ChatGPT Can Predict Environmental Research Trends, but Mostly by Repeating Them

Can a chatbot do what scientists pride themselves on most: innovate? A new study puts that question to a direct test in environmental research, and the answer is a carefully qualified no—at least for now. Researchers fed two decades of publications from the journal Environmental Science & Technology into ChatGPT-4 and asked the model to predict which topics would dominate the field’s future. The model performed surprisingly well at forecasting next year’s research hotspots, correctly anticipating up to 80 percent of them under the best experimental setup. But when the team dug into where that accuracy came from, the picture became far less flattering: the model succeeded largely by echoing the very keywords it had been given, revealing an aptitude for thematic continuity rather than genuine scientific creativity.

The study, published in March 2026 in the Springer Nature journal Carbon Neutral Systems, was motivated by a long-standing question in artificial intelligence research: whether machines can take over innovative work, or whether they remain tools for efficiency. Innovation has traditionally been regarded as a uniquely human capability, but the rapid rise of generative AI systems such as ChatGPT, Gemini, and DeepSeek has unsettled that assumption. These large language models, built on Transformer architectures and trained on massive text corpora, produce coherent, context-aware outputs that closely resemble human writing, and they have begun playing visible roles in scientific research and knowledge production. The authors set out to examine whether such a model could exhibit innovation-like behavior when applied to a concrete scientific task.

The researchers chose a proxy task with real stakes: predicting future research hotspots in environmental science. They selected Environmental Science & Technology, a leading journal in the field that has published more than 41,000 papers over recent decades, and extracted the complete publication data for 34,472 articles spanning 2000 to 2022 from the Web of Science database in April 2023. Using the bibliometric tool VOSviewer, they analyzed titles, keywords, and abstracts year by year, filtering out stopwords and merging plural forms, then ranked words by frequency of occurrence to identify the top recurring terms for each year and for aggregated multi-year windows. These high-frequency terms served a dual purpose: mapping the evolving priorities of environmental research and providing the curated input that would be fed to ChatGPT.

The experimental design was deliberately constrained. ChatGPT-4 was not given any full-text articles, publication metadata, or explicit temporal labels—only the aggregated keyword lists extracted from predefined historical windows. The team acknowledges an important caveat here: because ChatGPT-4 was trained on publicly available text up to April 2023, it may have been exposed during pre-training to environmental science literature or to general patterns of thematic change in the field. The task therefore does not constitute a strict out-of-sample forecast of unseen documents, but rather an evaluation of the model’s ability to infer short-term thematic continuity from limited historical signals. All experiments were repeated ten times and the results averaged, with accuracy defined as the share of correctly predicted words and repetition as the share of predictions matching the words originally provided.

The first results were humbling. When given the most frequent words from 2000 to 2009 and asked to predict the dominant terms of 2010 to 2019, ChatGPT’s accuracy was low: in predictions based on the title and abstract sections, only two words—emissions and toxicity—were correctly anticipated. Most forecasts gravitated toward broad, plausible-sounding terms such as climate, sustainability, renewable, and microplastic, which are relevant to environmental discourse but did not match the words that actually dominated the decade. When the team shortened the input window further, feeding only 2000–2004 data and asking for predictions across three subsequent five-year periods, the forecasts became even more generalized. Less information and a longer predictive horizon, the authors conclude, dilute precision considerably.

The breakthrough came from tightening the task. When the researchers supplied ChatGPT with the top ten most frequent words from every four-year segment and asked it to predict the predominant terms for the following year, accuracy improved markedly—reaching up to 80 percent for abstract terms. But deeper analysis revealed why. Environmental research topics exhibit strong temporal continuity: the dominant words in any given year tend to resemble those of the preceding four years. ChatGPT exploited this by frequently repeating the provided terms in its predictions, achieving high correctness simply by echoing its input. When the team explicitly prohibited repetition of previously provided words and repeated the experiment ten times, accuracy fell to zero in the vast majority of cases. The single exception was striking: in one run predicting abstract terms for 2010 to 2019, the model correctly produced toxicity without repeating any input keyword.

Those rare successes are what the authors call weak signals of novelty generation. In a few instances, ChatGPT predicted keywords such as toxicity, degradation, and emission that later became prominent despite being absent from the historical hotspot lists it received as input. The team is careful about interpretation: these occasional correct predictions do not demonstrate a stable or generalizable innovation capability, but they suggest that under highly constrained conditions the model can sometimes infer semantically relevant concepts beyond direct keyword repetition—likely a product of its training on large-scale text and its capacity to capture broad semantic associations, rather than autonomous creative reasoning. The authors also note that scientific innovation never means complete independence from existing knowledge; real innovation typically builds incrementally on prior work, so the no-duplication experiment should be read as a boundary test, not a definitive verdict on creativity.

The study’s limitations are candidly laid out. Beyond the pre-training contamination concern, the authors acknowledge that a frequency-based bibliometric approach may retain generic or method-related terms commonly used in scientific writing, and that such terms were deliberately not manually excluded to preserve objectivity—meaning generic high-frequency words may contribute to apparent continuity in hotspot evolution. The experimental setup, based on retrospective data analysis, also cannot fully mirror the unpredictable, multifaceted challenges of real-world scientific exploration. The observed prediction performance, the team writes, primarily reflects the model’s ability to reproduce and extrapolate dominant research themes under controlled input conditions, not evidence of true foresight or independent discovery of future scientific directions.

The implications extend beyond benchmark scores into ethics and governance. The authors argue that as AI-driven models increasingly support emission assessment, pollution mitigation, resource optimization, and climate-related decision-making, limitations in model transparency, data bias, or predictive reliability could directly influence carbon accounting accuracy and sustainability policy. They also flag the environmental footprint of AI itself: the energy consumption of data centers and large-scale computation adds a layer of ecological responsibility, and responsible deployment requires minimizing that footprint, potentially through sustainable energy sources. Transparency, accountability, inclusivity toward different ecosystems and indigenous data sovereignty, and honest acknowledgment that disinformation cannot be fully eliminated from these systems all feature in their framework for ethical AI in environmental science.

Looking forward, the researchers call for systematic cross-version and cross-model comparisons—testing Gemini, DeepSeek, and successive ChatGPT releases—to determine whether the heavy reliance on keyword repetition and the sporadic appearance of innovation-like signals are intrinsic to generative AI or specific to particular models. They also suggest exploring how different prompting strategies, input representations, and domain constraints shape predictive performance. Their bottom line is measured: in its current state, ChatGPT cannot substitute for human beings in scientific innovation, which involves hypothesis formulation, conceptual abstraction, and paradigm shifts beyond trend extrapolation. But the capabilities it did display, and the ethical questions they raise, deserve serious attention in advance—before, not after, such systems become embedded in the research enterprise.

Subject of Research: Testing whether ChatGPT can exhibit innovation-like behavior by predicting future research hotspots from two decades of environmental science literature.

Article Title: Will artificial intelligence challenge human in innovative works?– a perspective in environmental research

Article References: Chen, Y., Tao, J., Chen, G., Yan, B., Chen, C., Liang, R., Cheng, Z., & Hou, L. (2026). Will artificial intelligence challenge human in innovative works?– a perspective in environmental research. Carbon Neutral Systems, 2(1), Article 4. https://doi.org/10.1007/s44438-026-00025-3

Image Credits: AI Generated

DOI: 10.1007/s44438-026-00025-3

Keywords: ChatGPT, generative AI, environmental research, innovation, research hotspots, bibliometrics, large language models, scientific publishing, AI ethics, predictive modeling, Environmental Science & Technology, carbon neutrality