Ethics & Environmental Impact

Environmental Impact

It is easy to think of AI as invisible and weightless — just text on a screen. In reality, every interaction with an LLM requires substantial computational resources, with real-world environmental consequences.

Energy and Carbon

Training a large language model like GPT-3 is estimated to have produced around 85,000 kg of CO₂ — equivalent to roughly 112 cars running for a year. But training is a one-off event. The inference phase — every time a user sends a prompt and receives a response — can collectively surpass the emissions from training within just a few weeks of deployment.

Water

Data centres require enormous amounts of cooling. It is estimated that ChatGPT uses approximately half a litre of fresh water for every session of 20–50 prompts. Strikingly, around 40% of new data centres built in the USA since 2022 have been located in areas already experiencing water scarcity.

What Can You Do? Sustainable AI Practices
  • Batch your queries: Combine multiple related questions into one prompt rather than sending many separate ones
  • Use AI for high-value tasks: Avoid using it for things you could easily do yourself
  • Avoid unnecessary regeneration: Don’t keep asking for new versions if one is adequate
  • It is always acceptable not to use AI — knowing how LLMs work is valuable even if you choose not to use them

Bias and Fairness

AI systems do not emerge from nowhere — they are trained on text produced by humans, and human-generated text reflects human biases. These biases can be subtle, but in a scientific context they matter.

Bias in Science

Biological and medical research has historically over-represented certain populations (predominantly Western, high-income, and often male subjects). An LLM trained on this literature may treat these populations as the “default” and under-represent findings from other groups. For example:

  • Drug dosage information may reflect studies conducted predominantly on one sex or ethnicity
  • Disease prevalence data may be skewed toward populations with better-funded healthcare systems
  • “Normal” physiological ranges cited by AI may not reflect the full diversity of human biology

Other Forms of Bias

LLMs may also exhibit:

  • Gender bias — associating certain scientific roles (e.g., “the scientist”) with a particular gender
  • Geographic bias — defaulting to Western or English-language sources and perspectives
  • Socioeconomic bias — assuming access to resources, technologies, or contexts that are not universal

When reviewing AI outputs, ask yourself:

  • Who is included — and excluded — in the examples given?
  • What assumptions are being made about “typical” cases?
  • Whose perspective is being prioritised?

You can also prompt explicitly for diversity:

Give me examples of important scientists in genetics.

Give me examples of important scientists in genetics from diverse backgrounds, including different genders, ethnicities, and countries, explaining their specific contributions.

Broader Ethical Considerations

Academic Integrity

Using AI to write your essays or lab reports and submitting them as your own work is a form of academic misconduct at most universities, including Bristol. Beyond the disciplinary risk, there is a deeper issue: the process of writing is part of the learning. Outsourcing your thinking to an AI means missing the development of skills you will need as a scientist.

Data Rights and Invisible Labour

LLMs are trained on text collected from the internet, often without the explicit consent of the people who wrote it. Many creative writers, journalists, and researchers have had their work incorporated into models without attribution or compensation.

Additionally, AI systems are refined using human feedback — tasks like identifying harmful content or rating outputs. This work is often outsourced to workers in lower-income countries who receive minimal pay for psychologically difficult work. These are real people whose contributions remain invisible to most AI users.

The Value of Human Thinking

There are contexts where using AI — even if technically permitted — may not be the right choice:

  • Responding to a peer’s thoughtful question using AI-generated text may undermine the relationship
  • The cognitive effort of writing and reasoning has educational value that AI bypasses
  • Scientific integrity requires that you understand and stand behind what you write

Responsible AI use is not just about following rules — it requires ongoing reflection about whether AI is the right tool for a given situation.