As a global leader in AI infrastructure, Google should take the lead not just in technological performance, but also in honest, systemic environmental transparency.

Analysis of Google's AI Environmental Impact Report and Its Criticism

by ChatGPT-4o

I. Overview of Google’s Claims

In August 2025, Google released a study detailing the environmental footprint of its AI system, Gemini, with a key claim: a median AI text prompt consumes just 0.24 watt-hours of energy, emits 0.03 grams of CO₂, and uses 0.26 milliliters of water—roughly five drops. These figures represent a 33x improvement in energy efficiency and a 44x reduction in emissions over the previous year due to model and infrastructure optimizations.

Google emphasized its full-stack measurement methodology, which includes:

• Active AI accelerator use

• CPU/DRAM power

• Idle machine energy

• Data center overhead

• Water usage via Water Usage Effectiveness (WUE)

• Market-based emissions calculations under GHG Protocol

This approach contrasts with narrower estimates in prior literature, which often focus only on the energy consumed by AI chips (TPUs or GPUs) during active inference.

II. Summary of Criticism

Despite Google's extensive disclosures, critics—including experts like Shaolei Renand Alex de Vries-Gao—raise several concerns as outlined in an article published by The Verge:

1. Selective Omissions:

   • Google's study excludes indirect water consumption, such as water used for generating electricity via gas or nuclear power plants.

   • Only market-based (MB) carbon emissions are reported, not location-based (LB) ones, which are typically higher and reflect the actual energy mix at data center sites.

2. Lack of Transparency:

   • The report relies on median rather than mean usage, hiding the variability and high-resource prompts that significantly affect aggregate environmental impact.

   • There is no disclosure of token counts or prompt lengths—key to reproducing or comparing the estimates.

3. Misleading Comparisons:

   • Google claims prompt energy usage is “orders of magnitude” lower than prior studies—but those studies often included both direct and indirect impacts, while Google excludes indirect water use and only considers MB emissions.

   • The claim of “five drops of water per prompt” has been called symbolic rather than scientifically comprehensive, obscuring real systemic effects.

4. Jevons Paradox Ignored:

   • Efficiency gains can drive up overall resource usage due to increasing demand—a classic Jevons Paradoxsituation. Despite per-prompt efficiency, Google's total emissions increased by 11% last year due to AI scaling.

III. My Analysis

A. Google’s Strengths

• Google deserves credit for:

  • Publishing first-party operational data (a rare move in a secretive industry).

  • Adopting a comprehensive stack-based approach, including idle energy and data center overhead, which exceeds many academic benchmarks.

  • Disclosing progress over time, with quantifiable year-over-year improvements and detailing software and hardware innovations (e.g. speculative decoding, MoE, TPU efficiency).

B. Validity of the Criticism

• The criticism is well-founded and critical for trust in corporate sustainability claims:

  • Relying only on market-based emissions (MB) paints an overly optimistic picture. Location-based (LB)emissions show the actual grid mix, which can be fossil-heavy in key regions like Texas.

  • By focusing on the median prompt, Google may underrepresent outliers(e.g. long prompts, multimodal outputs), which have a disproportionately larger impact.

  • Water use from power generation is real and quantifiable, especially with gas and nuclear plant expansion for data center growth. Excluding this gives a misleadingly low estimate.

C. Broader Context

• The AI boom’s infrastructure impact is now tied to grid reliability, emissions goals, and local ecosystems:

  • Google's claim of 0.26 mL per prompt sounds low until one considers that billions of prompts are served daily, multiplying these "drops" into millions of liters.

  • AI’s environmental costs are not evenly distributed; they concentrate in areas with cheap water and power, often in already strained ecosystems.

• Google’s claims also reflect a classic corporate narrative strategy:

  • Emphasize relative improvements (e.g. “44x better than last year”) while downplaying absolute environmental costs.

  • Use analogies (“as much energy as watching TV for 9 seconds”) to reframethe narrative from environmental responsibility to harmlessness.

IV. Recommendations for Google

To enhance credibility, transparency, and leadership, Google should:

1. Include Location-Based Emissions:

   • Disclose both market-based and location-based emissions as per GHG Protocol best practices, especially when reporting to stakeholders and regulators.

2. Account for Indirect Water Use:

   • Factor in upstream water use from electricity generation, especially as reliance on thermoelectric power plants grows to meet AI-driven energy demand.

3. Publish Mean and Range in Addition to Median:

   • Report means, medians, and upper/lower bounds to accurately reflect the environmental distribution of AI prompts, including multimodal and long-form generations.

4. Open Methodology to Peer Review:

   • Submit the study for independent peer review. Transparency about data, models, prompt length distributions, and energy breakdowns will help ensure trust.

5. Disclose Prompt Characteristics:

   • Share average token counts, latency thresholds, and prompt types (e.g., code, image, long-text) to contextualize findings across different use cases.

6. Add Third-Party Verification:

   • Partner with academic or non-profit research institutions to validate claims and collaboratively develop benchmarks for industry-wide comparison.

7. Avoid Oversimplified Analogies:

   • Refrain from framing large-scale energy and water use through "drops" or “seconds of TV” analogies, which risk trivializing systemic infrastructure strain.

V. Conclusion

Google’s report is a strong technical achievement and an important contribution to environmental transparency in AI. However, the selective boundaries and framing choices undermine its credibility and risk misleading the public and policymakers. While per-prompt consumption is falling, overall environmental burden is increasing, and cherry-picking metrics cannot substitute for comprehensive, reproducible accountability.

As a global leader in AI infrastructure, Google should take the lead not just in technological performance, but also in honest, systemic environmental transparency. Future versions of its methodology should aim not only to defend its efficiency gains, but to acknowledge AI’s real environmental costs—and help others mitigate them.