AI, energy and water: What the assessment industry should be worried about

Separating the headlines from data: what responsible AI use looks like for high-stakes assessment

The environmental cost of AI has become a key concern raised by clients, regulators and the assessment community. It is also one of the most poorly understood topics. Numbers are often quoted from headlines that misread the underlying research, and AI is often compared with nothing, rather than comparing it to digital activities people already do without a second thought.

The goal of this article is to lay out what the current research actually shows, where the legitimate concerns lie, and why we believe responsible AI adoption is more environmentally credible than the alternative of inaction. It also explains how Surpass is approaching this as a supplier, including the offsetting and governance commitments we already have in place.

The per-prompt numbers are smaller than the headlines suggest

Two recent disclosures from AI providers themselves, with published methodology, have shifted what we can credibly say.

Google’s August 2025 technical paper, with published methodology, puts a median Gemini text prompt at 0.24 Wh of energy and 0.26 mL of water, covering the full stack including idle capacity and data centre overhead.¹ OpenAI disclosed in June 2025 that the average ChatGPT query uses 0.34 Wh and roughly one-fifteenth of a teaspoon of water.²

Independent academics, notably the University of California Riverside team behind “Making AI Less Thirsty”³, argue these figures understate indirect water consumption from electricity generation. Including those scopes, a realistic per-query range sits at 10 to 25 mL. The honest answer is somewhere between operator disclosures and academic worst-case estimates, but the order of magnitude is consistent: per-prompt impacts are small.

What matters more than any single snapshot is the direction of travel. Google reports a 33x reduction in energy and a 44x reduction in carbon footprint for the median Gemini prompt over a single year.¹ Efficiency is improving faster than usage is growing, at least for standard text models. Reasoning models are a separate and growing question, which we address below.

“Our position is that the responsible answer to AI’s environmental footprint is not to avoid AI. It is to govern it well, measure it honestly, and offset what we can’t yet eliminate.”

Andy McAnulla, Co-CEO, Surpass Assessment

Compared with everyday digital activity, AI looks ordinary

The most useful test isn’t whether AI uses resources. Everything does. It’s whether AI use is materially different from activities we already accept. The table below sets out the credible per-unit figures, drawn from the IEA, Google, OpenAI, EPRI and published academic work.

Activity (per hour or per unit)	Energy	Water
1 hour HD video streaming (Netflix / YouTube)	~80–150 Wh	~1000–15,000 mL (estimates vary widely)
1 hour video conferencing	~150 Wh	2000-12,000 mL
1 Google search	~0.3 Wh	~0.6 mL
1 Gemini text prompt (Google, 2025)	0.24 Wh	0.26 mL (~5 drops)
1 ChatGPT prompt (OpenAI, 2025)	0.34 Wh	0.32 mL (direct cooling)
1 ChatGPT prompt (UCR, full scope)	0.3–3 Wh	10–25 mL

Sources: Google Research ¹; OpenAI ²; Li et al ³; IEA and Data Centres and Data Transmission Networks ⁴; EPRI ⁵. Streaming figures vary widely with device, network and resolution; the ranges shown reflect that.

Put plainly, in energy terms, one hour of streaming uses roughly the same as 200 to 500 AI prompts, depending on device, network and resolution. A single prompt costs around 10 seconds of streaming.⁴

Water comparisons against streaming are contested in the literature, so the more useful anchor is something most readers do without thinking. A single cup of coffee carries around 130 litres of embedded water to grow, process and ship the beans.⁶ Even taking the worst-case academic estimate of 25 mL per AI prompt, that single cup of coffee is the water equivalent of around 5,600 prompts. On Google’s and OpenAI’s operator-disclosed figures, it is more than half a million. A typical knowledge worker’s daily AI usage, even with heavy reliance, sits well below the footprint of a single video meeting and a long way below the cup of coffee on the desk next to it ⁷. The marginal environmental cost of moving common professional tasks from manual to AI-assisted is, in most cases, neutral or favourable.

What should the real concerns be?

None of this means there is nothing to worry about. The substantive issues are:

Aggregate infrastructure growth. Global data centre electricity consumption reached around 415 TWh in 2024 (1.5% of global use) and is projected to roughly double by 2030. That growth is real and largely AI-driven. It is a question about grid planning, water rights and infrastructure location, not individual user behaviour.
Reasoning and multimodal models. Reasoning models such as o3 and DeepSeek-R1 consume significantly more energy on average than standard text models.⁸ Video and image generation are higher again. Industry impact projections built on standard text prompts will understate the trajectory.
Measurement opacity. Most operators do not yet publish like-for-like figures. Direct cooling, indirect electricity generation, training versus inference, and end-device load are inconsistently reported. This is where industry concern should focus.

The Surpass approach

Our position is that the responsible answer to AI’s environmental footprint is not to avoid AI. It is to govern it well, measure it honestly, and offset what we can’t yet eliminate. Four principles underpin this.

ISO/IEC 42001 certification for AI management In May 2026, Surpass became the first end-to-end assessment platform to achieve ISO/IEC 42001 certification, the international standard for Artificial Intelligence Management Systems. The certification, awarded after an independent third-party audit, confirms that we have structured policies, controls and oversight processes for AI risk, fairness, transparency, and reliability. For our customers, this is independent validation that AI in the Surpass Platform is managed within a recognised framework, not an internal claim.
A published AI Policy with environmental responsibility built in Our AI Policy commits us to Responsible AI Principles that include societal responsibility, explicitly covering environmental impact alongside fairness, transparency and security. It directs employees to use AI purposefully and efficiently, and to treat AI as one tool among several rather than a default. Governance sits with the Compliance Group and an external Surpass Community AI Advisory Board, both of which review AI implications carefully and regularly.⁹
Carbon offsetting that already covers our AI footprint Surpass measures GHG emissions under the GHG Protocol and uses ClimateCare to offset our carbon footprint. Our operations run on 100% renewable energy through our physical office supply contract. Because our AI consumption sits within our overall corporate energy and emissions footprint, it is already inside the scope of our existing offsetting programme. This is a deliberate position. We don’t treat AI as an environmental category requiring a separate excuse; we treat it as part of our total operational footprint and offset accordingly. As AI usage scales across our platform and our internal operations, the offsetting envelope scales with it. We will continue to report on this through our ESG framework.
Responsible procurement of AI infrastructure The largest lever a company of our size can pull is who we purchase from. Surpass hosts on Microsoft Azure, which has public commitments to be carbon negative by 2030, water positive by 2030, and zero waste by 2030, with annual public reporting.¹⁰ Our approved AI tooling sits within enterprise agreements that inherit those commitments. This isn’t a substitute for our own measurement, but it means the data centres serving our AI workloads operate under some of the most aggressive sustainability commitments in the industry.

What this means for the assessment industry

There are three implications worth taking seriously:

The “AI is environmentally irresponsible” objection should not, by itself, slow adoption. It echoes the early-2020s video streaming panic that the International Energy Agency (IEA) had to publicly correct. The numbers say AI use, at current scale and trajectory, is comparable to or below other accepted digital activity. The legitimate concerns are infrastructure and measurement, not the number of individual prompts.
Transparency is becoming a procurement criterion. Awarding and certification bodies, regulators and licensure clients will increasingly ask suppliers what they measure, how they offset, and what governance sits behind their AI claims. Vendors who cannot answer specifically will look evasive. Vendors who can answer will gain trust.
The right standard is measured and offset, not avoided. AI brings genuine improvements to assessment quality, accessibility, security and cost. Refusing to adopt it on environmental grounds, while continuing to use video conferencing and streaming, is not a coherent position. Adopting it responsibly, with measurement and offset, is.

Closing thought

The assessment industry has a long track record of taking responsibility seriously, on test security, fairness, candidate data, and accessibility. AI deserves the same treatment: clear principles, independent validation, honest measurement, and a willingness to offset what we cause. The conversation we should be having isn’t whether to use AI. It’s whether each of us can explain, with evidence, how we use it responsibly.

That’s the standard we hold ourselves to at Surpass, and the one we would encourage the wider industry to adopt.