Table of Contents

Note from the Author: I'm trying out a formatting experiment for gradually expanding on a topic of interest. First we give a TL;DR (Too Long, Didn't Read) 1-2 sentence blurb about it, then a 1-2 paragraph summary, then an outline of the basic points, and finally provide more deep-dive content into the topic and provide some key references with links, and some details on how we created this using a variety of AI based tools. Please let me know how you like this approach.

TL;DR

AI-driven work generally consumes more upfront energy, water, and resources compared to human labor but scales far more efficiently, making it cost-effective for large tasks. AI's environmental and economic sustainability depends on use case and energy source—when properly optimized, AI can reduce emissions in some industries but worsens energy and water demand if unchecked.

🔊Prefer audio over reading? Listen to the podcast summarising this article, hosted by two AI bots, Ethan and Sarah.

Summary

AI is resource-intensive upfront, requiring massive energy, water, and financial investment to train large models (e.g., GPT-4, AlphaFold). Training a major AI model can consume thousands of megawatt-hours of electricity and emit hundreds to thousands of tons of CO₂, while also using millions of liters of water for cooling. However, once trained, AI models can rapidly complete tasks at near-zero marginal cost, making them highly scalable and economically efficient for mass deployment.

In contrast, human work has much lower direct environmental costs, but is slower, more expensive per unit of work, and scales linearly—more tasks require more workers. AI outperforms humans in content generation, document review, coding, and scientific research when speed and scale matter, but still requires human oversight for quality, creativity, and ethical considerations. AI also lowers the cost of tasks that previously required significant human labor, such as scientific discovery, legal analysis, and software development.

The net sustainability of AI depends on energy sources and efficiency improvements. AI’s footprint can be mitigated by running models on renewable energy, improving hardware efficiency, and focusing AI on sustainability-positive tasks (e.g., climate modeling, energy optimization). While AI's environmental impact is significant, its ability to accelerate scientific breakthroughs and economic productivity suggests that when strategically deployed, it can offset its own footprint and drive long-term sustainability.

Outline

1. Introduction

• AI vs. Human Work Comparison: AI is faster, scalable, and cost-effective after training but has high upfront energy, water, and emissions costs. Humans are more environmentally efficient per task but limited in scale and speed.

• Key Metrics Considered: Energy use, carbon footprint, water consumption, costs (gross/net), and sustainability potential.

2. AI vs. Human Work in Traditional Tasks

• Content Generation (GPT-4 vs. Human Writers)

  • AI training (e.g., GPT-3) used 1,287 MWh of electricity, 502 tCO₂, and 5.4 million liters of water.

  • A human writer uses ~0.1 kWh per page and near-zero emissions.

  • AI scales at near-zero cost post-training, whereas human costs scale linearly.

  • TL;DR: AI writing is efficient for large-scale content but has a high training footprint.

• Document Review (AI NLP vs. Human Lawyers)

  • AI reduces review time by 90% and costs by 80%.

  • AI processing 1M docs uses ~2–5 MWh vs. ~8 MWh for humans.

  • TL;DR: AI-powered document review is faster and cheaper with a lower total footprint.

• Code Generation (AI Coding Assistants vs. Human Programmers)

  • AI coding models (e.g., Code Llama) used 3.3M GPU-hours, emitting 539 tCO₂.

  • AI-assisted developers work 55% faster and AI-generated code saves hours per week.

  • TL;DR: AI coding increases productivity but requires oversight for errors.

3. AI vs. Human Work in Deep Technical Applications

• Protein Folding (AlphaFold vs. Human Biologists)

  • AlphaFold training required ~1,000 MWh, 100+ tCO₂, but predicted 214M+ protein structures that would take centuries of human work.

  • AI’s impact vastly outweighs its energy cost in accelerating biotech breakthroughs.

  • TL;DR: AI-driven biology eliminates decades of experimental work with a net environmental benefit.

• AI as a Research Collaborator (Google AI Co-Scientist vs. Human Researchers)

  • AI proposes novel hypotheses and accelerates R&D (e.g., drug discovery).

  • AI literature review uses hundreds of GPU-hours vs. human researchers taking weeks/months.

  • TL;DR: AI boosts research productivity but doesn’t replace human innovation.

4. AI’s Environmental & Economic Sustainability

• Energy Use & Carbon Footprint

  • Data centers consumed 460 TWh in 2022, AI use is growing rapidly.

  • Microsoft’s emissions rose 30% from AI expansion, Google’s up 50% since 2019.

  • TL;DR: AI has a massive and growing energy demand, but efficiency improvements can help.

• Water Consumption

  • AI training uses millions of liters per model for cooling.

  • ChatGPT uses ~500 mL per 20–50 queries due to cooling needs.

  • TL;DR: AI’s hidden water footprint is substantial and needs attention.

• E-waste & Hardware Manufacturing

  • AI hardware (GPUs/TPUs) has a high carbon footprint from mining & manufacturing.

  • AI clusters need replacing every 3-5 years, leading to e-waste.

  • TL;DR: AI needs better hardware sustainability to reduce e-waste impact.

• Economic Sustainability (Cost vs. Value Created)

  • AI delivers $3.5–4 return per $1 invested (IDC study).

  • McKinsey: Generative AI could add $2.6–4.4T per year to the economy.

  • TL;DR: AI generates economic returns that justify its environmental costs.

5. Overall Conclusion

• AI consumes more upfront energy and resources but enables greater productivity.

• AI vs. Human: Humans are more energy-efficient per task, but AI is superior for large-scale, repetitive, or computationally complex tasks.

• AI sustainability depends on renewable energy, efficient hardware, and responsible deployment.

• Future AI should focus on problems where benefits exceed its footprint, like medicine, climate modeling, and efficiency improvements.

• TL;DR: AI is a resource-heavy investment that can be highly sustainable if used wisely.

Deep Dive

AI vs. Human Work: Environmental and Economic Comparison (2022–2025)

Introduction

Advances in artificial intelligence (AI) have enabled machines to perform tasks traditionally done by humans – from writing content and analyzing documents to coding and even scientific research. This report compares AI-driven work vs. human work for two categories: (1) Traditional AI applications (e.g. content generation, document review, code generation) and (2) Deep technical AI applications (e.g. DeepMind’s AlphaFold and Google’s “AI Co-Scientist”). We examine key sustainability metrics – water use, carbon footprint, energy consumption, costs, total environmental impact, and economic sustainability (output value vs. cost) – based on studies from the past ~3 years (primarily US/UK/EU). We ensure “apples-to-apples” comparisons by considering the resource and training investment for AI models versus educating/training humans for equivalent work.

Below, we analyze each area with cited findings, followed by summary tables and a discussion of overall environmental and economic impacts.

Traditional AI Applications vs. Human Work

1. Content Generation (AI Writers vs. Human Writers)

AI Approach: Large language models (LLMs) like OpenAI’s GPT-3/GPT-4 can generate articles, marketing copy, and other text content. Training these models is computationally intensive. For example, training GPT-3 (175 billion parameters) consumed an estimated 1,287 MWh of electricity and produced 502 metric tons of CO₂​. This is roughly the annual electricity use of 130 U.S. homes​. It also required substantial water for cooling data centers – on the order of 5.4 million liters of water (freshwater) consumed in training GPT-3 in Microsoft’s U.S. facilities​. According to a 2023 study, training GPT-3 can directly evaporate ~700,000 liters of water on-site in the data center​. Once deployed, inference (using the model to generate text) is less intensive per instance: about 0.004 kWh per page of content generated in one estimate. However, serving millions of users adds up – one report estimates ChatGPT’s daily use could emit ~50 lbs of CO₂ (~22 kg), or ~8–9 tons CO₂ per year​. Another analysis noted a single AI query can use 10× to 100× more energy than a standard Google search​.

Human Approach: A human writer’s brain and computer use only a tiny fraction of the energy of an AI supercomputer. A person working on a laptop might consume on the order of 50–100 W of power (0.05–0.1 kWh per hour). Writing an article (say 1 page) might use only 0.1–0.2 kWh of electricity – orders of magnitude less than what an AI uses to produce the same, if we include the AI’s share of training energy. Carbon emissions for a human writing are correspondingly negligible for one piece of content (perhaps a few dozen grams of CO₂, depending on the electricity source). Water use for an individual human writer is minimal (just indirect water used to generate the small amount of electricity they consume). However, humans are far slower – a single person can write maybe a few pages per hour, whereas an AI can generate drafts in seconds once trained. Scaling human content creation to millions of documents would require huge teams of writers (with commensurate energy for office lighting, computing, commuting, etc., though still likely less than AI datacenter requirements).

Costs: Training GPT-3 was expensive (~$4–5 million USD) by one estimate​. The ongoing cloud compute cost to host and run such a model is also substantial (OpenAI’s GPT-4 reportedly cost tens of millions to develop). By contrast, “training” a human writer involves education costs (e.g. university tuition or on-the-job training) and salary. For a rough apples-to-apples: $5M could fund ~100 writer-years (at $50k/year). But that single GPT-3 model can output text equivalent to thousands of writer-years once deployed, if used at scale. Thus, per unit of content, AI can be highly cost-efficient after deployment. In practice, businesses find generative AI more cost-effective for large-scale content: e.g. ChatGPT’s subscription is ~$20/month for unlimited writing, whereas a human freelancer might charge that for a single article. That said, humans often provide higher creativity or nuanced quality for certain content, which AI may need oversight or editing to match.

2. Document Analysis (AI Document Review vs. Human Review)

AI Approach: AI models (often natural language processing systems similar to LLMs or specialized classifiers) can analyze and summarize documents, or sift through large databases of text. This has big applications in legal e-discovery, contract analysis, and data mining. The environmental footprint of training a mid-sized NLP model (e.g. BERT with 110M parameters) is much smaller than GPT-3’s, but still non-trivial – one study from 2019 found training such models can emit tens or hundreds of kg of CO₂ depending on complexity​. For large-scale use (like an AI scanning millions of documents), the runtime energy in data centers becomes significant. For instance, running an AI to review 1,000,000 documents might consume a few MWh of electricity (spread across many servers) – still far less time and arguably less energy than a human team would use to read the same volume. The water usage parallels energy: data centers withdraw ~7,100 liters of water per MWh of energy used, according to a 2021 U.S. study​. So if an AI uses, say, 2 MWh to process a million docs, that might indirectly use ~14,000 liters of cooling water (again, much of that evaporated).

Human Approach: Human document review (e.g. in legal cases) is extremely labor-intensive. Dozens of lawyers or analysts might work for weeks to review a million documents. This has financial and some environmental costs: offices need lighting/AC, computers run for many hours, and employees commute. The energy per document for a human is low (reading a few pages might consume maybe 0.001–0.01 kWh in device and lighting), but because a human takes minutes per document versus an AI’s seconds, the total energy for a million-doc review can be sizable. For example, 1,000,000 docs * 5 minutes each = ~83,000 hours of work. If each human uses ~100 W while working, that’s ~8,300 kWh (~8.3 MWh) for the task – on the same order as an AI, though possibly a bit more. The human process, however, would be spread over time and people, not a burst of data center use. Carbon-wise, those 8.3 MWh might emit a few metric tons CO₂ (if grid power). Water use is again indirect (power generation and office cooling).

Costs & Efficiency: This is where AI shines. Studies show AI-assisted review can cut costs and time drastically. A 2022 analysis by Ernst & Young found that AI “document intelligence” solutions reduced review time by ~90% and costs by ~80% on average​. For example, one eDiscovery firm compared reviewing 1M documents with humans vs AI: human review was estimated around $1.7 million USD, whereas an AI-powered review cost about $450,000 USD​ – roughly a 3–4× cost reduction. In practice, the workflow often combines both (AI quickly filters or summarizes documents, humans do final checks), yielding both speed and accuracy improvements. From an economic sustainability view, the AI investment (software licenses, computing) pays off when large volumes of data need analysis, whereas humans become unfeasible or too costly at scale. Quality-wise, a landmark study even showed an AI algorithm could match or exceed experienced lawyers in spotting contract issues, with 95% accuracy for the AI vs 85% for humans, in a fraction of the time​. Overall, AI can massively increase productivity for document analysis, generating value by handling volumes impossible for small human teams – as long as the upfront computing resources (and their environmental costs) are available.

3. Code Generation (AI Coders vs. Human Programmers)

AI Approach: AI coding assistants like GitHub Copilot (powered by OpenAI’s Codex model) and DeepMind’s AlphaCode have shown the ability to generate computer code from prompts or solve programming challenges. These use AI models trained on billions of lines of code. The training process resembles that of large language models and is similarly resource-intensive. DeepMind’s AlphaCode (published 2022) used transformer models up to 41 billion parameters – likely requiring on the order of several thousand MWh of electricity and emitting hundreds of tons of CO₂ in training (comparable to GPT-3’s footprint)​. Meta’s newer Code Llama (2023) – a 34B parameter code-focused model – was trained on 3.3 million GPU-hours on A100 chips, resulting in 539 tCO₂e emissions (which Meta fully offset)​. Once trained, AI code generators run on cloud servers to assist developers. Each generated function or suggestion has a tiny energy cost (milliseconds of GPU time), but intensive use by many developers could mean more data center load.

Human Approach: Human programmers write, debug, and maintain code using brainpower and standard computers. A skilled software engineer might consume perhaps 100–200 W at their workstation (including their PC and share of office utilities). Over an 8-hour workday that’s ~0.8–1.6 kWh and perhaps ~0.5–1 kg CO₂ emitted (depending on energy source) – very small per day per programmer. The human training process (education, practice) spans years (a computer science degree, etc.), which has societal costs (education infrastructure, etc.) but these are hard to quantify per individual. Critically, humans are limited by working hours and experience – a single developer can only produce so much code in a given time. Scaling up a project means hiring more developers (with linear cost increase).

Productivity & Cost: Recent studies indicate AI coding tools can significantly boost developer productivity. For instance, Microsoft found that GitHub Copilot users completed tasks 55% faster than those without it​ (results vary by task). Economically, this means a company can get more software output with the same human team. The cost of running an AI coding assistant (a subscription or API fee, plus computing overhead) is relatively small compared to a developer’s salary. For example, Copilot costs ~$10/month per user – negligible next to an engineer’s wage – and early research suggests it can save a developer hours of work per week, which is hundreds or thousands of dollars in value. From a sustainability angle, if AI helps code get written more efficiently, it could also reduce the need for overtime or large teams (saving the energy and emissions those extra developers would have used). However, we must account for the massive upfront footprint of training the code model. A fair comparison might be: training a top-tier AI coder model might cost $10–$20 million (including energy, hardware, R&D), whereas training 100 junior programmers for a year might cost a few million in salary and education. But once the AI model is trained, it can potentially assist tens of thousands of programmers simultaneously. In effect, the AI is like a globally shared “worker” whose huge training cost is amortized over many uses. In terms of quality, AI can generate routine code quickly but may introduce errors (“hallucinations” or insecure code) that human engineers must fix. In one experiment, AI-generated code had bugs 16%of the time without human oversight​, underscoring that AI is a tool to augment, not fully replace, human programmers (at least for now).

Deep Technical AI Applications vs. Human Work

4. Scientific Discovery – Case: DeepMind’s AlphaFold vs. Traditional Protein Research

AI Approach (AlphaFold): DeepMind’s AlphaFold 2 (2020) was a groundbreaking AI that predicts 3D protein structures from amino acid sequences. Training AlphaFold was a computational marathon – it used 128 TPU v3 cores for ~11 days (plus additional tuning)​. This amounts to tens of thousands of TPU-hours; estimates put the energy consumption on the order of 1000+ MWhand carbon emissions of ~100+ tons CO₂ for training​. (One 2023 analysis noted AlphaFold and similar models can have carbon footprints “over 100 tCO₂e”​.) The water footprint for this training would also be significant (data centers in 2020 likely consumed ~1–2 million liters for that level of compute, assuming typical efficiency and cooling needs). However, once trained, AlphaFold can predict a protein structure in minutes or hours on much smaller compute (a single GPU). DeepMind released the pretrained AlphaFold model and collaborated to create the AlphaFold Protein Structure Database, which by 2022 contained ~214 million predicted structures​. This means one AI model effectively solved in bulk a task (structure determination) that would have taken human laboratories centuries. DeepMind noted that AlphaFold 2 was used to predict hundreds of millions of protein structures, “which would have taken hundreds of millions of researcher-years” using experimental methods​. In other words, the AI accomplished in a year or two what tens of thousands of human scientists could not realistically achieve at all.

Human Approach (Lab Experiments): Traditionally, determining a protein’s 3D structure requires labor- and resource-intensive experiments like X-ray crystallography or cryo-electron microscopy. These involve growing protein crystals, using large instruments (some are particle accelerators consuming megawatts of power for X-ray generation), and can take months or years per protein. It’s said an individual lab scientist might spend 4–5 years (the length of a PhD) to solve one protein structure, at a cost of hundreds of thousands of dollars per structure​. The environmental impact of experimental structure determination is not trivial: synthesizing proteins, running instruments and refrigeration, and international travel to synchrotron facilities all contribute. For example, a single synchrotron X-ray beamline can use several hundred kW of power; a few days of beam time might consume a few MWh of energy. Plus, reagents and lab equipment carry a chemical and waste footprint. So for each protein, you might imagine on the order of 0.1–1 ton CO₂ emitted and hundreds of liters of water used in various processes (very rough numbers) – multiplied by thousands of proteins, this is huge. Of course, not every protein would ever be solved experimentally – many would simply remain unknown without AI.

Impact and Cost: AlphaFold is a prime example where an AI’s gross resource cost is high, but the net benefit is even higher. The training phase emitted ~100 tCO₂, but by circumventing countless lab experiments, it likely saved far more in aggregate emissions and costs. One perspective: 100 tons CO₂ is equivalent to perhaps 20 experimental structures (if each experiment is ~5 tCO₂); AlphaFold delivered millions of structures at that cost. Economically, AlphaFold’s development was expensive (DeepMind reportedly spent upward of $fold millions on the project, though exact figures aren’t public), but the value generated is enormous – accelerating drug discovery, biology research, and biotechnology development worldwide. It’s hard to put a dollar figure on that, but consider that structure-based drug design can lead to new medicines; even one successful drug can be worth billions. In terms of apples-to-apples training: it takes ~10+ yearsto train a human structural biologist (PhD, postdoc) and perhaps $500k+ in education/employment costs; AlphaFold absorbed that knowledge from data for thousands of proteins in a matter of weeks. While AlphaFold does not “think” like a scientist, it provides a tool that massively augments human scientists’ capabilities. From a sustainability standpoint, if widely used, AI like AlphaFold could reduce the need for certain resource-intensive lab experiments, thereby offsetting its computational footprint. It’s a case where AI and human efforts together yield better outcomes: humans focus on experiments that truly need to be done (e.g. validating key targets), while AI handles the bulk prediction, saving time, money, and resources.

5. AI as a Research Collaborator – Google’s “AI Co-Scientist” vs. Human Researchers

AI Approach (Multi-Agent Research Assistant): In late 2024, Google introduced an “AI co-scientist” – a multi-agent system (built on Google DeepMind’s advanced models, reportedly using the Gemini AI model) designed to act as a virtual scientific collaborator​. This AI can ingest vast amounts of scientific literature, reason about hypotheses, and even debate and refine research ideas among simulated “agent” personas. In an experiment on liver fibrosis research, Google’s AI co-scientist generated multiple novel research hypotheses – “all the approaches suggested by the AI showed promising activity”, meaning the ideas were viable when tested in preliminary lab experiments​. The AI was also able to iterate and improve solutions over time, learning from feedback​. Essentially, this system can do months of literature review and brainstorming in a fraction of the time, by leveraging superhuman reading speed and memory. The computational cost: Running such an AI collaborator likely involves a cluster of machines using tens of kilowatts for hours or days. (For instance, an AI agent might consume a few hundred GPU-hours to deeply analyze thousands of papers and generate hypotheses.) The carbon and water impact of that would depend on the data center – if 100 GPU-hours were used, that might be ~0.05 MWh and a few hundred kg CO₂ (if not renewably powered), plus a few hundred liters of water for cooling. Training the underlying model (Gemini or similar) was far more intensive – likely 50× the energy of GPT-3 as per estimates for GPT-4​. Google hasn’t released details, but GPT-4’s training was speculated to use ~65,000 MWh and emit thousands of tons of CO₂ ​(if not fully mitigated) – an immense initial footprint. So the AI co-scientist stands on the shoulders of a very power-hungry model.

Human Approach: Scientific research is typically done by teams of human researchers using intuition, knowledge, and manual literature review. A human scientist might read dozens of papers, attend conferences, and spend weeks brainstorming to arrive at a new hypothesis. This process is slower and constrained by human cognitive limits – no one can read millions of papers, but an AI potentially can. The energy a human researcher uses (for their computer, office, travel to meetings, lab experiments etc.) is again relatively low on an absolute scale (a few MWh per year for an active researcher’s total footprint perhaps). However, human idea generation can be hit-or-miss; many dead-end experiments may be pursued, consuming lab resources, before a fruitful lead is found. There’s a hidden environmental cost to research inefficiency: lab experiments that fail still used energy and materials. Humans excel at creative and strategic thinking, but even top experts can’t sift through all data out there – meaning some opportunities or optimizations might be missed.

Collaboration & Sustainability: The AI co-scientist is meant to augment, not replace, human researchers​. Google’s scientists emphasize it will “increase, rather than decrease, scientific collaboration” and serve as a brainstorming partner​. The economic angle here is speeding up R&D: if an AI can cut down the time to discover a drug target or new material, that could save companies millions. For example, a pharma company might spend $100M+ over several years on target discovery; an AI that trims even 10% of that time could save tens of millions and get a product to market faster. From a resources view, the AI’s contributions could reduce the number of failed trials or experiments, which is an indirect sustainability win (less lab waste, less energy on fruitless experiments). The cost to run an AI co-scientist is not trivial (it might require cloud computing that costs thousands of dollars per day for large problems), but relative to R&D budgets, this can be acceptable if it yields breakthroughs. One peer-reviewed case of AI-human collaboration showed that using AI to guide experiments (in materials science) significantly reduced the number of trials needed to find optimal solutions​. In other words, AI can guide humans to more efficient experimentation, lowering the environmental impact of trial-and-error. However, training these advanced research AIs (like Gemini) does have a large one-time carbon footprint. It becomes sustainable economically if the model is re-used across many research problems (justifying the up-front cost), and environmentally if its use leads to innovations that help society (e.g. new climate solutions) or if its operation is powered by renewable energy (both Google and DeepMind have pledged to use carbon-free energy and even water-replenishment programs for their data centers​).

Summary of Environmental Metrics: AI vs. Human Work

To put the above into perspective, Table 1 summarizes key environmental and cost metrics for select AI use-cases vs human equivalents. These are approximate figures from recent studies and reports, illustrating the scale of resource usage:

Comparative Metrics: AI Models vs. Human Work

1. Content Generation (AI Writing vs. Human Writing)

• AI (GPT-3/4):

  • Training consumed approximately 1,287 MWh of electricity and emitted 502 metric tons of CO₂.

  • Water use for cooling was about 5.4 million liters during training.

  • After training, generating 100 pages of content consumes 0.4 kWh on average.

  • Estimated cost of training: $4–5 million, but per-page cost after training is nearly zero.

• Human Writers:

  • Writing one page manually consumes about 0.1 kWh (including laptop and lighting).

  • Near-zero emissions per page; a full-time writer emits approximately 0.5 metric tons of CO₂ per year(including indirect office use).

  • Water use is negligible per task, but a human’s lifetime water consumption (~80 years) is around 3–4 million liters for all personal and professional activities.

  • Salary costs scale linearly at $50K per year on average.

2. Document Review (AI NLP vs. Human Reviewers)

• AI NLP Model:

  • Can process 1 million documents using ~2–5 MWh of electricity.

  • Carbon footprint for processing that volume is around 1–3 metric tons of CO₂.

  • Water usage for cooling ranges from 15,000–35,000 liters for every 1 million documents processed.

  • Training time is in the range of days to weeks, and review costs at scale are $0.45 per document ($450,000 per 1M documents).

• Human Reviewers (Legal & Business Analysts):

  • Reviewing 1 million documents requires ~8 MWh of electricity spread across multiple workers.

  • Carbon footprint for that process is around 2–5 metric tons of CO₂, depending on office power sources.

  • Water usage is minimal, mostly indirect (linked to office AC and power generation).

  • The process requires teams working for weeks/months, and the total estimated cost is ~$1.7 million for 1M documents (assuming $50/hr labor).

3. Code Generation (AI Coders vs. Human Programmers)

• AI Coding Assistants (e.g., GitHub Copilot, Code Llama):

  • Training required 3.3 million GPU-hours, emitting 539 metric tons of CO₂.

  • Inference (code generation) requires only milliseconds per suggestion, consuming negligible energy per task.

  • A subscription for AI coding tools costs around $10–$20 per month per user.

• Human Programmers:

  • A typical programmer’s daily energy use is ~1.0 kWh, leading to ~250 kWh per year per developer.

  • Yearly emissions are approximately 0.1–0.2 metric tons of CO₂ per programmer (excluding personal emissions).

  • Education and training take years, and salaries are typically $100K+ per year for skilled developers.

4. Protein Folding & Scientific Research (AI vs. Human Lab Scientists)

• AI (AlphaFold2):

  • Training required ~1,000 MWh of power and emitted 100+ metric tons of CO₂.

  • Once trained, it can predict a single protein structure in minutes using ~0.1 kWh.

  • AI has solved over 214 million protein structures, eliminating the need for countless experimental procedures.

  • Cost of developing the AI model: Estimated $100K–$200K in cloud compute, but released freely for scientific research.

• Human Laboratory Research (Traditional Methods):

  • Determining a single protein structure can take months to years, using high-powered lab instruments.

  • Energy consumption per structure is roughly 1–2 MWh, leading to ~1 metric ton of CO₂ emissions per protein.

  • Lab research involves physical materials and waste, further increasing environmental impact.

  • Estimated cost: ~$100K+ per protein structure due to staff salaries, equipment, and materials.

As the table shows, **AI systems typically incur much higher upfront resource costs (energy, carbon, water, dollars) during training, whereas humans incur relatively low ongoing resource usage but high time and salary costs. Once deployed, AIs can perform many tasks rapidly with moderate incremental energy, whereas a human’s output is limited by working hours but each additional task they do has minimal environmental cost.

Total Environmental Impact and Sustainability Considerations

Energy & Carbon: The expansion of AI has non-trivial energy demands. Data centers worldwide consumed an estimated 460 TWh in 2022 (all purposes) – if they were a country, that would rank ~11th in electricity use​. Generative AI workloads are a fast-growing part of this: training clusters for AI can draw 5–8× more power than typical data center workloads​. The result is surging carbon emissions for tech companies. Microsoft reported its emissions rose ~30% from 2020–2022largely due to AI data center growth​. Google’s emissions in 2023 were nearly 50% higher than 2019 for similar reasons​. Each new generation of AI model is more power-hungry – GPT-4’s training used perhaps 50× more electricity than GPT-3​. If that electricity is from fossil fuels, the carbon footprint scales accordingly. However, big AI firms are increasingly mitigating this by purchasing renewable energy and buying carbon offsets. Meta, for example, fully offset 100% of the 539 tCO₂ from training its LLaMA-2 model​. The trend in policy is toward transparency: the EU’s draft AI Act even considers requiring reporting of energy use and carbon emissions for AI systems. In short, AI’s carbon footprint can be alarming if unchecked​, but there is momentum to manage it (through efficiency and green energy). By contrast, human work has a steady but lower carbon output – mostly from powering offices and commuting. A typical office worker might indirectly be responsible for a few tons of CO₂ per year (including commuting), whereas an AI model might burn that in a few training runs.

Water Usage: Often overlooked until recently, water is critical for cooling the servers that run AI. A 2023 study highlighted that training GPT-3 in Microsoft’s U.S. data centers consumed about 700,000 liters of clean freshwateron-site ​(evaporated for cooling), and a total of 5.4 million L including indirect water (power generation)​. Moreover, inference “drinks” water too: depending on location, each 20–50 questions to ChatGPT might consume about 500 mL of water through cooling needs​. These numbers came as a surprise to many, leading researchers to call AI’s water footprint a “secret” cost​. Data centers in the US already withdraw billions of liters annually – Google’s centers used an estimated 12.7 billion L in 2021​– and AI is a key driver of growing demand​. By comparison, humans “use” water mainly in the supply chain of their activities (electricity, manufacturing goods, etc.). A human office worker might indirectly use a few cubic meters of water per year for their share of electricity and coffee, whereas a single AI training run can evaporate many cubic meters in days. Water scarcity is location-specific, so an AI trained in a drought-prone region has a larger sustainability impact. Companies like Google, Amazon, and Meta have pledged to become “water positive” by 2030, replenishing more water than they use​. This indicates a recognition that AI’s water use must be managed alongside energy.

Manufacturing & E-waste: One often ignored comparison is the environmental cost of manufacturing AI hardware vs. supporting humans. Training cutting-edge AI requires thousands of specialized chips (GPUs/TPUs). Manufacturing these chips and servers has a carbon footprint and resource cost: mining rare materials, fabrication in chip factories, etc. A recent MIT study noted the embodied carbon in a single high-end GPU is significant, and large AI clusters can have tons of electronics that eventually become e-waste​. By contrast, the “hardware” for humans is just our bodies and some computers/desks – far less complex. Of course, humans need infrastructure too (buildings, schools, transportation), which has its own embodied environmental impact. But on a per-worker basis, this is spread over decades. The lifespan of AI hardware might be only ~3-5 years before upgrade, meaning a cycle of e-waste that must be recycled or disposed. Efforts like Green AI and efficient hardware design (e.g. Google using liquid cooling and energy-efficient TPUs) are aiming to reduce the footprint per computation​. Additionally, strategies like using sparsely activated models (which use only parts of the network for any given task) can cut energy use by 10× while retaining performance​ – an active research area to make AI more sustainable.

Economic Sustainability (Value vs. Cost): Despite the above costs, the economic argument for AI is strong when looking at output. Numerous analyses show that, after the hefty initial investment, AI can generate significant value. A 2023 McKinsey report estimated that generative AI could add $2.6–4.4 trillion per year to the global economy across 63 use cases​. That’s a 15–40% increase on top of the value from non-generative AI, effectively doubling AI’s economic impact when GenAI is fully adopted​. In terms of ROI, an IDC global study in 2023 (of ~2,000 businesses) found that for every $1 companies invest in AI, they are seeing about $3.5–4 in returns on average​. Financial services reported even higher (~4.2× ROI)​. This suggests that, economically, AI can be self-sustaining or better – the revenue or savings generated outweigh the costs of development and deployment. For example, OpenAI’s GPT-4 might have cost ~$100M+ to build, but Microsoft’s $10B investment and the proliferation of GPT-based services imply a belief that it will return many-fold that in value (through advanced products, efficiencies, etc.). By comparison, human labor generally scales linearly with cost – 10× more work usually means ~10× more salary expense. AI, however, has a high fixed cost but low marginal cost: once you have a model, using it 10× more only slightly increases cost (mostly energy). In terms of AI as a work product creator, we already see businesses using generative AI to produce content, software, designs, even artwork, and monetize those products. Entire new services (AI-powered chatbots, copilots, image generators for media) are generating revenue streams. The sustainability of this from a business perspective hinges on maintaining low variable costs (compute) relative to output value. Cloud providers are addressing this by offering specialized AI chips and optimizing datacenter cooling to cut ongoing costs. Google even uses AI to optimize its data center energy usage, creating a virtuous loop where AI makes running AI cheaper and greener​.

Comparative Perspective: When considering training an AI model vs. training a human for a job, the investments are different in nature. Training a human involves years of education, mentorship, and experience – a process powered by our societal infrastructure. Its “energy cost” is diffuse (schools, food, etc. over decades) and the reward is a versatile brain that can adapt to many tasks (but only one brain per person). Training an AI is an upfront concentrated burst of energy and capital, resulting in a system that may exceed human ability in narrow domains but lacks adaptability beyond what it was trained for. One could argue that as AI gets more general, the cost to train a top AI might approach the cost (financial and environmental) of raising and educating a top expert, but that AI expert can then serve millions at once. In cases like AlphaFold, we see that dynamic: it took a huge compute investment akin to training many scientists, but now acts as a multiplier for all biologists. On the other hand, humans are still more energy-efficient in absolute terms for many tasks. The human brain runs on ~20 W – nature’s 86 billion-neuron “model” is amazingly efficient compared to a 175B-parameter transformer needing megawatts. In scenarios where one doesn’t need massive scale, a few humans might be the more sustainable option. For instance, if you only need ten well-written articles, hiring a human writer might have virtually no environmental impact compared to firing up a giant model. There’s a breakeven point: AI’s environmental ROI improves as task scale increases.

Total Environmental Impact: Summing up, AI’s total environmental impact includes direct factors (electricity, water, hardware) and indirect factors (the consequences of what AI enables). Training and running large AI models currently tend to increase a company’s carbon footprint in the short term​, unless mitigated, simply due to the energy intensity. If AI is used to replace or reduce human activity in a process, we then must consider what human activity is avoided. In some cases, AI might help cut overall emissions – e.g., AI-optimized logistics can reduce truck fuel usage (a clear net positive) or AI might help design more energy-efficient materials. These positive use-cases can offset the AI’s own footprint many times over. A consulting study by PwC projected that AI applications in energy, transportation, and industry could help reduce global greenhouse gases by 4% by 2030, even accounting for AI’s footprint​. So, context matters: AI as a tool for climate solutions (smart grids, climate modeling, etc.) can yield a net environmental benefit, whereas AI purely as a generator of endless content can seem less justified in sustainability terms. Economic sustainability is closely tied to this – if AI is creating genuine economic value (productivity, new capabilities), some of that value can be reinvested in greener infrastructure (e.g., using AI profits to build solar farms to power data centers).

Apples-to-Apples Conclusion: In an equivalent work comparison, AI generally uses more energy and resources upfront than humans to achieve a similar outcome, but then can scale to produce vastly more output with lower marginal cost. Humans use relatively little energy and water in performing knowledge work, but their throughput is limited and cost scales with quantity. From an environmental perspective, a single human brain is “greener” than a single massive AI for one unit of task; however, to match an AI that can do 1 million tasks, you’d need an army of humans whose combined footprint (millions of computers, offices, and lives) might then rival or exceed the AI’s footprint. The studies reviewed (2022–2025) suggest that for large-scale problems, AI can be more economically sustainable (in dollars output vs input), but we must work to make it more environmentally sustainable by improving energy efficiency and using clean power. There is active research into “Green AI” – algorithms that require less computation – and life-cycle assessments of AI to fully capture training, deployment, and disposal impacts​. The human vs AI balance will likely co-evolve: humans will focus on high-level judgment, creativity, and oversight, while delegating high-volume, repetitive, or extremely complex tasks to AI. The key is ensuring the benefits created by AI outweigh the resources it consumes. If, for example, an AI system costs $10M and 500 tCO₂ to develop but helps cure a disease or save $100M in efficiency, one can argue it’s a worthwhile trade-off – especially if future iterations can be done with less carbon. In summary, AI is an investment of resources up front to hopefully reduce total resources used in the long run. Achieving true sustainability will require continued innovation in how we build and power AI, and thoughtful use of AI where it makes economic and environmental sense.

References (2019–2025 Studies and Reports)

• Patterson et al. (2021) – “Carbon Emissions and Large Neural Network Training.” arXiv (Apr 2021).
  Link: arxiv.org/abs/2104.10350
  Summary: Detailed study by Google and academic researchers quantifying energy use and CO₂ emissions for training several large AI models (T5, Meena, GPT-3, etc.). Found GPT-3 consumed ~1,287 MWh, emitting ~502 tCO₂ (assuming average US grid mix). Proposes strategies for reducing carbon footprint (e.g. scheduling workloads in low-carbon regions, using efficient hardware). Quality: 90% (High – authored by leading experts, rigorous data, not peer-reviewed formally but well-regarded).

• Li et al. (2023) – “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models.” arXiv preprint (v4 Jan 2025).
  Link: arxiv.org/pdf/2304.03271 (UC Riverside & UT Arlington)
  Summary: First comprehensive look at water consumption in AI model training and inference. Uses GPT-3 as a case study: estimates 5.4 million liters water consumed in training (direct + indirect), and 0.5L of water per 10–50 queries during use due to data center cooling needs. Calls for transparency and “water-efficient” AI scheduling (training when/where water stress is low). Quality: 80% (Good – solid methodology and novel findings, but as a preprint it hasn’t undergone full peer review yet).

• Lannelongue & Inouye (2023) – “Environmental Impacts of Machine Learning Applications in Protein Science.”Cold Spring Harbor Perspectives in Biology 15(12): a041473.
  Link: doi.org/10.1101/cshperspect.a041473
  Summary: Examines the carbon footprint of computational biology tools. Reports that large models for protein folding (like AlphaFold2) and protein NLP (ESMFold) can each have >100 tonnes CO₂e emissions in training. Emphasizes need for sustainable computing practices in science. Provides comparisons with other protein simulation methods. Quality: 95% (Excellent – peer-reviewed article by Cambridge researchers, thorough and authoritative on the niche topic of ML in biology).

• Noble & Berry (2024) – “Power-hungry AI is driving a surge in tech giants’ carbon emissions – nobody knows how to fully offset it.” The Conversation (May 13, 2024).
  Link: theconversation.com/power-hungry-ai... (Institute for Sustainable Futures, UTS Sydney)
  Summary: Analysis of how the AI boom is impacting energy grids and emissions. Notes that data centers use ~7,100 L of water per MWh and cites increases in Google’s and Microsoft’s water and energy use due to AI. Discusses survey results: 72% of IT managers are concerned about AI energy consumption. Urges better reporting from data center operators. Quality: 85% (High – written by sustainability researchers, uses up-to-date data and surveys; not a formal study but informative and credible).

• Muvija M. (2025) – “Google develops AI co-scientist to aid researchers.” Reuters News (Feb 19, 2025).
  Link: reuters.com/technology/google-develops-ai-co-scientist-2025-02-19/
  Summary: News report on Google’s unveiling of an “AI co-scientist” tool for biomedical research. Tested at Stanford and Imperial College, it uses advanced reasoning on scientific literature to propose hypotheses. In one experiment (liver fibrosis), all AI-suggested approaches showed promising results. Google states it’s meant to augment human researchers, not replace them, highlighting collaboration. Quality: 90% (Reliable – Reuters reporting with quotes from Google and scientists; limited technical detail but factual).

• Ernst & Young (2022) – AI Document Review Study (via LexCheck).
  Link: LexCheck Blog – “Benefits of Relying on AI Document Review…” (blog.lexcheck.com)
  Summary: Cites an internal EY study finding that deploying AI-based document intelligence in legal review cut processing time by ~90% and costs by ~80%, while also improving accuracy by 25%. Demonstrates the efficiency gains and financial savings from AI in document review workflows. Quality: 85% (Good – EY is a reputable firm, though the exact report isn’t public; data is believable and aligns with other case studies).

• McKinsey Global Institute (2023) – “The economic potential of generative AI: The next productivity frontier.”McKinsey & Co. (June 2023).
  Link: mckinsey.com – Our Insights (Generative AI report)
  Summary: Extensive research report estimating generative AI’s impact on economy and labor. Key figure: GenAI could add $2.6–4.4 trillion in value annually across industries (on top of other AI). Notes greatest potential in functions like customer operations, marketing/sales, and software engineering. Discusses how 50%+ of work activities could be automated to some degree. Quality: 80% (High – thorough analysis by a top consulting firm, based on modeling and expert input; not academic peer-review, but widely cited and vetted internally).

• World Economic Forum (2024) – “AI and energy: Will AI reduce emissions or increase demand?” WEF article (July 12, 2024).
  Link: weforum.org/stories/2024/07/generative-ai-energy-emissions/
  Summary: Examines the paradox of AI’s environmental impact. Reports that AI might use 33× more energy to do the same task as specialized software​. Notes Microsoft’s CO₂ up 30% (2020–22) and Google’s up 50% (2019–23) due to AI data center growth​. Also highlights opportunities: AI could also optimize energy use (mentions Google using AI to reduce data center cooling energy by 30%). Concludes that AI’s net effect on emissions will depend on mitigation and use-case (good vs bad). Quality: 80% (Good – an informed summary with data from credible sources (OECD, arXiv, etc.), though as a magazine-style piece it simplifies some assumptions).

• Microsoft & IDC (2023) – “New study validates the business value and opportunity of AI.” Microsoft Official Blog (Nov 2, 2023).
  Link: blogs.microsoft.com (commissioned IDC study results)
  Summary: Presents findings from an IDC survey of 2,000+ companies on AI ROI. Key stats: 71% of companies are using AI in some form; 92% of AI deployments reached production in under 12 months; average ROI is $3.5 per $1 invested, achieved within ~14 months​. Predicts generative AI could add $10 trillion to GDP over 10 years​. Underscores that rapid adoption is occurring and those who invest early see significant returns. Quality: 85% (High – while on Microsoft’s blog, it references a robust IDC study; the statistics are recent and from a large sample, likely reliable).