The UK government's AI Knowledge Hub has published draft guidance for staff on using AI "ethically and sustainably." Among its examples of ways to reduce interactions with AI, it says there is no need to send an extra message just to say thank you.
But the document's main point is not to change how people are polite to AI.
First consider whether AI is needed at all. If it is, choose the smallest model that can still do the job, and give clear instructions to cut unnecessary back-and-forth. The draft is an attempt to translate the environmental impact of AI use into the everyday actions of staff.
The draft does not, however, include any measurements of how much electricity or water would be saved by omitting a single "thank you."
Has thanking AI been banned?
The AI Knowledge Hub page "Using AI ethically and sustainably" is clearly marked "draft" at the top.
Its disclaimer also says it is not a comprehensive guide or legal advice, but a starting point for thinking about AI use. Staff are asked to check the AI policies set by their own organisations and follow those.
It would therefore be inaccurate to read the document as a new rule in which the UK government has banned all staff from thanking AI.
The mention of "thank you" that drew attention appears in a section on using AI efficiently.
That section recommends practices such as:
- Keeping prompts short, clear and direct
- Using as few prompts as possible
- Using AI only when needed
It gives "there is no need to say thank you" as a concrete example.
In other words, the idea is not to treat greetings or courtesy as a problem, but to cut unnecessary exchanges, such as sending a bare "thank you" after a task is finished and triggering yet another AI response.
The draft does not give figures for how much power or water use would fall from skipping that one message.
Some items also take priority over reducing interactions.
The draft points out that AI can generate plausible-looking misinformation or biased answers, and asks staff to verify figures and sources and to check for fairness and accessibility.
When content has been created or edited with AI, that fact should be disclosed, and people must be able to explain the final decisions themselves.
The advice to reduce prompts does not mean skipping necessary checks.
First, ask whether AI is needed
The first thing the draft asks is not that staff pick a smaller model.
It asks whether AI is needed for the task in the first place.
For example, if a spreadsheet or an ordinary search engine is enough, it recommends considering not using AI.
If AI is needed and the service lets you choose a model, use "the smallest model that can do the job."
Its examples include choosing Gemini Flash rather than Gemini Pro, and GPT Instant rather than GPT Thinking. It also states that with Microsoft Copilot, users cannot change the model themselves.
The key point is that using a small model is not itself the goal.
If a model lacks the capability and users must re-ask questions repeatedly or make extensive manual corrections, total processing could end up increasing.
The draft accordingly sets a condition: not simply "the smallest model," but "the smallest model that can do the job."
The same applies to shortening prompts.
If necessary conditions are cut and vague instructions have to be redone repeatedly, shortening the first message does not make anything more efficient.
Indeed, another section of the draft notes that AI can give different answers to the same question, so humans need to verify sources and figures.
To judge efficiency in practice, it is closer to reality to look at how many exchanges it took to finish the work, including necessary checks, than at the character count of a single prompt.
Information protection and human responsibility come before environmental impact
Government work also places limits on which AI can be used and what information can be entered.
The draft asks staff to use the AI systems provided by their departments, and not to enter personal information or information that could identify individuals.
It also warns that some AI services may use inputs for training.
So before comparing which model is most energy-efficient, it is necessary to decide:
- Whether the tool may be used for work
- What information may be entered
- How outputs will be checked
The draft's environmental considerations are not rules that override security, accuracy or accountability.
The UK government's AI environmental measures are not new
This is not the first time the UK government has addressed the environmental impact of AI.
The "AI Playbook for the UK Government," published in February 2025, asks departments to assess the environmental impact of training and operation before adopting AI, and to consider whether the same or similar results can be achieved with less energy.
The "Data and AI Ethics Framework" likewise does not treat AI's environmental impact as a matter of electricity alone.
It asks that the whole lifecycle of an AI system be considered, including water used in data centres, raw materials needed for chips and equipment, and electronic waste from discarded hardware.
The new AI Knowledge Hub draft can be seen as turning these existing principles into concrete actions for staff who use AI day to day.
| Publication | Published / updated | When it applies | Main content |
|---|---|---|---|
| AI Knowledge Hub "Using AI ethically and sustainably" | No publication date given (checked September 29, 2026) | When staff use AI day to day | Consider whether AI is needed, choose an appropriate model, reduce unnecessary exchanges |
| AI Knowledge Hub "Building AI sustainably" | No publication date given (checked the same day) | When building and operating AI systems | Consider approach, suppliers, computing facilities and data storage, and measure environmental impact on an ongoing basis |
| AI Playbook for the UK Government | Published February 10, 2025 | When planning AI adoption | Weigh benefits against environmental impact, consider lower-impact methods and suppliers |
| Data and AI Ethics Framework | Updated December 18, 2025 | Across the data and AI lifecycle | Consider electricity, water, raw materials, electronic waste and more |
The materials aimed at developers go beyond model size.
For example, they include asking suppliers about their use of renewable energy and disclosure of environmental information, choosing computing facilities suited to the processing required, and not storing unnecessary data for longer than needed.
The Data and AI Ethics Framework also points out that treating electricity use as the sole indicator of environmental impact risks overlooking water use, raw materials and effects on local communities.
In other words, a measure such as cutting one "thank you" is only a small part of the overall approach to AI's environmental impact.
Decisions about which AI services to procure, which data centres and equipment to run them on, and which power to use cannot be changed by individual staff simply shortening their prompts.
AI power use is rising, but the effect of one "thank you" is unknown
According to the International Energy Agency's 2026 "Key Questions on Energy and AI," global data centre electricity consumption is projected to nearly double from about 485 TWh in 2025 to about 950 TWh in 2030.
Electricity consumption by AI-focused data centres is expected to grow even faster.
At the same time, the IEA notes that the energy needed for a single AI task is improving rapidly.
For simple text generation, efficiency gains in models and hardware have sharply reduced per-task energy use over the past few years.
The problem is that the number of users and uses is growing at the same time, and more computation-heavy applications are spreading.
The IEA explains that models that do a lot of reasoning, AI agents and video generation can use far more energy per task than simple text generation.
For that reason, it is hard to describe current AI as a whole with a single figure for how much electricity one use consumes.
The same goes for a short "thank you" message.
Sending it triggers additional inference, so it is plausible that it increases computation compared with sending nothing.
But the UK government's draft includes no measured figures for the electricity or water use of that one message.
Even small models can use more, depending on how they are used
When comparing the environmental impact of models, model size alone is not enough to go on.
Hugging Face's "AI Energy Score" compares the inference-time energy consumption of publicly available models using the same evaluation method.
The second version, released in 2025, reported that models with reasoning enabled used substantially more energy on average than standard models.
A major factor in that gap is the number of tokens generated during reasoning.
Even with the same model, energy use rises if it reasons for longer and generates a large number of tokens.
However, the AI Energy Score mainly evaluates models with publicly available weights.
It does not compare commercial services such as Gemini, GPT and Microsoft Copilot, cited as examples in the UK draft, under the same conditions, nor does it show that for a given task one will always save a certain percentage of energy.
Moreover, even if the energy used for a single response were known, that alone would not indicate the environmental impact of the work as a whole.
If switching to a smaller model lowers accuracy and requires three re-asks, it must be compared with finishing in one go on a larger model.
The extra time people spend checking and correcting also needs to be factored in.
In that sense, it matters that the draft says not "a small model" but "the smallest model that can do the job."
Can the whole job be measured, rather than just skipping "thank you"?
The draft does not include a publication date, an effective date, or information on how far staff are following this advice.
Nor does it estimate how much electricity or water the UK government as a whole could save by skipping additional "thank you" messages.
At this point, therefore, it cannot be assessed as delivering a major environmental benefit from not thanking AI.
What the draft sets out is a broader way of thinking.
Don't use AI for work that can be done without it. If it is needed, choose a lower-impact model among those capable enough for the task. Convey the necessary information clearly and avoid adding pointless exchanges.
And don't cut processing at the expense of accuracy or safety.
To evaluate real environmental effects, it is necessary to look beyond the length of a single prompt to the total cost of finishing the work, including the model used, the amount of processing, the quality of answers, and checking and retries.
The small habit of not thanking AI drew attention, but that is not the essence of the draft. If AI is to be built into daily work, it asks that people use it with awareness not just of convenience but of necessity and computing resources, and it tries to make that thinking concrete in staff's everyday actions.