Responsible, transparent, human-centred AI

Ethical Standards for Large Language Models

A visually enhanced guide for teaching, research, and innovation. This page outlines the core responsibilities institutions and practitioners should follow when designing, evaluating, and using LLMs in academic and professional settings.

Explore the Principles See the Workflow
Bias Awareness Verification Privacy Sustainability Human Oversight
Ethical
LLM Use
Privacy
Accuracy
Fairness
Transparency
Sustainability
Oversight

Why this guide matters

Large Language Models can amplify productivity, creativity, and knowledge discovery, but they can also introduce bias, misinformation, privacy risks, and misuse. Ethical practice ensures that benefits are maximised while harms are reduced.

Human oversight

People remain accountable for interpreting, checking, and approving outputs in teaching, research, and operational settings.

Critical evaluation

LLM outputs should be treated as generated suggestions, not guaranteed facts, especially in high-stakes contexts.

Responsible data use

Data handling must align with privacy law, institutional policy, and ethical expectations around consent and confidentiality.

Broader impact

Ethical use includes fairness, cultural sensitivity, accessibility, transparency, sustainability, and public trust.

Ten core ethical principles

These principles provide a practical framework for responsible development, deployment, and use of LLMs in academic and research environments.

1. Mitigating bias and promoting inclusivity

LLMs can reflect historical and social biases embedded in training data. Ethical use requires active steps to reduce discrimination and improve representation.

  • Review outputs for biased or exclusionary language.
  • Use diverse and representative datasets where possible.
  • Refine prompts and methods to reduce skewed responses.
  • Promote fairness, respect, and inclusive language.

2. Ensuring accuracy and verifiability

Generated text can sound convincing while still being inaccurate. Validation is essential, especially when outputs affect decisions or public knowledge.

  • Fact-check critical outputs against trusted sources.
  • Use human review in sensitive domains.
  • Disclose that outputs may require verification.
  • Encourage evidence-based judgement.

3. Respecting privacy and confidentiality

LLM use must protect personal, confidential, and sensitive information throughout input, processing, storage, sharing, and disposal.

  • Avoid uploading personally identifiable or sensitive data unnecessarily.
  • Prefer anonymised or aggregated information.
  • Follow institutional and legal data protection requirements.
  • Apply secure data governance practices.

4. Minimising environmental impact

AI systems consume energy and computing resources. Responsible use considers efficiency and sustainability alongside capability.

  • Choose models proportionate to the task.
  • Avoid unnecessary repeated generation.
  • Prefer smaller or specialised models when appropriate.
  • Consider carbon and resource efficiency in deployment.

5. Managing hallucinations and misinterpretations

LLMs can invent citations, facts, or relationships. Users should understand these limitations and establish safeguards against misleading outputs.

  • Introduce review steps before sharing results.
  • Use prompts that request uncertainty or evidence.
  • Educate users about hallucination risks.
  • Encourage critical reading and verification.

6. Respecting intellectual property and ownership

Ethical use requires awareness of copyright, licensing, attribution, provenance, and the ownership of both training and generated content.

  • Use appropriately licensed content and data.
  • Maintain transparency around sources and provenance.
  • Avoid unauthorised use of copyrighted material.
  • Clarify ownership and reuse expectations.

7. Addressing security and safety concerns

LLMs can be misused for harmful, deceptive, or unsafe purposes. Risk management should be built into usage and governance.

  • Monitor for malicious or inappropriate use.
  • Use safeguards against adversarial abuse.
  • Develop incident response protocols.
  • Protect systems, users, and organisational reputation.

8. Promoting ethical and transparent use

Trust depends on openness about capabilities, limitations, intended use, and governance responsibilities.

  • Align practice with institutional values and policy.
  • Engage diverse stakeholders in decision-making.
  • Document limitations, risks, and assumptions.
  • Support transparent deployment and communication.

9. Educating and empowering users

Ethical practice depends on informed users who understand how LLMs work, where they fail, and how to use them responsibly.

  • Provide training on capabilities and limitations.
  • Promote critical thinking and ethical reflection.
  • Support workshops, guidance, and exemplars.
  • Encourage responsible experimentation.

10. Final recommendations

LLMs are valuable assistants when used carefully. Productivity gains should be matched with sound judgement, oversight, and documentation.

  • Experiment thoughtfully with prompts and workflows.
  • Validate outputs before relying on them.
  • Apply stronger checks in critical applications.
  • Use AI to enhance, not replace, human expertise.

Responsible LLM workflow

A simple operational sequence for safe and effective use in educational, research, and administrative practice.

1. Define the purpose

Clarify the task, intended audience, risk level, and whether LLM use is appropriate for the context.

2. Prepare inputs responsibly

Remove sensitive data, reduce ambiguity, and provide balanced, well-structured prompts.

3. Generate with awareness

Use the model as a drafting, brainstorming, or summarisation aid rather than an unquestioned authority.

4. Verify and review

Check facts, inspect reasoning, identify omissions, and assess ethical implications before use.

5. Document and reflect

Record assumptions, limitations, and lessons learned to improve future responsible practice.

Key risk areas to monitor

Institutions should establish review criteria and governance controls around the most common ethical and operational risks.

Bias and exclusion
Outputs may marginalise groups, misrepresent cultures, or reinforce stereotypes if not checked carefully.
Inaccuracy and fabricated content
Models may invent references, statistics, and facts while sounding highly credible.
Privacy leakage
Sensitive personal, research, or organisational data may be exposed through careless prompting.
Copyright and provenance issues
Generated or source material may raise uncertainty around ownership, permission, and attribution.
Misuse and unsafe deployment
Unreviewed automation can create harmful content, flawed decisions, or reputational damage.
Environmental cost
Model selection and repeated large-scale use can increase energy and compute demands unnecessarily.

Practical guidance for users

The following questions can be used in classrooms, research teams, and project reviews to support consistent ethical practice.

Confirm the purpose, audience, and risk level of the task. Decide whether the task involves sensitive information, high-stakes advice, copyrighted material, or content that requires expert verification. If the risk is high, establish a stronger human review process before using the output.

Use clear instructions, avoid confidential details, ask the model to identify uncertainty, and request sources or reasoning where appropriate. Prompts should be designed to reduce bias, clarify expectations, and discourage overconfident or speculative responses.

Human oversight is essential whenever outputs influence grades, research claims, policy, wellbeing, legal interpretation, financial advice, hiring, admissions, or public communication. In such contexts, the model should support human judgement rather than replace it.

Build trust through clear policy, staff and student training, documented governance, transparent disclosure, feedback channels, auditing practices, and ongoing review of bias, privacy, and performance issues. Ethical AI adoption is a continuous process, not a one-time checklist.

Final recommendations

By adhering to these ethical standards, institutions can foster a responsible, transparent, and inclusive environment for using LLMs in teaching, research, and innovation. The goal is not simply to use AI more often, but to use it more wisely.

  • Use LLMs to augment human capability, not to bypass accountability.
  • Validate outputs, especially in critical or sensitive contexts.
  • Protect privacy, intellectual property, and institutional trust.
  • Promote fairness, inclusivity, and culturally respectful practice.
  • Choose efficient, proportionate, and sustainable workflows.
  • Educate users continuously as tools and risks evolve.

Guiding idea

Responsible LLM practice combines capability with care: fairness, accuracy, privacy, transparency, sustainability, and human judgement.