Carbon per Accuracy as a sustainability metric likely rewards under-trained models because accuracy scales non-linearly with compute¶

Claim · All claims
Evidence · 1 study · theoretical r2 · q1
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q1 i? The proposed Carbon per Accuracy metric rests on a false assumption of linear accuracy scaling and will likely reward under-trained models rather than strong performance. → Eimler 2026

Evidence¶

Eimler 2026¶

Eimler, S.C., Erle, L., Flood, D., Haiman, A., Häckert, L., Helgert, A., McGinness, L., Yapici, B. (2026). The Environmental Cost of LLMs in AIED: Reporting and Practices. https://arxiv.org/abs/2606.11215 doi:10.48550/arxiv.2606.11215

q1 · i? · theoretical · r2

This is the authors' analytical argument about their own proposed metric, not an empirical test. They note Carbon per Accuracy interprets CO2 required to reach accuracy 1.0 "under the (false) assumption that accuracy scales linearly with computational expense", so it "will likely reward under-trained models".

"In reality, the performance of ML models plateaus during training, and the relation between accuracy and training compute is highly non-linear. This means that carbon per accuracy will likely reward under-trained models and not reward strong performance."

Discussion¶

Related Claims¶

• Few AIED 2025 papers report computational costs or discuss environmental impacts, and reporting is non-standardized — related
• Training large AI models carries substantial environmental and human labor costs — related
• The review reports that training a single deep learning model may have a carbon footprint comparable to several automobiles over their lifespan
• Under identical LoRA hyperparameters and training data, the 27B Gemma model outperforms the 70B LLaMA model on every computed essay-scoring metric — related