Guide to Best Practices

Guide to good practices in the use of generative AI for teaching at the Miguel Hernández University of Elche

September, 2024

Rector’s Delegation for Digital Transformation of the UMH

Download the Guide to good practices in the use of Generative AI for teaching at the UMH (PDF).

Download the Guidelines for good practices in the use of Generative AI for teaching in the MHU (PDF).

Table of Contents

ChatGPT and large language models.

Generative AI.

Reliability of generative AI.

Use of Generative AI in Universities.

Alternative Approaches to Assessment.

Use of alternative generative AI.

References.

ChatGPT and large language models

ChatGPT is a large language model (LLM, by its English acronym) of artificial intelligence developed by the company OpenAI. It is an advanced chatbot designed to maintain natural conversations and answer questions in a coherent and plausible manner.

ChatGPT’s operation is based on generative AI algorithms, trained with immense datasets that include millions of web pages, books and articles. These algorithms analyze patterns and relationships between words to produce coherent responses. The text generation process is iterative: at each step, the model determines which word or symbol is most likely to follow, based on the context provided by the user.

ChatGPT’s training involves significant costs due to the enormous amount of data and computing power required. It is estimated that the training cost of models like GPT-3 (with 175B parameters), on which ChatGPT is based, was between 4-5 million dollars (ref. Visual Capitalist).ChatGPT was launched on November 30, 2022 by OpenAI, a technology company with significant investment from Microsoft. Since its launch, it has had a notable impact on various fields, including teaching. In the educational field, it has transformed teaching and assessment methodologies, driving the need to integrate new technologies and adapt to generative AI tools.

The launch of ChatGPT has led to a review of teaching practices. Educators have had to develop new strategies to leverage AI capabilities, such as generating drafts, offering personalized tutoring, and fostering self-reflection in students. At the same time, the need has emerged to address ethical and academic integrity issues related to AI use. Many educators have begun to incorporate ChatGPT as a learning tool, while others express concern about its potential misuse.

Currently, ChatGPT offers a free version that allows users to interact with the basic model. There are also paid versions, such as ChatGPT-4o, which provide access to more advanced models and enhanced capabilities. It is expected that in the future new versions will be launched with additional functionalities and greater accuracy, continuously adapting to user needs and technological advances. Other companies have launched their chatbots such as Google with its Gemini, which integrates with its Google Workspace products (such as Gmail, Google Docs or Google Sheets), or Anthropic with its Claude 3.5 chatbot, which defines itself as an LLM model trained to be a helpful, honest and harmless assistant, with a conversational tone.

Generative AI

Generative AI is a type of artificial intelligence capable of creating new and original content from training data. It is used to produce various types of content in an automated manner.

Its main functions include:

  1. Text generation: It can write articles, stories, scripts, etc.
  2. Image creation: Produces original images based on textual descriptions.
  3. Musical composition: Creates melodies and complete musical pieces.
  4. Design: Generates graphic designs, logos, or even architectural plans.
  5. Programming: Can write code and develop software.
  6. Translation and natural language processing: Translates between languages and understands language context.
  7. Data prediction and analysis: Generates predictions based on existing data patterns.
  8. Video and animation creation: Produces original audiovisual content.

A generative AI works through several key components and processes. It begins with training using large datasets relevant to its task, such as texts, images, or sounds, from which it learns patterns and features. It employs complex neural network architectures, such as deep neural networks (DNN), Transformers (GPT) or generative adversarial networks (GAN), which allow it to capture intricate relationships in the data.

The generation process uses what it has learned to create new content that is similar but not identical to the original, being able to generate from scratch or complete partial content. To control the randomness and creativity of outputs, sampling techniques such as temperature sampling or nucleus sampling are employed.

Many systems use iterative processes to gradually improve the quality of the output. Additionally, these AIs can be adjusted for specific tasks or particular styles through techniques such as fine-tuning. Often, they allow user inputs to guide or modify the generation process, which increases their versatility and applicability in different contexts.

Reliability of generative AI

The reliability of generative AI is a complex and nuanced topic. In general, these AIs can be very useful and produce impressive results, but they also have limitations and risks that are important to consider.

On one hand, generative AI can be very reliable for certain tasks:

  • Generation of creative content such as images or text
  • Assistance in programming tasks or data analysis
  • Translation and natural language processing

However, there are several aspects that can affect its reliability:

  1. Data bias: If the training data contains biases, these will be reflected in the outputs.
  2. Lack of real understanding: Although they can generate coherent content, they do not have a true understanding of the real world.
  3. Hallucinations: They can generate false or inaccurate information, especially when asked for something outside their knowledge base.
  4. Inconsistency: Responses can vary between different executions.
  5. Lack of updated context: Their knowledge is limited to the date of their last training.
  6. Ethical and legal issues: Problems can arise from an ethical perspective, and violations of regulations, particularly regarding data protection and intellectual property.

Therefore, it is crucial to use generative AI with judgment, verify information when necessary, and understand its limitations. Its reliability depends largely on the context of use and how it is implemented and monitored.

Use of Generative AI in Universities

Generative AI is significantly transforming the educational landscape in universities, offering a variety of applications that can improve both teaching and learning.

Students can improve their learning through reinforcement with generative AI tools that will offer them help in answering questions and obtaining explanations of complex topics. Students could use generative AI to reinforce and learn complex topics, not just to obtain answers and use them directly in their work. Other uses of generative AI can be support in drafting academic papers or scientific articles, or translating materials into other languages, or answering previous years’ exams, or creating concept maps. Generative AI enables the creation of accessible materials for students with disabilities, such as text-to-speech conversion and vice versa.

Teachers can use generative AI to support the creation of content or the adaptation of scientific materials to different levels of student comprehension, the automation of administrative tasks, such as grade management, or the analysis of student performance data to identify areas for improvement in teaching. It can also help in creating interactive and/or gamified content that motivates student learning.

Generative AI has come to stay with us, and rather than prohibiting it, we should assume that it is another tool that will improve learning and teaching as happened with the arrival of other tools such as calculators, personal computers, mobile phones, or the Google search engine.

Therefore, we must take into account the following considerations regarding the use of generative AI at UMH:

  • We should promote ethical use of AI, avoiding plagiarism and ensuring the authenticity of academic work.
  • Data privacy is fundamental in any interaction with any AI.
  • The technological divide should be avoided whenever possible, trying to ensure that students have equal access to AI technologies, avoiding inequalities in academic performance.

· Guidelines on the use of AI by UMH students and faculty

  1. In the context of Generative AI use, UMH students and faculty are strictly prohibited from sharing any type of personal data with it. Likewise, you must ensure that your questions and content do not contain personal data or information that could result in the inference or deduction of personal data.
  2. UMH students and faculty must be aware that personal data includes, among others, personal identifiers in text, photographs, audio recordings, video, or any document or file containing personal data. Furthermore, it should be considered that the combination of different types of data can lead to the identification of individuals.
  3. All of the above is categorically prohibited in the use of generative AI applications, and compliance with these restrictions is mandatory for UMH students and faculty.
  4. In this sense, and if the use of data is necessary, UMH students and faculty must use completely anonymous data. In the context of the GDPR, anonymous data is that which cannot be associated with an identified or identifiable person, due to the removal of identifiers or techniques that prevent re-identification. To be considered truly anonymous, data must be irreversible, impossible to identify directly or indirectly, and maintain its anonymity in any context. If these criteria are met, the data ceases to be personal and is not subject to GDPR provisions

Alternative Approaches to Assessment

The literal use of responses generated by generative AI by students in academic work or final degree or master’s thesis work may be considered as the use of illicit or fraudulent means according to article 9.4) of the Student Assessment Regulations of Miguel Hernández University of Elche.

However, the use of AI in teaching cannot be radically prohibited, since other uses of AI are perfectly lawful as previously described.

Given this current scenario we find ourselves in with the use of generative AI in teaching, we must find alternative approaches to the assessment of our students where AI is used legitimately and academic integrity is guaranteed.

The following recommendations for the use of AI by students, in their learning process, and by faculty, in their teaching process, can be considered as a guide to best practices in the use of generative AI:

  1. Academic work should not be easily generatable by generative AI: work that a generative AI can write will not allow for proper student evaluation. Faculty should previously check whether a generative AI writes a piece of work well or not. An appropriate assessment test should:
    • Be sufficiently specific: not limited to the presentation of an overly generic topic that can be easily copied or generated.
    • Require higher-order cognitive skills: analyze, compare, justify, criticize, and not just transfer already existing information.
    • Involve practical tasks, such as collecting testimonies or real data, processing or developing data, analyzing concrete cases, etc.
  2. Evaluate the process, not just the result: Conduct proper monitoring of work during its completion, without waiting until the final submission of the task, with periodic reviews in which students report to faculty about the work development process.
  3. Include self-reflection components in the assessment: require the student to answer questions regarding their learning and the work completed:
    • How have you planned and structured the work? What problems have you encountered and what decisions have you had to make?
    • What have you learned from the work? What new knowledge or skills have you acquired? What prior knowledge or skills have you needed to apply?
  4. Assess the suitability of written work to learning objectives and consider other options and assessment strategies:
    • Consider other types of evaluable work: videos, presentations, concept maps or infographics.
    • Consider other assessment strategies, such as project-based learning.
  5. Explore with students the possibilities and limitations of technology:
    • You can work with students using information generated by generative AI on course topics, identifying errors or verifying data.
    • Especially in subjects focused on text writing or computer code, you can try to integrate generative AI tools into the usual work cycle.
    • For example, you could ask students to use text or code generated by generative AI and explain and justify the changes they have had to make to reach their final version.

Use of alternative generative AI

Testing different Generative AI tools such as ChatGPT, Gemini, Copilot or Claude will allow you to obtain a clearer understanding of the scope and limitations of this technology in your field of knowledge. However, remember that this is a service external to the university and that you will need to accept the personal data processing that it may carry out. Therefore, it should not be used with students in mandatory teaching activities.

Especially in teaching activities, the application we recommend is Google’s Gemini or Microsoft’s Copilot using your UMH corporate email account. If you use Gemini or Copilot with your UMH account, both Google and Microsoft guarantee that they will not save your conversation data, will delete it at the end of the session, and will not use it to train their AI models.

ChatGPT or Claude do not offer UMH users the same level of protection. In any case, avoid entering personal data or confidential information in your interactions with any Generative AI.

References

This guide is framed within the collaboration agreement between the Generalitat Valenciana, through the Office of the President, and the Miguel Hernández University of Elche to promote activities in matters of transparency, access to public information, good governance, open data and institutional integrity, during the 2024 fiscal year.