IA generativa en educación: una visión práctica | UOC | Robert Clarisó

Name: IA generativa en educación: una visión práctica | UOC | Robert Clarisó
Uploaded: 2023-07-18T16:48:47.000Z
Duration: 2 h 48 min 26 s

Welcome to the AI Seminar

Introduction and Overview

The seminar on Artificial Intelligence (AI) has attracted 3,000 participants from various countries, including Spain and Latin America.

Participants are encouraged to engage through different channels of the webinar, highlighting a diverse international audience.

The focus is on identifying elements that constitute new education methods using AI tools, especially in light of experiences during the pandemic.

Goals of the Webinar Series

The webinars aim to create a cycle of reflections and materials related to AI in education, fostering knowledge sharing among educators and institutions.

Key questions include how to effectively utilize AI tools for student benefit and how to educate university staff and students about productive use of these technologies.

Institutional Support

The UOC's Center for Innovation aims to support learning processes not only for its faculty but also for other universities and private companies undergoing digital transformation.

Attendees are invited to share their questions or topics they wish to explore in future webinars, aiming for a collaborative resource bank for educators.

Introduction of Speaker: Rubén Clariso

Speaker Background

Rubén Clariso is introduced as a professor at the University of Catalonia with expertise in software application techniques and research in relevant fields.

He leads a research group focused on software engineering within an innovation center known as IM3.

Expectations from the Session

Clariso expresses gratitude towards Manel (the host), emphasizing excitement about discussing generative AI's impact on education.

He notes that recent advancements in generative AI tools have significantly affected universities, likening it to an earthquake due to their unprecedented capabilities.

Practical Insights into Generative AI

Current Landscape

Clariso plans to present practical insights based on UOC's experiences with generative AI tools since their introduction late last year.

Engagement During Presentation

Introduction to Generative Artificial Intelligence in Education

Overview of the Presentation Structure

The presentation will cover generative artificial intelligence (AI), focusing on its educational relevance without delving into overly technical details.

Discussion will include various applications of generative AI in education, emphasizing practical scenarios for implementation.

A significant concern addressed will be the impact of generative AI on evaluation processes within education.

The speaker aims to summarize key conclusions and future expectations regarding generative AI's role in education.

Understanding Generative Artificial Intelligence

Three metaphors are introduced: a Swiss Army knife (versatile tool), a magic lamp (fulfilling requests), and a parrot (repeating information without understanding).

Generative AI is simplified as computational learning where data is provided to train a model that represents knowledge from those data sets.

Mechanism of Learning and Prediction

The training process involves identifying relevant features from data, revealing patterns stored as internal parameters or weights within the model.

These patterns enable the model to make predictions or complete tasks, such as classifying messages as spam or not.

Tasks Performed by Generative AI

Generative AI can emulate fulfilling specific instructions, producing outputs based on textual prompts received from users.

Outputs can vary widely, including text, images, code, videos, and 3D graphics depending on input types and user requirements.

Types of Generative AI Tools

Different tools exist for generating various outputs; examples include text-to-text models like ChatGPT which predict subsequent words based on given sequences.

Understanding Generative AI and Its Versatility

The Functionality of Generative AI

The concept of "defense is a good attack" illustrates the versatility of generative AI, allowing it to respond effectively to various types of prompts, including questions.

ChatGPT operates on language models known as GPT (Generative Pre-trained Transformer), with version 3.5 being commonly used; its internal parameters have significantly increased since 2018.

The growth in data used for training these models has led to substantial improvements in their accuracy over time.

Examples of Generative AI Applications

Users can request diverse outputs from ChatGPT, such as poems or reports, showcasing its ability to handle different formats and styles.

A poem generated by ChatGPT addressed themes like ethics and privacy relevant to artificial intelligence, demonstrating its capability for creative writing.

Compared to manual efforts, the speed and quality of output from ChatGPT are impressive; it can produce results that would take humans much longer.

Control and Customization in Outputs

Tools like Stable Diffusion allow users to generate images based on text prompts while providing extensive customization options regarding style and quality.

Key factors contributing to the success of generative AI include versatility across tasks not specifically trained for, unlike previous AI tools limited by their training scope.

Accessibility and Ease of Use

Generative AI's user-friendliness means no programming knowledge is required; users can interact using natural language instructions.

Despite occasional errors, generative AI generally performs well across various tasks with minimal revisions needed for effective use.

Creativity and Human Interaction

The creativity exhibited by generative AI often involves unexpected combinations that may not be directly present in the training data but reflect learned patterns.

Discussions about whether this creativity is genuine lead into philosophical debates about the nature of creativity itself.

Barriers to Entry and User Engagement

The accessibility of generative AI tools through web browsers eliminates installation barriers; they are increasingly integrated into everyday applications like text editors.

Free access encourages experimentation among users who wish to explore the capabilities of these technologies without financial commitment.

Limitations and Misconceptions

Challenges of Generative AI

Issues with Image Generation

Generative AI can produce confusing outputs, such as images where features appear distorted or misplaced, exemplified by a bunny with an ear growing from its shoulder.

A notable challenge in text-to-image systems is generating human hands accurately; they often result in unrealistic depictions, like a handshake featuring six fingers.

Limitations of Current AI Models

While generative AI models are reasonably effective, they can misinterpret instructions and may overlook critical details if too many instructions are provided.

The convincing nature of the generated text can lead to misinformation; users need domain knowledge to verify the accuracy of the information presented.

Problems Arising from Training Data

When lacking specific data, generative models may fabricate facts, leading to entirely invented references or dates in their responses.

Biases present in training data can manifest in model outputs. For instance, if trained on biased texts, the model may generate racially biased salary recommendations based on race.

Lack of Introspection and Awareness

Generative models like ChatGPT predict probable next words without forming a mental model of the task at hand. This leads to errors such as not recognizing that humans typically have five fingers.

The clarity and specificity of user instructions significantly impact output quality; vague or overly complex prompts can yield unsatisfactory results.

Broader Implications and Misuse Potential

Generative AI is a powerful tool but poses risks for misuse, including creating fake news or impersonating individuals for malicious purposes.

There are concerns regarding transparency about training data used by organizations developing these models; undisclosed biases could affect reliability and fairness.

Recurrence and Ethical Concerns

The issue of recurrence arises when generative models train on content produced by other generative systems, potentially perpetuating original errors.

Impact of Generative AI in Education

Energy Costs and Efficiency

The execution of generative AI tasks incurs significant energy costs due to computational demands, suggesting that simpler programs may be more efficient for certain educational tasks.

Principles for Implementing Generative AI

It's crucial to introduce generative AI in education with a specific purpose, aiming to enhance particular aspects rather than merely innovating for the sake of it.

Innovations should be measurable; having clear objectives allows educators to assess whether the implementation has achieved its intended goals.

Cautions Against Overuse

There is a risk of over-reliance on generative AI tools by both teachers and students; these tools should serve as assistants rather than replacements for human interaction.

Educators must review outputs generated by AI before sharing them with students, especially when accuracy is critical, such as in assessments.

Transparency and Student Awareness

Educators should inform students when using generative AI in their activities, fostering awareness about potential errors or misinterpretations in the content produced by these tools.

Practical Applications of Generative AI

One effective use case is providing real-time feedback to students on their work or questions regarding course materials, enhancing learning through immediate responses.

Impact of Generative AI on Education

The Role of Generative AI in Learning Tools

The integration of generative AI tools into educational settings is becoming commonplace, serving as a supportive resource rather than a replacement for traditional methods.

Generative AI allows for personalized learning experiences by creating unique activities and resources tailored to each student's needs, adapting difficulty levels based on individual characteristics.

Flexibility and Challenges in Assessment

A significant advantage of generative AI is its flexibility; however, it poses challenges in assessment due to the need for multiple versions of exercises that require careful review.

Educators face the dilemma of evaluating numerous outputs generated by AI, necessitating a system where the AI can also provide correct solutions alongside generated problems.

Teaching Effective Use of Generative AI

It is crucial to teach students how to effectively utilize generative AI, focusing on skills like prompt engineering to enhance their problem-solving strategies.

Students should learn various approaches to construct prompts, similar to how they are taught programming tools and debugging techniques.

Critical Perspective on Generative AI Outputs

Encouraging students to critically analyze outputs from generative AI helps them identify errors and fosters a critical mindset towards technology use.

Evaluating Competence with Generative AI

Educators must balance using generative AI as an aid while ensuring that it does not hinder the assessment of student competencies.

The challenge lies in measuring how well students achieve learning objectives when they have access to powerful tools that can produce varied results.

Ethical Considerations in Using Generative AI

While some uses of generative AI may be beneficial (e.g., grammar correction), others may raise concerns about academic integrity if students overly rely on these tools without contributing their own intellectual effort.

Different assignments may warrant different levels of acceptable tool usage; educators must define what constitutes appropriate assistance versus complete reliance on technology.

Conclusion: Navigating the Future with Generative AI

As educators adapt to incorporating generative AI into curricula, they must remain vigilant about maintaining academic standards while leveraging technological advancements.

Understanding Generative Ideas in Education

The Concept of Generative Ideas

Importance of recognizing generative ideas and their capabilities, as discussed in the initial minutes of the presentation.

Emphasis on assessing the risk level associated with course content and activities to understand potential challenges.

Establishing Norms and Evaluation Strategies

Need for clear guidelines so students know permissible actions regarding generative tools.

Discussion on implementing alternative evaluation mechanisms to adapt to generative AI's influence.

Risk Levels in Assignments

Certain tasks are more susceptible to resolution by generative AI, impacting assignment difficulty based on task type.

Differences in complexity between simple text-based tasks versus those requiring reasoning or programming skills.

Language Considerations and Contextual Risks

Generative AI performs better in English contexts; however, translation tools mitigate some language barriers.

In high-risk scenarios, increased scrutiny is necessary for AI usage compared to low-risk environments.

Guidelines for Student Use of Generative Tools

Students must consult instructors before using generative guides; misuse can lead to academic penalties.

Clear referencing of any generative tool use is required, ensuring students take responsibility for their submissions.

Multiple Assessment Strategies

Employing diverse assessment methods helps prevent reliance on a single point of failure during evaluations.

Regularly comparing student submissions aids in identifying shifts in writing style potentially influenced by AI tools.

Designing Activities with AI Limitations in Mind

Focus should be placed on assignments that highlight human value-add over machine capabilities.

Encouraging reflective activities allows students to articulate challenges faced during task completion.

Using Images in Educational Contexts

Importance of Visual Aids

Utilizing images in educational prompts can enhance understanding; for instance, providing a screenshot and asking students to explain it limits reliance on text-only models.

Incorporating multiple formats (like graphs and texts) encourages reasoning and logic, especially in tasks where generative AI may not be effective.

Challenges with Generative AI Detection

Detecting AI-generated text is complex; students often rewrite content using various tools, making it hard to identify original sources.

Style changes between different submissions can indicate AI use, but detection tools have high rates of false positives and negatives.

Evaluating the Use of Generative AI

Control Mechanisms

Multiple checkpoints are necessary when suspecting inappropriate activity; these tools should serve as additional evidence rather than conclusive proof.

Alternative evaluation methods should be employed when there's suspicion of misconduct or high-risk scenarios.

Types of Activities for Assessment

Effective assessment strategies include in-person tests or synchronous interviews where access to generative tools is restricted.

Opportunities and Concerns with Generative AI

Benefits for Teaching

Generative AI offers significant opportunities by automating low-value tasks, enhancing learning processes through personalized feedback.

It introduces new competencies that educators must instill in students regarding effective use of generative AI technologies.

Risks Associated with Overreliance

The powerful nature of generative AI necessitates adaptations in teaching methods, which could lead to misuse by both educators and students.

There’s concern that reliance on such tools may hinder essential skills like writing, which are foundational for other competencies.

Future Trends in Generative AI

Integration into Daily Tools

Expect widespread integration of generative AI across mobile devices and content creation tools, enhancing accessibility.

Advancements in Model Capabilities

Future models will likely be larger with more data precision; they will incorporate plugins for specific problem-solving capabilities.

Multimodal Inputs and Outputs

Increasingly multimodal models will accept diverse inputs (text, image, video), allowing users to generate comprehensive presentations without direct interaction.

Anticipated Regulation

The Business Behind Generative AI

Impact of Data Access Restrictions

Generative AI models rely on large volumes of data, but platforms like Twitter are limiting access to tweets to prevent training these models.

This may lead to an increase in paid models and a decrease in free options, with advanced functionalities becoming exclusive to paid services.

Transitioning to Q&A Session

The speaker thanks the audience and indicates that a Q&A session will begin shortly, along with resources from the Learning Innovation Center.

Challenges in Distance Education

Obsolescence of Certain Activities

A question arises about which educational activities might become obsolete due to generative AI tools.

Similar to how Wikipedia changed research tasks, general assignments may need redefinition as students can easily find information online.

Adapting Assignments for AI Context

Assignments must be more specific and analytical rather than broad topics that can be easily researched.

Long essay writing may not be feasible without oversight; alternative assessments like asynchronous interviews could supplement evaluations.

Limitations of Generative AI Tools

Evaluation Challenges

Some valuable educational activities may no longer fit within an evaluation framework due to the unregulated use of generative AI.

Updates and Availability Issues

Generative AI tools frequently update without user awareness; this can affect their performance unpredictably over time.

Relying heavily on specific tools poses risks if they become unavailable or shift to a paid model unexpectedly.

Future of Generative AI Problems

Inherent Limitations

Some issues with generative AI are likely solvable through larger datasets, while others stem from fundamental architectural limitations of current models.

Ongoing Debates

Understanding the Limitations and Applications of Generative AI

Short-term Problem Resolution

The speaker discusses that while generative AI may not resolve certain foundational problems in the short term, new architectures and training methods could emerge to address these issues.

Improving Research Education with Generative AI

A question arises about using generative AI in research education, emphasizing the importance of adhering to established norms and guidelines.

Journals like Nature have set criteria for using generative AI, such as prohibiting it from being listed as a co-author on articles, highlighting ethical considerations in academic writing.

Tools for Enhancing Understanding

Students should explore tools like Explain Paper, which allows them to ask questions about articles' content, results, and conclusions. This aids in critical thinking but does not replace reading the original material.

Such tools can help students prioritize literature during their reviews by providing an initial filter for relevant articles.

Data Quality Concerns

The discussion touches on whether generative AIs use their output data for retraining. Ideally, they should rely on human-reviewed data to ensure quality and reduce biases.

However, if generative AI is trained on social media outputs, it may lead to a feedback loop of inaccuracies due to self-reinforcement.

Challenges in Legal Applications

The speaker highlights challenges faced when using generative AI in legal contexts. An example is given where an attorney relied on incorrect case law generated by an AI tool.

While generative AI can assist with summarizing extensive legal documents or guiding information searches, users must remain cautious as errors are ultimately their responsibility.

Impact on Employment

There’s speculation about whether generative AI will replace human jobs. While some roles may shift towards machine reliance (e.g., translation), complete replacement is unlikely.

Automated translation services are increasingly used due to cost-effectiveness and speed despite potential quality issues compared to human translators.

Differentiating Human Work from AI Output

The challenge of distinguishing between human-generated work and that produced by generative AIs is discussed. Implementing varied assessment activities throughout a semester can help educators gauge student understanding better.

Understanding the Role of AI in Education

Identifying AI-Generated Responses

The use of specific vocabulary and technology in academic responses can indicate whether a response is human-generated or produced by generative AI. Excessive length in explanations may also raise suspicion.

Educators can compare student responses with those generated by generative AI to identify similarities, which helps differentiate between human and machine outputs.

Incorporating AI into Academic Settings

While there are benefits to integrating generative AI into certain subjects, it should not be universally applied across all courses. The decision depends on the professor's expertise and the course structure.

Innovation in education must have a clear purpose; thus, incorporating AI should only occur where it adds real value rather than being implemented indiscriminately.

Companies behind AI models typically avoid using copyrighted content; however, they may inadvertently collect such material through web crawling techniques.

Instances have occurred where users recognized their own code within suggestions made by an AI tool, highlighting potential copyright issues that need addressing.

User Responsibility with Sensitive Information

Users must exercise caution when providing licensed texts or confidential information to generative AIs, as this data could be used for training future models without consent.

It is crucial for users to read terms of service regarding how their uploaded content will be utilized by these tools to safeguard personal and proprietary information.

Benefits and Limitations of Generative AI

Generative AI can significantly enhance self-study opportunities for students due to its interactive nature but comes with inherent risks of inaccuracies that learners must recognize.

Students should remain aware that while generative AIs are powerful tools for learning, they require careful review and validation of the information provided.

Citing Generative AI in Academic Work

Different universities have guidelines on citing generative AIs. For example, APA style recommends including details about the tool used, date accessed, version number, and ideally a complete prompt used for generating responses.

Proper citation practices help maintain academic integrity by ensuring transparency about sources used in research or assignments involving generative AIs.

Utilizing Generative AI for Educational Tools

How Do Evaluation Rubrics Function in Generative AI Contexts?

Understanding Evaluation Rubrics

The effectiveness of evaluation rubrics in generative AI environments is discussed, emphasizing their role as powerful assessment tools for measuring student performance and providing feedback.

Caution is advised regarding the weight given to certain evaluation activities that may be easily resolved through generative AI, suggesting some components should not heavily influence overall assessments.

The Role of Teaching Assistants

A participant notes that teaching assistants will likely become standard across universities, serving as a first line of support for students alongside educational materials.

The assistant's role includes providing immediate feedback to students, enhancing the learning experience by addressing queries before escalating them to instructors.

Resources for Generative AI Tools

A student inquires about guides on generative AI tools; a compilation prepared by the eLearning Innovation Center is mentioned as available with presentation materials.

Prioritizing Use Cases in Education

There’s a discussion on whether teaching about generative AI should take precedence over other use cases. While it’s important, the impact of teaching assistants is also highlighted as crucial.

The speaker clarifies that the order presented was based on immediacy rather than importance, indicating all aspects are relevant to current educational standards.

Information Retrieval and Critical Thinking

Questions arise about whether increasing data availability from generative tools will diminish traditional documentation searches. Current search engines already limit information retrieval methods.

It’s noted that while search engines can summarize information effectively, critical thinking remains essential when evaluating sources beyond just accepting top results from searches.

Risks and Reliability of Generative Tools

Concerns are raised regarding risks associated with using generative AI for sourcing references and bibliographies. Specific mention is made of ChatGPT's limitations compared to Bing Chat's source citation capabilities.

Emphasis is placed on ensuring reliability by checking sources provided by generative tools, highlighting the need for users to verify information accuracy independently.

Creative Aspects and Copyright Issues

The conversation shifts towards creative applications of AI, particularly concerning copyright issues related to artists' works used in training models for image generation.

Regulation of Generative AI Training

Importance of Licensing and Regulation

The discussion emphasizes the need for awareness regarding the content shared online, as it may be used to train generative artificial intelligence (AI).

It is crucial to establish clear intellectual property licenses that restrict such uses, ensuring individuals are informed about how their content might be utilized.

Regulatory measures should include punitive actions against misuse of material, even if individuals have explicitly requested non-use for training purposes.

Educational Initiatives on Generative AI

A question arises about whether the web offers training in generative AI for educators; it is noted that the Learning Innovation Center is preparing relevant courses.

Although details are not fully disclosed, there is an indication that more information will be provided by the Learning Innovation Center in due course.

Future Webinars and Engagement

The session concludes with a note that this webinar is part of a series focused on significant educational topics, including generative AI.