Transparency in Climate Analytics Methods

Just read fascinating research on AI energy consumption - but the methodology is almost more interesting than the results. The Challenge: How do you measure the carbon footprint of Claude, GPT-4, or any commercial AI when the companies won't share their infrastructure data? The Solution: Researchers combined public API performance data with reverse engineering: ⌚️ Step 1: Scraped latency and tokens-per-second from artificialanalysis(dot)ai across 30 models 📈 Step 2: Used statistical inference to estimate hardware (Claude likely runs on AWS H100/H200 based on performance patterns) ⚡️ Step 3: Applied the formula: Energy = (inference time) × (estimated GPU power + system overhead) × datacenter efficiency 💦 Step 4: Layered in region-specific multipliers for carbon intensity and water usage They didn't need Anthropic's internal data. They used publicly observable performance to work backward to energy consumption. Key Finding: Claude-3.7 Sonnet scored highest (0.886) in eco-efficiency, while DeepSeek models used 70x more energy than GPT-4.1 nano. When companies won't publish sustainability metrics, researchers find creative ways to find out anyway. It's the core philosophy of the Impact Framework, if you can observe something, you can measure it's impact. Even if an organizations is not disclosing, they might be leaking enough observable information that you can model the impacts anyway. Transparency through ingenuity. 👏 Great work Nidhal Jegham, Marwan F. Abdelatti, El Moubarki Lassaad and Hend Abdeltawab

🌿 New Research: AI, Climate, and Transparency in EU Regulation 🌡️ Happy to share a new paper I've co-authored with Nicolas Alder, Kai Ebert and Ralf Herbrich: "AI, Climate, and Transparency: Operationalizing and Improving the AI Act." In this short interdisciplinary paper, we critically examined the EU AI Act's climate-related provisions, particularly on climate reporting. Our key findings include: 1. Significant gaps in the Act, such as: - Overlooking energy consumption during AI inference - Neglecting indirect greenhouse gas emissions from AI applications - Lack of access for the general public to climate disclosures - Lack of standardized reporting methods (hopefully being adequately addressed as part of the standardization process) 2. A novel interpretation to potentially include inference-related energy use 3. Advocacy for public access to climate-related AI disclosures 4. Recommendation for energy reporting at the cumulative server level We also suggest policy changes like sustainability risk assessments (arguably included under my reading of the AI Act) and renewable energy targets for data centers for a more comprehensive approach to AI's environmental impact. This work represents our collective effort to contribute to the ongoing dialogue on AI regulation and sustainability. We hope it sparks further discussion and research in this crucial area. You can download and read the full paper here: https://lnkd.in/eEcFE-Nf Looking forward to your thoughts and feedback. How can we better balance AI and technical innovation with environmental responsibility? BTW, my longer paper on Sustainable AI Regulation, including thoughts on sustainability impact assessments and emissions trading, is here: https://lnkd.in/eH5W7Mj5 published: https://lnkd.in/eVxC2Hr4 #AI #ClimateChange #Regulation #ResponsibleAI #aiact cc: European New School of Digital Studies

Behind the #AI Numbers: How Google, OpenAI & Academia Measure the Climate Cost of Every Prompt. Let's start with the data: Google’s recent transparency on Gemini’s inference footprint is a useful step for the sector: their team reports a median Gemini text prompt uses ~0.24 Wh, emits ~0.03 g CO₂e, and consumes ~0.26 mL of water — figures derived from a full-stack, in-production measurement framework. For comparison, Sam Altman's public commentary estimates an average ChatGPT query at ~0.34 Wh and a very small water footprint (~0.000085 gallons, or ≈0.32 mL) — broadly the same order of magnitude but not identical in method or boundary. Independent benchmarking shows a wide spread across models and deployment choices based on the infrastructure-aware study: "How Hungry is AI?" reports examples from ~0.43 Wh for a short GPT-4o query to many Wh (tens of Wh) for some long, inefficient model deployments, underscoring how model architecture, batching, hardware, and operational choices drive outcome variance. Key Takeways here: a/ Technical transparency is a positive step, as shown by Google and OpenAI. Publishing data—and methodology—is the foundation for accountability. b/ Median efficiency doesn’t tell the whole story. We must also examine outliers, indirect resource use (like infrastructure and power production), and water sourcing. c/ Scale magnifies even small inefficiencies. With billions of prompts generated daily, cumulative energy—even with tiny per-query footprints—becomes significant. d/ We need shared industry benchmarks. Google’s call for standardized, full-stack metrics is timely; infrastructure-aware comparisons like those from the University of Rhode Island research help signal the way forward. e/ Proactive research and policy alignment are key. As AI becomes more embedded in everyday life, a balanced approach—innovating for capability and sustainability—is not only smart but essential. References (and you should read the actual papers to think about it): 1/ Measuring the environmental impact of delivering AI at Google Scale by Google 2/ How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference by Nidhal Jegham et al 3/ The Gentle Singularity by Sam Altman

Data Transparency and Verification in Global Temperature Archives. A recurring question in climate science is not whether temperature records require adjustment, but whether those adjustments can be independently verified. Berkeley Earth (BE) publishes a raw temperature archive with checksums, allowing third parties to confirm data integrity and reproduce results. This level of transparency enables genuine auditing and fact-checking. NOAA’s GHCN-M v4 archive, by contrast, aggregates daily station data from global sources, converts daily min/max values into monthly means. NOAA provides a clear and detailed methodology paper describing this process (https://lnkd.in/ge-F2cG8), but access to the underlying unadjusted station archive appears limited, making independent replication difficult. 📄 Methodology: Observations from GHCN-M v4 (2025 release) for Rio de Janeiro : ~77% of data archived in 2013 remains in the 2025 archive ~4% of observations receive positive adjustments ~52% receive negative adjustments ~51% of adjustments are < 1.0°C ~3% exceed 1.0°C Maximum positive adjustment: +4.59°C in February 2003 and November 2009 When comparing monthly temperatures: Berkeley Earth data vs NOAA GHCN-M v4 data. The difference between the two…the resulting series diverge more than expected, given that both rely on many of the same underlying stations. Site-level corrections are typically reported as <0.3°C, which does not fully explain the observed differences. Questions worth discussing: Why do datasets derived from largely overlapping station records yield materially different temperature histories? Where is public access to NOAA’s full unadjusted station archive to allow independent verification? How can researchers reproduce NOAA’s results end-to-end without access to the complete raw inputs? As an example, Berkeley Earth data for Rio de Janeiro suggest cooling through the mid-19th century, followed by gradual warming after ~1920. Some portion of recent warming may reasonably reflect urbanization effects, highlighting the importance of station-level transparency. The core issue is not trust, but reproducibility. Scientific confidence is strongest when results can be independently validated using openly available data.

Many passenger aviation carbon calculators still fall short — they’re often too narrow, not transparent enough, and communicate impacts poorly. The result? Systematic underestimation and declining trust. Our latest study takes these gaps head-on. We’ve developed a comprehensive methodology that captures a much fuller picture of aviation’s climate impact — including NOx, water vapour, contrail-induced cloudiness, upstream emissions from in-flight services, and full life-cycle emissions from aircraft and airports. We also push accuracy further with detailed modelling of flight distance, fuel burn, and emissions allocation by passenger class, luggage, and cargo. A historical adjustment factor, validated against more than 30,000 flights, brings real-world variation into pre-flight estimates — achieving an average error of just ~0.5%. The result is a tool that provides unmatched transparency and granularity, offering a clear breakdown of where emissions come from and why they matter. And it confirms what many suspected: current methods consistently underestimate aviation emissions. If you’re interested in where aviation carbon accounting needs to go next, our new video explainer walks through the approach and what it means for industry stakeholders.

I believe that radical transparency should be a requirement for models in the climate and natural resources arenas, because these models ultimately serve the common good. Not necessarily the internals of the models, but transparency around how well they perform and how much they can be trusted. In that spirit, as a first step, I am sharing the results of a big effort to validate our ML models against every recorded wildfire that occurred anywhere in the western US this year (Jan-July). Two versions of the ML models predicted the fire size classes correctly in 60% and 75% of the nearly 12K fire incidents. The models predicted the fire cause correctly in 82% of the cases. To put the complexity in context, this involves predicting the sizes and causes of fires ignited at random locations on random start dates (with very specific weather and dryness on those dates at those locations) across 11 states covering 1.2 million sq miles. The validation also pointed to very specific areas for further investigation and improvement. Models by definition are never perfect and never complete, they are always a work in progress, so this kind of feedback is a nice benefit of validation. #WildFire #WildfirePrediction #DeepLearning #MachineLearning #modeltesting #modelvalidation #RadicalTransparency