Guidance for Generative AI in Education and Research: An Extended Introduction

On 7 September 2023, UNESCO published Guidance for generative AI in education and research to support action, policy planning and human capacity development. This article retains the scope of the site’s extended introduction of 24 November 2023, covering risks, educational uses and child protection.

Scope

This is a selected introduction and interpretation, rather than the full guidance or an official translation. Its human-centred perspective emphasizes agency, inclusion, equity, gender equality, and cultural and linguistic diversity.

The 2023 publication context

The guidance recorded that ChatGPT reached approximately 100 million monthly active users in January 2023, while only one country had issued dedicated generative-AI regulations by July 2023. These are observations from the report’s historical context, not present-day regulatory statistics.

The original introduction compared the guidance with China’s Interim Measures for the Management of Generative Artificial Intelligence Services, published in July 2023, discussing content labelling, ethics, accuracy and human review. The documents differ in status and scope; the Chinese measures took effect on 15 August 2023, and their provisions should be read in the official text.

At the launch, then UNESCO Director-General Audrey Azoulay described opportunities for human development alongside risks of harm and prejudice. Educational adoption, she argued, requires public participation, governmental safeguards and appropriate rules, with learners’ interests at its centre.

Six controversies and educational risks

The guidance examines frequently overlooked controversies. A prominent educational risk is the mixing of bias and error into apparently credible content, complicating learners’ judgement of knowledge and sources. Six issues highlighted in the original introduction follow.

Data poverty and inequality: models require extensive data, computing and continued technical investment, concentrated in a few major companies and economies. The report referred to the USA, China and European countries. Groups unable to access or afford data may fall further behind. Educators, researchers and learners should examine training data’s values and cultural norms, while policymakers address structural disparities.

Regulatory adaptation: rapid iteration and security challenges can outpace legislation and oversight. Enhancing human capabilities does not automatically make providers accountable. Access to and use of data concerning domestic institutions, individuals and local contexts require appropriate governance and rights protection.

Content used without consent: training may involve texts, sounds, code and images collected without owners’ permission. Users need to understand data rights. Generated images or code may infringe others’ intellectual property, while work shared publicly by users may itself be used by other models.

Source contamination and recursive risks: discrimination and bias in training material may enter generated content; confident fluency is no guarantee of accuracy. Young learners with limited prior knowledge are particularly vulnerable. Generated errors may be collected again for training, creating feedback loops; curricular material therefore requires human verification.

Distance from real-world understanding: the guidance uses the “stochastic parrot” analogy to discuss imitation of language patterns. Plausible text does not establish observation, scientific reasoning or human value judgement. Fluency should not invite unwarranted trust from teachers and students.

Deepfakes and misinformation: generative AI can alter images, audio and video, making fabricated content harder to distinguish. Educational use requires identification, checking and disclosure of synthetic material so that it is not mistaken for genuine evidence.

Copyright in its historical context

The introduction also drew on the guidance’s 2023 copyright comparisons: the USA’s focus on human creativity, disclosure of training material in the then proposed EU AI Act, and Chinese requirements to label specified generated content. These historical discussions should not be presented as a universal current legal conclusion about every AI-assisted work.

Creative uses start with learning needs

ChatGPT’s launch prompted concerns about ghostwriting and cheating, alongside experiments with ideas, varied examples, lesson plans, presentations, summaries and images. The guidance prioritizes educational and research needs and ethical principles over technology-driven enthusiasm.

Four institutional actions

Education and research institutions need to design, implement and validate appropriate strategies and ethical frameworks. The original introduction identifies four complementary actions:

• Institutional ethics: define responsible use and rigorously check outputs for accuracy and validity.
• Guidance and training: understand data-labelling and algorithmic biases, and comply with applicable privacy and intellectual-property rules.
• Prompt design and evaluation: combine subject expertise with writing and improving prompts, critical evaluation of answers, and high-quality training and support.
• Academic integrity and assessment: the guidance noted limited evidence for watermarking and detection tools. Human judgement, accountability and redesigned assignments should foreground empathy, creativity and human capabilities in complex real-world problems.

Exploring research and teaching uses

Through co-design in use, the guidance considers research, teacher support, basic-skills practice, higher-order thinking and special learning needs. These are avenues requiring review and validation, rather than guarantees about any tool’s effectiveness.

Research support: generative AI may broaden research ideas, assist data exploration and literature reviews. Research questions, expected outcomes and tests of accuracy and effectiveness still need definition, without weakening researchers’ understanding of the real world.

Teaching support: teachers may co-design lesson plans, course packages and curricula to strengthen subject and pedagogical understanding. Conversational assistants trained with experienced teachers’ knowledge and resource collections have been explored, but both potential and ethical risks require scrutiny, iteration and human oversight.

Individual basic-skills practice: listening, pronunciation and writing in first or additional languages, numeracy, arts and coding remain foundations of development. Practice need not become obsolete; suitable tools may support repetition at an individual pace while preserving higher-order thinking and creativity.

Special education and accessibility

Special and accessible education: possible uses include captions for deaf and hard-of-hearing learners, audio descriptions for blind learners, and text-to-speech or speech-to-text to help people with visual, hearing or speech impairments access content, ask questions and communicate with peers.

A 2023 government survey cited by the guidance found that only China, Jordan, Malaysia and Qatar reported governmental validation and recommendation of AI-assisted tools for inclusive access by learners with disabilities. This concerned AI-assisted tools; it does not establish the effectiveness of all generative-AI applications.

For proposals to identify psychological, social-emotional or learning difficulties through dialogue, the guidance stresses limited evidence of effectiveness or safety and the need for professional interpretation of any diagnosis. Such exploration should not be described as a mature substitute for professional services.

Child protection and age-appropriate use

Most general-purpose generative-AI applications were initially designed for adults and may expose children to inappropriate content or manipulation. With rapid iteration and uncertainty, the guidance recommends age-appropriate design, restrictions and supervision to protect children’s rights and wellbeing.

For independent conversations with generative-AI platforms, the guidance recommends a minimum threshold of 13 and public consideration of suitable limits and verification. This is a policy recommendation, not a directly binding age law imposed by UNESCO on every country.

The report illustrated its discussion with ChatGPT’s terms at the time: a minimum age of 13 and parental or guardian permission below 18. It also considered 13- and 16-year thresholds and consent arrangements. These 2023 comparisons do not replace checking current service conditions and applicable local rules.

The guidance calls for clear provider responsibility for age verification and parental or guardian responsibility for supervising minors’ independent conversations. Age thresholds need to be considered alongside educational suitability, privacy and effective supervision.

A human-centred conclusion

A human-centred approach seeks to extend intellectual and social capabilities. Building on the 2021 Recommendation on the Ethics of Artificial Intelligence, scrutiny is needed at individual, institutional and system levels. Changes to education and research should advance inclusive digital futures for all.

Related reading

Related reading: compare the site’s guidance launch brief, final teacher and student AI competency frameworks, and special and inclusive education resources. These offer different scopes from this extended historical introduction.

Original publication and sources

Original publication: Fengchun Miao and Wayne Holmes, Guidance for generative AI in education and research, UNESCO, 2023. DOI: 10.54675/EWZM9535. The report uses CC BY-SA 3.0 IGO, with stated exceptions for third-party images; the official publication page offers different language editions.

Editorial note: the original Chinese title, URL and 2023 archive date are retained. The user count, age recommendation and mistranslations of GenAI have been corrected against the report. Policy and product examples retain their historical context.