Responsible AI Use in Higher Education

Responsible use starts before a prompt is written. The educator or institution remains accountable for the purpose, the data, the decision, and its effects on students. AI can support professional judgment; it should not replace it.

A proportionate decision process

Data protection and privacy

Do not paste identifiable student work, grades, disability or health information, disciplinary records, private communications, unpublished research, or confidential institutional material into an unapproved service.

Consent alone does not automatically make a tool or use compliant. In an educational relationship, consent may not always be freely given, and an institution still needs an appropriate legal basis, purpose, security measures, retention rules, and processing arrangements. The EU’s GDPR principles include lawfulness, fairness and transparency, purpose limitation, data minimisation, accuracy, storage limitation, security, and accountability.¹

Practical safeguards include:

• remove names, student numbers, and indirect identifiers;
• share only the minimum information needed for the task;
• prefer synthetic examples when real records are unnecessary;
• check contracts and settings rather than relying on an “incognito” label;
• keep source material and reviewed final work in approved storage; and
• report accidental disclosure through institutional procedures.

Removing a name does not necessarily make work anonymous. Distinctive personal experiences, small cohorts, or a separately held identifier key can still allow identification. Treat pseudonymised material as personal data and check approval for the actual material and purpose.

Running a model locally can reduce disclosure to a third-party provider, but it does not by itself resolve lawful-basis, access-control, security, bias, or copyright questions.

High-impact educational uses

Distinguish prohibited uses from uses requiring assessment

The EU AI Act prohibits AI used to infer emotions in educational institutions, except for systems intended for medical or safety reasons. This concerns emotion recognition based on biometric data, such as facial or vocal signals; it should not be read as a blanket legal classification of every text-analysis tool. Do not introduce emotion-recognition monitoring of students under the label of engagement or wellbeing. Human approval does not remove the prohibition; refer any proposed exception to institutional specialists.²

Obtain an institutional classification for high-risk uses

Under the EU AI Act, some systems used to determine access or admission, evaluate learning outcomes, including to steer learning, assess an appropriate educational level, or monitor prohibited behaviour during tests may be classified as high-risk.³ Classification depends on the system’s intended purpose and how it is used—not simply on whether it is called “generative AI”. Annex III must be read with Article 6’s classification rules. Do not self-certify an exception because a person checks the result. Ask the institution to establish classification, applicable obligations, and current application dates before procurement or use.⁴

Do not introduce AI as the sole or decisive basis for:

• admitting, ranking, placing, passing, failing, or awarding a grade;
• alleging plagiarism, cheating, or other misconduct;
• making disciplinary, welfare, or disability-support decisions; or
• inferring a student’s ability, motivation, emotion, or future performance.

These workflows require specialist institutional assessment. They may also require risk management, data governance, documentation, oversight, transparency, monitoring, and a meaningful route for people to challenge the outcome. The AI Act also requires providers and deployers to take measures that support an appropriate level of AI literacy among relevant staff.⁵

Human oversight that is actually meaningful

Research on automation bias in decision-support systems shows that people can follow incorrect automated advice and that workload and verification conditions matter. This evidence has a healthcare focus and predates GenAI; it supports attention to oversight, not a measured effect for university chatbot use.⁶

A person merely clicking “approve” is not meaningful oversight. The reviewer should be able to:

• inspect the original evidence and the criteria used;
• recognise common model errors and automation bias;
• change or reject the output without penalty;
• explain the final decision in their own words; and
• offer an accessible correction or appeal process.

For consequential uses, periodically test whether errors are concentrated among particular groups or types of work. Stop the workflow if the reviewer cannot independently establish that it is reliable and fair enough for its purpose. Use the workflow evaluation guide to record error patterns, comparisons, and stop conditions.

Transparency and academic integrity

Tell students when AI materially shapes teaching materials, feedback, or a student-facing interaction. Explain its role, the checks performed, who is responsible, and how to report an error. Disclosure should help students judge and question the result.

For student work, publish allowed actions, evidence requirements, and access arrangements in advance. Use Setting course and assignment AI rules for the four-level planning model, copyable assignment block, and proportionate disclosure template.

Disclosing AI use does not make fabricated citations, undisclosed outsourcing, or prohibited assistance acceptable. Apply the learning outcomes and the local academic-integrity policy.

Copyright and licensing

Provider terms, copyright protection, permission to reuse source material, and academic attribution are different questions. A service’s terms may allocate contractual rights in an output, but they do not guarantee that the output is copyright-protected, non-infringing, accurate, or acceptable under an institution’s rules.

Keep track of important source material, licences, and substantial human contributions. Avoid prompts intended to imitate a living creator or reproduce a protected work. Review outputs for recognisable passages, images, code, or other material before publishing them. The EU Intellectual Property Office maintains current guidance and resources on copyright and generative AI.⁷

Tool-selection checklist

Before adopting a tool, confirm:

• Purpose: it solves a defined teaching or operational need;
• Approval: procurement, privacy, security, and accessibility requirements are met for the intended use;
• Data controls: retention, training use, deletion, location, and access are understood;
• Evidence: users can inspect sources or compare outputs with authoritative material;
• Equity: cost, language, accessibility, and account requirements do not create an avoidable barrier;
• Oversight: a named person owns review and final decisions;
• Exit: work can be exported and the activity can continue if the tool changes or becomes unavailable; and
• Review: the use has an owner and a date for reassessment.

A reusable review record

Copy this into a course or project file:

Purpose and expected benefit:
People affected:
Tool and version:
Data entered and its classification:
Consequence if the output is wrong:
How outputs will be verified:
Human decision owner:
How the use will be disclosed:
Accessibility or non-AI alternative:
Institutional approvals or consultations:
Review date and stop conditions:

Further guidance

• Guidelines on the ethical use of artificial intelligence and data in teaching and learning for educators , European Commission, updated 2026
• AI Act Service Desk: education and vocational training , European Commission
• When to use generative AI, SAGE’s shorter task-level checklist

References & Footnotes

• European Commission. (2026). Guidelines on the ethical use of artificial intelligence and data in teaching and learning for educators. https://doi.org/10.2766/9548952