💻 computer science

The Environmental Cost of LLMs in AIED: Reporting and Practices

This paper highlights the lack of standardized reporting on the hidden computational and environmental costs of Large Language Models in Artificial Intelligence in Education (AIED) and proposes an open-source framework, including software tools and calculation formulas, to systematically measure and transparently report these impacts.

Original authors: Sabrina C. Eimler, Lukas Erle, Daniel Flood, Aditi Haiman, Luca Häckert, André Helgert, Lachlan McGinness, Büsra Yapici

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sabrina C. Eimler, Lukas Erle, Daniel Flood, Aditi Haiman, Luca Häckert, André Helgert, Lachlan McGinness, Büsra Yapici

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of Artificial Intelligence in Education (AIED) as a bustling, high-tech school. For a long time, everyone has been excited about the new "smart tutors" (Large Language Models, or LLMs) that can write essays, grade papers, and chat with students. The school has been so focused on how good these tutors are at teaching that they've completely forgotten to look at the bill for the electricity and water needed to run them.

This paper is like a group of concerned teachers and engineers walking into that school and saying, "Hey, we need to start reading the meter."

Here is a breakdown of what they found and what they propose, using simple analogies:

1. The "Hidden Bill" Problem

The researchers looked at 396 recent papers from the AIED 2025 conference. It's like checking the report cards of 396 different school projects.

What they found: Almost everyone (257 out of 396) is using these smart AI tutors.
The Gap: Only a tiny handful mentioned how much energy these tools consume. Even fewer talked about the environmental "smoke" (carbon emissions) they produce.
The Analogy: Imagine if every student in the school was asked to build a giant robot to help them study. They would all write a report on how cool the robot is, but almost no one would mention that the robot eats 500 pounds of coal a day. The paper argues that we need to start listing that coal consumption on the report card.

2. The Current Mess: No Standard Ruler

Right now, if a researcher does mention the cost, they all do it differently.

Some say, "It took 2 hours."
Others say, "We used a specific computer."
Some just say, "It's bad for the planet" without numbers.
The Analogy: It's like everyone measuring the length of their desk in different units—some use inches, some use "hand spans," and some use "pizza slices." You can't compare them. The paper says we need a single, standard ruler so we can actually compare who is being more efficient.

3. The Solution: A New "Energy Calculator"

The authors propose a toolkit to fix this, which they call TerraFlops (think of it as a "Green Calculator").

For Local Computers (The Home Office):
If you run an AI on your own laptop, the paper suggests using a tool called CodeCarbon.
- The Analogy: This is like a smart plug that tells you exactly how much electricity your coffee maker uses. The paper notes that this tool sometimes underestimates the bill because it forgets to count the energy used by the fan cooling the computer or the monitor screen. To fix this, they propose a "PUE-equivalent" factor. Think of this as a "waste multiplier." If your computer is only working at 10% capacity but the whole machine is humming along, you multiply the energy cost to account for that wasted idle power.
For Cloud Computers (The Big Data Centers):
Many researchers use big servers (like AWS or Google) instead of their own laptops.
- The Analogy: Since you can't plug a "smart plug" into a server farm in another country, the authors suggest a theoretical estimate. They created a formula based on how many words (tokens) the AI reads and writes.
- The "Guessing Game": If you are using a secret, proprietary AI (like a super-smart model from a big tech company) and you don't know how big it is, the paper suggests a "rule of thumb." They assume big models are roughly the size of known open-source giants (like 100 billion "brain cells" or parameters). It's not perfect, but it gives you a consistent way to say, "This task cost roughly X amount of energy," so you can compare it to other tasks.

4. The New Report Card: "Carbon per Accuracy"

The paper suggests that when researchers publish their work, they shouldn't just say, "Our AI got 90% accuracy." They should also say, "Our AI got 90% accuracy, but it cost 5 kilograms of CO2 to get there."

They propose a few new metrics:

Corrected Carbon Intensity: The total pollution, adjusted for how efficiently the computer was used.
Carbon per Accuracy: How much pollution was needed to get one point of improvement?
Sustainability Score (1-10): A simple grade, like a report card, that combines the pollution and the performance. A score of 10 means "Super efficient," and a 1 means "Very wasteful."

5. The Call to Action

The authors aren't just complaining; they are handing out the tools. They have made their software open-source (free for anyone to use) and are asking the AIED community to adopt these standards.

The Goal: They want future conferences to require these "energy reports" just like they require authors to be anonymous.
The Vision: If everyone starts reporting the "hidden bill," the community can start making smarter choices. Maybe we'll realize that a slightly less accurate AI is worth using because it saves a massive amount of energy, or maybe we'll find better ways to run the ones we have.

In short: The paper says, "We love the new AI tutors, but we can't keep ignoring the electricity bill. Let's measure it, report it, and start making our educational AI greener."

Technical Summary: The Environmental Cost of LLMs in AIED: Reporting and Practices

Problem Statement

The integration of Large Language Models (LLMs) into Artificial Intelligence in Education (AIED) has accelerated, yet the computational and environmental costs associated with their usage remain largely unreported and unstandardized. While ethical discussions in AIED frequently address bias, privacy, and student over-reliance, the energy consumption and carbon footprint of training and deploying LLMs are often ignored. This lack of transparency is exacerbated by the absence of standardized procedures to measure and report these impacts. Consequently, the community lacks a unified understanding of the environmental burden of educational AI, making it difficult to compare the efficiency of different systems or to establish norms for sustainable research.

Methodology

The authors employed a mixed-method approach combining a systematic literature review with the development of open-source technical frameworks for measurement and estimation.

1. Literature Review

The authors conducted a comprehensive review of all 396 papers published in the AIED 2025 conference proceedings. They analyzed submissions across all tracks (full papers, workshops, doctoral consortium, etc.) using a coding framework to extract:

LLM Usage: Whether an LLM was used (binary) and the specific role (Subject of Experiment, Data Analysis, Product Design, or User Interaction).
Reporting Practices: Whether computational costs (runtime, energy, hardware) and environmental sustainability were explicitly reported.
Model Details: Specific model versions and parameter information.

2. Development of Measurement Frameworks

To address the identified gaps, the authors proposed and implemented three technical solutions:

Local Hardware Measurement (CodeCarbon): The authors utilized the open-source Python package CodeCarbon to estimate carbon emissions by measuring local hardware power draw and converting it to CO2 based on local electricity grid mixes. They noted that while CodeCarbon provides a baseline, it often underestimates total consumption by ~20% because it does not account for power supply inefficiencies or peripheral cooling.
The TerraFlops Framework: To improve accuracy for local development environments, the authors introduced TerraFlops, an open-source wrapper for tracking tools. This framework introduces a PUE-equivalent (Power Usage Effectiveness) metric for local machines. Unlike data centers with high utilization, local devices have significant baseline overhead (OS, display, cooling) regardless of ML workload. TerraFlops dynamically adjusts the PUE-equivalent (ranging from 1.60 for low-load to 1.08 for high-load) to account for this wasted energy.
- The framework also includes a Cloud Mode that applies fixed PUE values based on major providers (AWS, GCP) to standardize deployment cost projections.
- It generates composite metrics: Corrected Carbon Intensity, Carbon per Accuracy, and a Sustainability Score (1–10 scale using logarithmic transformation).
Theoretical FLOP Estimation for LLMs: For scenarios where direct measurement is impossible (e.g., proprietary models via API), the authors proposed a method to estimate computational cost using Floating Point Operations (FLOPs).
- They derived formulas for decoder-only transformers (including Pre-Layer Normalization, RoPE, and KV caching) to calculate FLOPs for the first token and subsequent tokens.
- They proposed a simplified approximation: $FLOPs = 2Nn$ (where $N$ is active parameters and $n$ is total tokens).
- For proprietary models where $N$ is unknown, they suggested order-of-magnitude estimates (e.g., ~100B parameters for frontier models, ~30B for fast models) and a conversion factor to CO2 (approx. $6 \times 10^{-5}$ grams CO2 per T-FLOP, assuming H100 GPUs and US grid mix).

Key Results

The literature review of the 396 AIED 2025 papers yielded the following findings:

High Adoption: 257 papers (65%) utilized LLMs.
Usage Patterns: The most common use case was User Interaction (176 papers), primarily as tutors or assistants. Other uses included Data Analysis (79), Content Design (108), and Subject of Experiment (100).
Reporting Deficit: Only 85 papers provided any details on computational costs, and only 57 papers mentioned sustainability or environmental issues.
Inconsistency: Among those that did report, there was no standardization in metrics, units, or depth of analysis.
Model Prevalence: OpenAI models (GPT-4/4o) were the most frequently cited, followed by Anthropic Claude, Google Gemini, and various open-source models (Llama, Mistral, etc.).

Key Contributions

Empirical Evidence of the Gap: The paper provides the first systematic quantification of the lack of environmental reporting in the AIED community, demonstrating that while LLM usage is ubiquitous, its cost is invisible in the vast majority of research.
TerraFlops Tool: The introduction of an open-source tool that extends carbon tracking to include PUE-equivalent adjustments for local hardware, offering a more realistic estimate of the environmental footprint of research conducted on consumer-grade machines.
Standardized Reporting Proposals: The authors propose a set of composite metrics (Corrected Carbon Intensity, Carbon per Accuracy, Sustainability Score) to be reported alongside traditional accuracy metrics.
Estimation Methodology for Proprietary Models: A practical, three-step procedure for researchers using API-based LLMs to estimate their computational cost and carbon footprint without access to internal hardware data, utilizing token counts and parameter approximations.

Significance and Claims

The authors argue that the environmental cost of LLMs in AIED is a critical, yet neglected, ethical dimension. They claim that by making these costs visible through standardized reporting, the community can:

Establish norms for transparent reporting of hidden costs.
Enable the comparison of models based on efficiency (performance per unit of CO2) rather than accuracy alone.
Motivate researchers to optimize for sustainability, potentially reducing the environmental impact of educational AI.

The paper concludes with a call to action for the AIED community to adopt these reporting practices, suggesting that future conferences could mandate sustainability metrics as a condition of submission, similar to anonymity requirements. The authors position their work not as a definitive solution, but as a foundational step toward a more sustainable and accountable AIED research ecosystem.