Labor, Environmental, and Supply-Chain Impacts

The footprint of a system reaches well past the data center it runs in. Five stages carry it: annotation labor, workplace deployment, compute operations, hardware supply chain, rebound and scale. Most of them have already been measured by somebody.

Annotation labor: around three dozen Kenyan workers labeling descriptions of sexual abuse, hate speech and violence, at quotas of 150–250 passages per nine-hour shift. Compute operations: 176 TWh of US data-centre electricity in 2023, 4.4% of US electricity. Hardware supply chain: the distance between BLOOM's 24.7 tonnes and its 50.5 is what counting equipment manufacturing adds. Workplace deployment: a jurisdiction fight at Nairobi over who employs 185 former Facebook content moderators.

Rebound and scale is the stage nobody has a meter for. It is also the stage that drops off the list.

Efficiency per request can improve while the absolute footprint grows. Only a lifecycle view puts both numbers on the same page.

Running 1,000 inferences costs 0.002 kWh if the task is text classification. It costs 2.907 kWh if the task is image generation. That is a spread of over 1,450× inside the single column labelled joules per completed task. It is settled by which task you chose, not by how well you optimised it. Luccioni and colleagues measured it across 88 models in 2024.

The same study lays the efficiency metric alongside the absolute footprint. BLOOMz-7B would need 592,570,000 inferences before its deployment energy equalled the 59,257 kWh it took to train (51,686 kWh) and fine-tune (7,571 kWh) it. That sounds unreachable until you count users. The authors write: “Even assuming a single query per user, which is rarely the case, the energy costs of deploying it would surpass its training costs after a few weeks or months of deployment.” Training is a number you pay once. Serving is a number that keeps arriving.

Responsible AI includes labor and environmental conditions across data collection, annotation, content review. It includes them across hardware production, training, serving, maintenance, and disposal. Automation often redistributes work and cost rather than eliminating them. A lifecycle assessment should consider job quality, surveillance, exposure to harmful material, fair compensation, worker voice. It should then consider energy source, water use, utilization, embodied hardware impact, electronic waste, and rebound effects. Absolute impact and marginal product impact are different questions.

Training one Transformer (big) model emitted 192 lbs CO2e. Reaching that same architecture through neural architecture search emitted 626,155 lbs CO2e, at a cloud cost of $942,973–$3,201,722. That is roughly 3,260 times the published cost of the run itself. For scale, an average car emits 126,000 lbs including fuel over its whole lifetime.

Strubell and colleagues published both figures in 2019. They also stated what all that searching bought: “So et al. (2019) report that NAS achieves a new state-of-the-art BLEU score of 29.7 for English to German machine translation, an increase of just 0.1 BLEU at the cost of at least $150k in on-demand compute time and non-trivial carbon emissions.” Thousands of training runs' worth of compute. A tenth of a BLEU point.

Rebound effects close the list and are the item most often left off it. Cheaper moderation tends to buy more moderation, not less.

Around three dozen Kenyan workers read and labeled descriptions of sexual abuse, hate speech and violence, at quotas of 150–250 passages per nine-hour shift. They worked under three contracts worth about $200,000, signed by OpenAI with the outsourcing firm Sama in late 2021. TIME documented the arrangement on 18 January 2023: “The data labelers employed by Sama on behalf of OpenAI were paid a take-home wage of between around $1.32 and $2 per hour depending on seniority and performance.” Sama ended the work in February 2022, eight months early.

Who answers for such conditions is not a rhetorical question. It has been litigated. Meta argued that Kenya's Employment and Labour Relations Court had no jurisdiction over it. On 20 September 2024 the Court of Appeal at Nairobi dismissed that argument with costs. “The upshot of our above findings is that the appellants’ appeals numbers E232 of 2023 and E445 of 2023 are devoid of merit and both appeals are hereby dismissed with costs to the respondents,” the court held.

Daniel Motaung and the 185 former Facebook content moderators engaged through Sama may now proceed in that court. Motaung's petition covers working conditions, unfair labour practices and constitutional violations. The moderators have a separate constitutional petition over their mass redundancy.

Three of these drills count things. The fourth asks the product question nobody schedules: whether this feature should generate anything at all.

The vendor-evidence drill is no longer a request for goodwill. The EU AI Act requires a general-purpose AI model provider's technical documentation to record the “known or estimated energy consumption of the model. With regard to point (e), where the energy consumption of the model is unknown, the energy consumption may be based on information about computational resources used.” The law has been in force since 1 August 2024, and it supplies its own fallback method. A provider who says the figure is unavailable is answering a question the law has already anticipated.

The lifecycle traced here includes contractors and disposal. That is exactly where most published impact figures stop.

Step one is where the omission usually happens. The paper announcing a model reports the run, not the neural architecture search that found it. That is how 192 lbs CO2e and 626,155 lbs CO2e come to describe the same architecture.

Step four is what a court ends up doing when nobody wrote the protections in advance. Meta's position was that the Employment and Labour Relations Court had no jurisdiction over it. On 20 September 2024 that argument was dismissed with costs.

Step five now has a statutory calendar beside it. The AI Act obliges the Commission to report by 2 August 2028, and every four years thereafter, on progress on standardisation deliverables for the energy-efficient development of general-purpose AI models. The same regulation presumes the high-impact capabilities that trigger systemic-risk classification wherever training compute exceeds 10^25 floating point operations.

Efficiency gains that lower unit cost tend to raise total use. That is why the absolute number is the one worth reporting.

The United States has that claim as a measured national series rather than an assertion. Data-centre electricity use held near 60 TWh from 2014 to 2016, while efficiency work absorbed the growth. It reached 76 TWh in 2018 — 1.9% of US electricity. In 2023 it reached 176 TWh, or 4.4%. The projection for 2028 is 325–580 TWh, between 6.7% and 12.0%.

The numbers are Lawrence Berkeley National Laboratory's, published in 2024. Its executive summary names the break: “While many of these efficiency strategies continue to provide significant energy efficiency improvements in data center design and operation, the expansion of data center services into areas that require new types of hardware has ended the era of generally flat data center energy use.” Efficiency did not stop working. It stopped being enough to hold the total down.

Define when labor and environmental impact governance requires the team to redesign, restrict, remedy, or retire the system.