Ethics & Environmental Impact
- Environmental impact is primarily an electricity and infrastructure issue; water impacts are important but depend strongly on location, cooling technology, and what is included in the calculation.
- AI has biases - Watch for gender, racial, socioeconomic, and geographic biases; request diverse examples explicitly.
- Maintain human oversight - AI should enhance, not replace, your expertise and critical judgment.
- Consider ethical implications - Be mindful of data rights, labor practices, and the value of human communication.
We must be mindful of “environmental impacts, risks of bias and stereotyping, and ethical concerns about data privacy and security” when using AI tools.
Environmental Impact
AI is not an immaterial service: every interaction depends on physical infrastructure including data centres, specialised chips, networks, cooling systems, and electricity grids. The environmental impact of AI is best understood at system scale rather than by focusing only on the footprint of a single prompt (Prof. Chris Preist, Everything, Everywhere, All at Once: The Hidden Cost of AI, 2026). A ClimateAction.tech summary identifies electricity consumption as the standout concern, while water use and hardware production remain important secondary impacts.
Electricity: the main pressure
The International Energy Agency estimates that data centres used around 485 TWh of electricity in 2025 and projects roughly 950 TWh by 2030 in its central case, about 3% of global electricity demand. AI-focused data centres are growing faster than data centres overall. The global percentage can hide much larger local effects because data centres are geographically concentrated and can place substantial new demand on particular electricity grids (IEA 2026).
The impact of an individual AI request is not a fixed number. A 2025 production study by Google measured a median Gemini text prompt at 0.24 Wh, but energy use varies with model, prompt length, output length, hardware, and the type of task (Elsworth et al. 2025). Reasoning, agentic workflows, image generation, and video generation can be much more computationally intensive than a short text interaction. Efficiency per task is improving rapidly, but increased adoption and more demanding workloads can outweigh those gains.
Water: important, but highly context-dependent
AI-related water use can occur in several places:
- Directly at the data centre, for cooling and humidification.
- Indirectly through electricity generation, where power stations may withdraw or consume water.
- During hardware manufacturing, including semiconductor production.
These different boundaries matter. A widely cited study estimated that GPT-3 could consume about 500 ml of water for roughly 10–50 medium-length responses, depending on where and when it ran; the estimate includes both on-site and electricity-related water and should not be treated as a universal “water per prompt” figure (Li et al. 2025). By contrast, Google’s 2025 study reported 0.26 ml of on-site cooling water per median Gemini text prompt. The figures are not directly comparable because they concern different models, facilities, dates, and system boundaries.
Cooling technology also changes the picture. Modern data centres may use free-air cooling, adiabatic cooling, or closed-loop liquid cooling, which can greatly reduce on-site water use compared with older water-intensive cooling designs. Microsoft reports a global Water Usage Effectiveness (WUE) of 0.27 L/kWh for FY25 across the data centres it owns and controls, illustrating how operators now track water alongside energy efficiency (Microsoft datacentre efficiency).
A particularly relevant local example is Isambard-AI at the University of Bristol. This 5 MW national AI research facility uses direct liquid cooling and reports a Power Usage Effectiveness (PUE) of about 1.08, indicating that only a relatively small proportion of its electricity is used for facility overhead rather than computation (University of Bristol, 2026). Its design shows how newer AI infrastructure can reduce cooling and power overheads through tightly integrated engineering. It also provides a useful reminder that environmental impacts should be assessed using the characteristics of the actual facility, rather than extrapolating from older or less efficient data-centre designs.
Water demand that is manageable in one location can be damaging in a water-stressed catchment. Claims about AI water use should be interpreted carefully. Reported figures may refer to water withdrawal or water consumption, and may include only on-site cooling or also water used to generate electricity. Their significance also depends on the cooling technology, the local level of water stress, and whether the figures are measured directly, modelled, or extrapolated from other facilities.
Sustainable AI Practices
Individual interactions are usually small compared with the system as a whole, but billions of interactions and increasingly compute-intensive applications add up. Useful practices include:
Use the lightest tool that meets the need:
- Use simpler text interactions rather than reasoning, agentic, image, or video workflows when the extra capability is not needed.
- Choose smaller or task-specific models when they are adequate.
- Avoid repeated regeneration and reuse suitable outputs where possible.
Look for transparent infrastructure:
- Prefer providers that disclose energy use, electricity sources, PUE, WUE, and model-level resource information.
- Remember that “renewable energy” claims and water metrics can use different accounting boundaries; check what is actually being measured.
- Where you control infrastructure, consider grid carbon intensity, local water stress, and cooling design as well as model efficiency.
Keep the decision in context:
- The computational cost of AI should be considered alongside the value it adds to the task.
- Choosing not to use an LLM can be appropriate where the environmental or ethical costs outweigh the benefits.
For a detailed set of notes from the talk, see AI sweat: data centre emissions.
Bias and Fairness in AI
AI systems inherit and can amplify existing societal inequalities. The existing guidance on watching for gender, racial, socioeconomic, and geographic biases remains paramount.
Gender Bias:
- Associating certain professions with specific genders
- Using gendered language inappropriately
- Making assumptions about capabilities based on gender
Racial and Ethnic Bias:
- Stereotypical associations with names or cultural references
- Underrepresenting certain groups in examples
- Making assumptions about backgrounds or capabilities
Socioeconomic Bias:
- Assuming access to resources or opportunities
- Using examples that exclude certain economic backgrounds
- Privileging certain educational or professional experiences
Geographic Bias:
- Focusing on Western/English-speaking perspectives
- Making assumptions about local contexts
- Overlooking global south perspectives
When reviewing AI outputs, ask:
- Representation: Who is included and excluded in examples?
- Language: Are descriptions fair and respectful to all groups?
- Assumptions: What unstated assumptions are being made?
- Perspectives: Whose viewpoints are prioritized?
- Stereotypes: Are any harmful generalizations present?
Mitigating Bias
There are some prevention strategies that we can follow to mitigate the presence of bias in LLM answers:
- Request diverse examples explicitly
- Ask for multiple perspectives on controversial topics
- Challenge AI outputs that seem stereotypical
- Include diverse voices in your verification process
For example, instead of:
Provide examples of successful leaders,
try:
Provide examples of successful leaders from diverse backgrounds, including different genders, ethnicities, and cultural contexts, explaining their varied leadership styles.
Other Ethical Considerations
Beyond bias and environmental concerns, generative AI systems raise significant ethical questions regarding intellectual property, labor practices, and the value of human communication.
Data Rights and Labor Practices
Large language models are trained on vast datasets typically collected without explicit consent from content creators or copyright holders. This practice raises concerns about attribution, compensation, and the rights of original authors whose work contributes to model training.
Additionally, AI development relies heavily on labor that is often invisible to end users. Tasks such as Reinforcement Learning from Human Feedback (RLHF)—which involves rating outputs and identifying harmful content—are frequently outsourced to workers in lower-income countries who receive inadequate compensation. These patterns reflect broader inequities in how AI benefits and burdens are distributed globally.
The Value of Human Process
The appropriateness of AI use extends beyond output quality to consider the intrinsic value of human cognitive and communicative processes.
Reliability: AI-generated content may contain errors or unverifiable claims. The effort required to verify and correct outputs may exceed the effort of completing tasks independently, particularly for complex work requiring domain expertise.
Communication and respect: Using AI to generate responses to thoughtfully composed correspondence may signal that you do not value the communicative exchange equally, potentially undermining professional relationships and collegial trust.
Process versus product: Many academic contexts value the cognitive processes involved in creating work—critical thinking, synthesis, and personal reflection—not merely the final output. Delegating these processes to AI may diminish learning and professional development outcomes.
Responsible AI use requires critical awareness of these broader implications and contextual judgment about when AI tools align with your professional values and the expectations of your academic community.
Exercise 5: Reflective Exercise
To explore the complex human dimensions of AI use, consider how others might feel about being on the receiving end of AI-generated content in sensitive contexts. You could hold a discussion exploring various scenarios, such as:
- Using AI to summarize CVs for job applicants.
- Using AI (without sharing any personal details) to draft an email response to a student about their extenuating circumstances.
- Using AI to write the first draft of a dissertation (you are the supervisor).