, paying <$2 per hour for the identification of harmful text samples. Some workers were made physically ill by these images and descriptions of violence and abuse (Ellis-Soto et al., Reference Ellis-Soto, Chapman and Locke2023). Unfortunately, this outsourcing is a common practice for training AI models that need large sets of labeled data, presenting a social and ethical cost largely unseen by end users (Miceli and Posada, Reference Miceli and Posada2021). As an alternative censoring approach, rule-based, or “constitutional,” AI might help reduce reliance on large amounts of human-classified training data, but would also require subjective decisions about what to include in these “constitutions” (Anthropic, 2023).

Developing, training, and running ChatGPT has an environmental cost as well. The training of GPT-3 led to an estimated 552 metric tons of CO₂ equivalent emissions, roughly the same as generated by three round-trip jet plane flights between San Francisco and New York. ChatGPT has an estimated daily carbon footprint at the scale of 23 kg CO₂e (Patterson et al., Reference Patterson, Gonzalez, Hölzle, Le, Liang, Munguia, Rothchild, So, Texier and Dean2022; Ludvigsen, Reference Ludvigsen2023). These costs will likely grow with the use of ChatGPT, but are still currently orders of magnitude lower than the carbon emissions attributed to airborne travel to conferences (Patterson et al., Reference Patterson, Gonzalez, Le, Liang, Munguia, Rothchild, So, Texier and Dean2021). Nevertheless, more transparent and rigorous accounting of efficiency gains versus energy usage is needed, especially for LLMs hosted by private companies that currently have no mandatory emissions reporting requirements (Bender et al., Reference Bender, Gebru, McMillan-Major and Shmitchell2021).

In addition to the human and environmental costs of developing, training, and maintaining ChatGPT, its use, and implementation in environmental data science raises several considerations around policy dynamics and imbalances. Leveraging ChatGPT for addressing our most pressing planetary crises risks handing over power historically held by government agencies and resource users to the companies that develop these tools (Kalluri, Reference Kalluri2020). Disparities in underlying data and decisions about model design and training stand to create not only biased content but also biased research agendas and biased researcher ideologies. For example, it is not difficult to imagine a case where the political and financial interests of big tech companies are misaligned with global environmental stewardship.

Finally, access to LLMs will likely be inequitable. Currently, ChatGPT is free to use (assuming access to a computing device and the internet) but OpenAI also sells a paid tier subscription (20 USD/month) that provides priority access, faster response times, and access to GPT-4 (vs. 3.5 in the free version). For larger queries to the OpenAI application programming interface (API), OpenAI charges per-token (currently 3 USD for every 1,000 tokens for GPT-4 and 15 USD for every 1,000 tokens for GPT-3.5) (https://openai.com/pricing). While prices are low per-token for now, with ever-larger datasets and more complicated queries and data manipulation steps, costs can quickly add up and introduce disparities between the volume of data that researchers can afford to manipulate with privately-held LLMs. Furthermore, it is unclear how the costs of using these tools will evolve in the future and whether or not access to higher-performing tools will mirror current inequities in access to proprietary computational software (Ghose and Welcenbach, Reference Ghose and Welcenbach2018). On top of many other ethical issues with these products, access to these tools might be the most immediate way that this technology reproduces or exacerbates existing equity issues in the field of EDS.

8. Where do we go from here?

While the future of LLMs is not yet clear, what is certain is that these technologies are here to stay. There are substantial social, ethical, and environmental issues with the creation and maintenance of these tools, and their adoption will have potentially profound implications for the practice of science. Thus, our community needs to not only proactively consider the implications of their use for our own research and training tasks, but to also consider how we might leverage our uniquely human perspectives to guide the development of these tools and their use more broadly. Tools like ChatGPT can assist EDS practitioners in the development of ideas and translating them into code. However, creating meaning from the output still requires expert knowledge. While LLMs may save time on certain tasks, their use runs the risk of propagating biases embedded in their training and therefore requires caution. As a community, we see the need for greater dialogue about the issues raised by the emerging use of LLMs. We offer a series of discussion prompts for engaging with community members within a range of contexts: within labs, classrooms, departments, and conferences (Box 2, 3).

Taken together, these changes pose a fundamental question about the nature of our roles as environmental data scientists going forward. Ultimately, LLMs are built off of human ideas, and the continued relevance of these tools depends on the sustained generation and distribution of novel ideas. Without this, scientific insight will stagnate and scientific thinking will become increasingly homogenized. Even while we as scientists continue to do what models cannot yet accomplish—to innovate and to think critically about the challenging environmental issues of our time—we need to comprehensively reassess the skills that are needed to succeed in these roles in the age of LLMs.