Removing “please”, “thank you” and other unnecessary words from AI prompts could save 87 to 98 gigawatt‑hours of electricity per year, according to a new UN University Institute for Water, Environment and Health (UNU‑INWEH) report. That’s roughly the annual residential electricity use of up to 760,000 people in sub‑Saharan Africa, and it underscores why AI prompt efficiency matters more than most developers realise.
Key Takeaways
- Shorter prompts can shave up to 25 percent off ChatGPT’s energy consumption.
- AI now accounts for about 20 percent of data‑centre power, a share that could double by 2030.
- Generating images uses 60 times more energy than a text query; complex video can be 8,000 times more.
- UN researchers urge public education, mandatory energy disclosures, and caps on AI usage.
- Developers can cut emissions by writing concise prompts, avoiding conversational loops, and opting for smaller models when possible.
AI Prompt Efficiency: Cutting Politeness to Cut Energy Use
Madani and his colleagues aren’t asking us to be rude to our digital assistants, but they are warning against what they call the “interaction trap.” In the report, they show that every extra word adds tokens, and every token forces the model to fire more transistors. When you trim the fluff, you cut both the number of tokens the model has to ingest and the number it spits out in response. That double‑reduction can translate into measurable power savings, especially at the scale of billions of daily queries.
“We are not saying be rude to your AI. But don’t fall into the interaction trap and don’t go falling in love with it either,” says Kaveh Madani at UNU‑INWEH.
Large language models break text into tokens that are usually a few characters long. A concise prompt might be 10‑15 tokens, whereas a polite, conversational request could balloon to 30‑40. The extra processing isn’t trivial – each token requires a forward pass through the neural network, which consumes electricity. In some cases, a shorter prompt also simplifies the underlying task, meaning the model can generate an answer with fewer layers of computation.
Why Tokens Matter for Data‑Centre Power
ChatGPT alone processes around 2.5 billion queries every day, while Google handles roughly 16 billion. Those figures aren’t just vanity metrics; they’re the raw material that powers the UN’s energy calculations. Because tech firms rarely disclose exact power draws, the researchers had to piece together public data‑centre specs and extrapolate from known hardware efficiencies.
“You’re looking at something on a global scale that is being adopted faster than any other technology in the history of technology, so the energy use is increasing very rapidly,” says Miriam Aczel at UNU‑INWEH.
The study estimates that AI already consumes about 20 percent of data‑centre electricity, a share poised to climb to 40 percent in the next few years. By 2030, AI’s share could amount to 378 terawatt‑hours a year, while total data‑centre demand might reach 945 TWh – roughly three per cent of projected global electricity use.
Policy and Industry Recommendations
Madani’s team argues that the fastest way to curb this growth isn’t to ban AI, but to make its use more transparent and disciplined. They call for mandatory energy‑consumption reporting from AI firms, reminiscent of climate‑risk disclosures in the finance sector. Governments, they say, should consider caps on both corporate and individual AI usage until reliable metrics are in place.
Public education is another pillar of their proposal. If users understand that a polite “please” can cost an extra few watt‑hours, they’ll be more likely to trim it. The researchers also suggest that developers provide low‑power alternatives – for example, offering a “lite” model for simple text tasks instead of defaulting to the most powerful, energy‑hungry engine.
Historical Context
When the first transformer‑based language models appeared, token counts were modest and hardware footprints were small. The leap to the transformer‑heavy architectures that dominate today brought a proportional rise in per‑query compute. Early papers highlighted the trade‑off between model size and performance, but few quantified the electricity bill attached to each extra token. The UN report builds on that gap, turning abstract efficiency discussions into concrete gigawatt‑hour numbers.
At the same time, data‑centre design has shifted from a focus on raw throughput to a broader sustainability agenda. Operators now track Power Usage Effectiveness (PUE) and invest in cooling innovations, yet the surge in AI workloads can outpace those gains. By anchoring the conversation in token‑level economics, the report gives operators a new lever: ask users to be concise, and the hardware can stay cooler while delivering the same answers.
Previous industry‑wide audits hinted at rising AI electricity draws, but they lacked the granularity to link everyday phrasing to power consumption. The new study’s method of mapping word count to transistor activity bridges that gap, showing that a simple linguistic habit can ripple through the global power grid.
Practical Steps for Developers
From a developer’s perspective, the guidance is straightforward: write concise prompts, avoid unnecessary conversational loops, and pick the smallest model that meets your accuracy needs. When you do need richer interactions, consider caching frequent responses or pre‑generating templates to reduce repeated token processing.
Energy of Text vs Images
Generating an image consumes about 60 times more energy than a text query, enough to power a 10‑watt LED bulb for roughly 17 minutes. A complex video can be 8,000 times more energy‑intensive, lighting that same bulb for about 1.7 days. Those numbers make it clear why the UN’s report urges users to reserve visual generation for cases where it truly adds value, rather than treating it as a default output mode.
- Concise prompts cut token count, reducing both compute and energy.
- Avoiding endless back‑and‑forth loops prevents runaway power draw.
- Choosing smaller models for routine tasks can halve the carbon footprint.
- Limiting image or video generation to essential use‑cases saves orders of magnitude in energy.
What This Means For You
If you’re building a chatbot, start by auditing your prompt templates. Strip out every superfluous word – “please” and “thanks” are polite, but they’re also expendable tokens. Test the same request with a stripped‑down version and compare latency; you’ll often find the leaner prompt runs faster and cheaper.
For product managers, the takeaway is to educate your users. Show them a side‑by‑side comparison of a polite versus a concise request, and let the energy savings do the talking. When you embed AI into SaaS tools, consider offering a toggle that switches the backend to a smaller model for low‑stakes queries – you’ll be lowering your carbon bill without sacrificing core functionality.
Startup founders can weave prompt efficiency into their go‑to‑market narrative. Pitch investors with a concrete figure: a 25 percent reduction in token usage translates into measurable cost savings at scale. Those numbers resonate with anyone watching operating expenses, and they also position the company as environmentally aware.
Architects of internal tooling should revisit any “assist‑me” features that automatically prepend polite phrasing. A tiny change in the codebase – removing a static “please” – can shave megajoules off each request. Multiply that by millions of daily interactions, and the aggregate impact aligns with the UN’s 87‑to‑98 GWh estimate.
As AI continues to infiltrate every corner of the digital economy, the question isn’t whether we can afford to be less polite, but whether we can afford to stay as polite as we are now without draining the planet.
Key Questions Remaining
How will mandatory energy disclosures shape the competitive landscape? Will vendors rush to certify “low‑token” models, or will they bundle efficiency metrics into existing service‑level agreements?
What standards will emerge for measuring token‑level power draw? The report hints at the need for a common methodology, but industry consensus remains elusive.
Can user‑facing interfaces be designed to nudge people toward brevity without sacrificing user experience? Balancing friendliness with efficiency will likely become a design discipline of its own.
Will regulatory caps on AI usage prove practical, or will they drive innovation toward more compact architectures? The answers will unfold as policymakers and technologists iterate on the recommendations.
Sources: New Scientist Tech, original report
