Environmental Knowledge Test EKT-19

• Environmental Knowledge Test (EKT-19) is a standardized multiple-choice tool designed to quantitatively evaluate understanding across seven key environmental science domains.
• It employs a corpus-based, curriculum-driven methodology to ensure balanced, data-driven content reflecting contemporary environmental discourse.
• The tool benchmarks both human and AI performance, revealing significant score discrepancies that inform curriculum enhancements and AI model development.

The Environmental Knowledge Test (EKT-19) is a standardized assessment tool designed to quantitatively evaluate knowledge across key domains of environmental science. Extensively applied in both education and artificial intelligence research, EKT-19 serves as a robust benchmark for measuring competency in environmental concepts among university students and LLMs, enabling comparative analyses and informing the development of advanced educational and AI evaluation frameworks.

1. Definition and Structure

EKT-19 is constructed as a multiple-choice questionnaire that comprehensively spans the principal thematic areas within environmental science. The test consists of 30 items that systematically address the following categories: Ecology, Climate, Resources, Consumption Behavior, Society & Politics, Economy, and Environmental Contamination. Each domain is represented by multiple dedicated questions to ensure balanced coverage. The content is intended to probe both factual knowledge and applied understanding, with domains and items curated to reflect both academic rigor and contemporary relevance (Smail et al., 5 Aug 2025).

2. Methodological Foundations and Construction

The design of EKT-19 leverages corpus-based and curriculum-driven approaches to ensure validity and representativeness. Essential resources for assembling key terminologies and environmental concepts include the EcoLexicon English Corpus (EEC), a 23.1-million-word corpus of contemporary environmental texts compiled by the LexiCon research group and available in Sketch Engine (Leon-Arauz et al., 2018). Utilizing EEC, developers employ advanced querying capabilities, such as Concordances, Corpus Query Language (CQL), semantic word sketches, and the EcoLexicon Semantic Sketch Grammar (ESSG), to extract context-rich terminological data. This enables:

• Filtering texts by domain and target audience to capture nuanced conceptual usage.
• Employing collocational and semantic analyses to ensure each test item reflects authentic scientific discourse.
• Deriving frequency and relation statistics for terms and phraseology, facilitating data-driven validation of content selection.

Mathematical representations for semantic relation and frequency are formally expressed as follows:

$$\text{Term} \xrightarrow{\text{relation}} \text{Related Concept}$$

$$RF = \frac{f_{\text{term}}}{\sum f_{\text{all terms}}} \times 100$$

where $RF$ is relative frequency, $f_{\text{term}}$ is the frequency of the targeted term, and $\sum f_{\text{all terms}}$ is the total word count in the relevant corpus or subdomain (Leon-Arauz et al., 2018).

3. Benchmarking AI and Human Acquisition

EKT-19 is prominently used in evaluating not only student attainment but also the environmental knowledge of LLMs. In recent investigations, prominent LLMs such as GPT-3.5, GPT-4, GPT-4o, Gemini, Claude, and Llama 2 were administered EKT-19 alongside cohorts of university students representing both environmental and non-environmental majors (Smail et al., 5 Aug 2025). Key findings include:

• Mean score for LLMs: 26.67/30 ($\sim$88.9%)
• Mean student score: 13.20/30 ($\sim$44%)
• Best model (Claude): 30/30
• Domain-specific LLM outperformance in all categories; e.g., Resources, AI=2.67, Students=0.87
• Statistical analyses:
  • $t = -2.69$, $p = 0.009$ (GPT-3.5 vs. Claude)
  • $t = 2.12$, $p = 0.04$ (GPT-4 vs. Llama 2)
  • $\chi^2 = 15.30$, $p = 0.0092$ (correct rates across LLMs)

These results demonstrate both the high degree of AI knowledge retention and the limitations of human acquisition given typical curricular exposure. However, they also underscore the continued necessity of human experts for validation, contextual adaptability, and critical engagement with ambiguous or open-ended questions (Smail et al., 5 Aug 2025).

4. Integration with Environmental Education and Interdisciplinary Approaches

Reflection on educational best practices indicates that EKT-19 encapsulates not just rote recall but the application of interdisciplinary methods. Research into the integration of physics and environmental education highlights how conceptual frameworks (such as energy conservation, thermodynamic efficiency, and quantitative modeling) can be incorporated into EKT-19's question structure (Valderrama et al., 6 Jan 2025). Example formulae employed in such integration include:

$$E = K + U,\quad \eta = \frac{E_{\text{out}}}{E_{\text{in}}},\quad Q = mc\Delta T$$

This approach enables test items to probe both knowledge and applied reasoning, assessing the ability to interpret data and solve environment-centric quantitative problems—a feature that is essential for fostering critical thinking and practical competence in prospective environmental professionals.

5. Comparative Datasets and Evaluation Frameworks

The emergence of specialized benchmarks such as EnviroExam (Huang et al., 18 May 2024) and the Environmental LLM Evaluation (ELLE) dataset (Guo et al., 10 Jan 2025) situates EKT-19 within a larger context of environmental knowledge evaluation. EnviroExam, for example, features 936 multiple-choice questions across 42 environmental science courses, employing rigorous statistical methodologies including the mean score $(M)$, standard deviation $(\sigma)$, coefficient of variation $(CV = \sigma/M)$, and aggregate performance index $(I = M \times (1 - CV))$ to assess model competence and consistency. ELLE comprises 1,130 QA pairs segmented by domain, difficulty, and cognitive type (knowledge, calculation, reasoning), with evaluation structured along professionalism, clarity, and practical feasibility. EKT-19 shares foundational similarities with these benchmarks but is distinct in its standardized, human-oriented deployment and its recognition in studies directly comparing human and LLM performance (Smail et al., 5 Aug 2025).

Evaluation Tool	Number of Questions	Target Subjects	Distinctive Metrics/Features
EKT-19	30	7 domains	MCQ, standardized test, human+LLM benchmarking
EnviroExam	936	42 courses	MCQ, 0/5-shot, $CV$, $I$
ELLE-QA	1,130	16 domains	Knowledge/calc/reasoning, expert-labeled

6. Implications for AI Development, Education, and Policy

EKT-19, particularly in tandem with tools such as EEC, EnviroExam, and ELLE, enables researchers and practitioners to:

• Distinguish LLMs’ factual competence from their adaptability and ability to generalize across environmental subfields.
• Identify significant performance gaps between automated systems and students taught using traditional content delivery.
• Justify curricular enhancements and the adoption of interactive, AI-supported teaching methodologies in environmental education.
• Deploy nuanced statistical screening via metrics such as $CV$ or composite scoring for robust model selection and targeted fine-tuning (Huang et al., 18 May 2024).

In policy and professional settings, EKT-19 may serve as an element in certification processes, professional development, and large-scale educational diagnostics, contingent on periodic validation and refreshment of its item pool using data-driven terminological corpora and current curricular standards.

7. Future Directions

Contemporary research advocates for the continual refinement and expansion of EKT-19, incorporating new item sets from diverse cultural, linguistic, and research traditions. A plausible implication is the development of domain-specific variants based on specialized environmental textbooks or the incorporation of dynamic item generation for adaptive assessment. Furthermore, future integration with AI benchmarks like EnviroExam and ELLE is anticipated to enhance both the sophistication and granularity of AI and human performance diagnostics in environmental science (Huang et al., 18 May 2024, Guo et al., 10 Jan 2025). The convergence of corpus-driven test construction, interdisciplinary content, and advanced model benchmarking will likely shape the next generation of environmental knowledge assessment tools.