AI in Instructional Design

Adaptive learning platforms use Artificial Intelligence and data analytics to create personalized learning experiences. How these platforms can improve the way we learn?  Is there a risk that constant evaluation and immediate adjustments could result in an overly surveilled learning experience?  Moreover, as we develop AI-driven learning activities, how do we strike a balance between innovative pedagogy and the protection of privacy in adaptive learning contexts, considering both educational and organizational implications?

People illustrations by Storyset

Adaptive learning platforms are technologically advanced tools that allow creating personalized learning experiences by using artificial intelligence (AI) and data analytics. Here are some key aspects of adaptive learning platforms:

• Personalized Learning Paths: Adaptive learning platforms tailor the learning experience to each individual’s needs, strengths, and areas for improvement. For example, a math app adjusts the difficulty of problems, providing more challenging questions as learners progress based on their performance.

• Real-Time Adaptation: Learner performance is continuously analyzed, and the content and activities are adapted in real-time. For example, an online language course offers additional practice exercises for grammar topics where the learner shows weaknesses.

• Diverse Content Formats: Adaptive learning platforms often provide content in various formats, such as text, video, and interactive simulations, to cater to different learning styles. For example, a science course includes video lectures, interactive lab simulations, and text-based readings.

• Continuous Assessment and Feedback: These platforms offer ongoing assessments and instant feedback to help learners understand their progress and areas needing improvement. For example, an online coding platform provides immediate feedback on code submissions, highlighting errors and suggesting corrections.

• Data-Driven Insights: Adaptive learning platforms use data analytics to provide insights into learning patterns, helping educators and learners make informed decisions. For example, a dashboard tracks student progress and identifies topics where multiple learners are struggling, allowing instructors to address these areas in more detail.

Review an example of presentation with interactive knowledge checks generated by an adaptive learning platform: SC Training.

The computer icon below leads to the presentation Engagement as a UDL Principle (21-slides long, it opens in a new tab).

💻

What are your initial impressions, and what could we consider adding or improving, keeping in mind that we’ve already removed about 10 redundant slides?

Practice:

Design a compact learning object using an adaptive learning tool such as SC Training. Evaluate the AI-generated learning object according to pedagogical, social, and technological quality aspects. Share the observations and the URL of the learning object with peers for comments.

1. Select a Subject or Topic: Choose a subject or topic you want to teach. This could be anything from a math concept to a historical event or a scientific principle.
   
   

2. Identify Learner Needs: Understand learner needs related to their level of preparedness, unique interests, cultural affiliations, and identities.
   
   

3. Design the Learning Object: Consider the following examples for your prompts:
   
   

   • Math Concept: “Create a learning object that explains the Pythagorean theorem using interactive diagrams and practice problems.”

   • Historical Event: “Design a learning object that covers the causes and effects of the French Revolution, including videos and primary source documents.”

   • Scientific Principle: “Generate a learning object that demonstrates the principles of Newton’s laws of motion with simulations and real-world examples.”

4. Publish the AI-Generated Learning Object: Obtain a public preview link.
   
   

5. Evaluate the Learning Object:
   
   Pedagogical Aspects:
   
   

   • Engagement: Evaluate how well the learning object captures and maintains learners’ interest. Look for the use of interactive elements that make learning enjoyable and engaging.

   • Differentiation: Assess the variety of content formats included in the learning object. Check if the text, video, and interactive simulations are properly used to cater to different learning needs.

   • Feedback: Analyze the quality and timeliness of feedback provided to learners. Ensure that the learning object offers immediate and constructive feedback to help learners understand their progress and areas for improvement.

   • Assessment: Evaluate the effectiveness of the assessments included in the learning object. Determine if the assessments accurately measure learners’ progress, and if they provide meaningful insights into learners’ performance.

6. Social Aspects:
   
   

   • Privacy: Check for compliance with relevant regulations on users’ privacy and the transparency of data usage.

   • Ethical Considerations: Assess the balance between the benefits of better learning and the potential impact on the learners’ rights, autonomy and privacy.

7. Technological Aspects:
   
   

   • Usability: Evaluate the ease of use and accessibility of the learning object. Ensure that it is user-friendly and accessible to all learners, including those with disabilities.

   • Data Security: Analyze how the platform handles and secures learner data. Check for robust security measures and protocols to protect sensitive information.

   • Adaptability: Assess how well the learning object adapts to individual learner performance. Look for features that personalize the learning experience based on continuous assessment and feedback.

8. Share the description of your designed learning object, indicating the pedagogical, social, and technological aspects it addresses. Provide the URL of the AI-generated learning object and invite peers to comment and provide feedback.
   
   

Resources:

• SC Training – Creator tool  designed to streamline the creation process of the learning objects that can be accessed anytime, anywhere, on any internet-connected device. Generate course content using AI, which can help brainstorm ideas, reduce research time, and build bite-sized information. Incorporate multimedia, interactive simulations, and quizzes to make learning engaging and fun.

• Meet Khanmigo: Khan Academy’s AI-powered teaching assistant & tutor Khan Academy: Khan Academy uses AI to personalize learning experiences based on individual student performance. It provides tailored practice exercises and instructional videos to help students progress at their own pace. Khan Academy offers free options for AI-powered learning through their platform. They have integrated AI tools like Khanmigo, which is powered by GPT-4, to enhance the learning experience. Khanmigo acts as a personal tutor and teaching assistant, providing personalized support and adapting to the learner’s needs
  
  

• Duolingo: leverages AI to adapt language learning exercises to the user’s proficiency level. It offers a gamified learning experience that adjusts the difficulty of exercises based on the learner’s performance.
  
  

• Quizlet: uses AI to create adaptive study sets and learning tools. It includes features like flashcards, quizzes, and games that adapt to the learner’s progress and help reinforce learning. Quizlet offers free options for both learners and teachers. Students can access study sets, flashcards, and various learning modes without any cost. Teachers can also use Quizlet to create and share educational resources, and they can engage students with in-class games like Quizlet Live
  
  

Customized assistance for learning and research 

NotebookLM, introduced by Google as personalized research assistant, which can on demand transform text and multimedia sources into personalized multimedia instruction (audio and video) with interactive functionalities.

examples of AI toold for rapid ideation and content generation:

Mindsmith.ai

Mindsmith is an AI-powered course authoring tool that takes a holistic approach to instructional design. Instead of just generating text, it helps you structure entire courses from scratch.

• You can input existing documents, web articles, or a simple course outline, and Mindsmith will generate a full storyboard and course content. It's designed to streamline the entire process, from ideation to creating interactive elements like quizzes, scenarios, and flashcards. This makes it an ideal tool for quickly prototyping a new course or converting existing, static content into an engaging e-learning module.

Synthesia

Synthesia is a tool for creating video-based learning content. It allows you to transform text into professional-looking videos with AI-generated avatars.

• Video is a key component of modern e-learning, but production can be time-consuming and expensive. Synthesia removes these barriers by letting you create videos without a camera, actors, or a studio. You simply write the script, choose an AI avatar, and the tool generates a high-quality video with a voiceover in over 130 languages. This is perfect for creating consistent, on-brand explainer videos, welcome messages, or training modules at scale.

Eduaide.Ai

Eduaide.Ai is a generative AI platform built specifically for educators. It focuses on creating a wide range of instructional materials, from lesson plans to assessment tools.

• This tool is particularly useful for the more granular aspects of instructional design. You can use it to generate a variety of learning activities, assessment questions, or even graphic organizers based on a topic or a document you provide. A key feature is its ability to differentiate content with one click, creating variations of a resource to support different learning needs. This saves a tremendous amount of time on the detailed, hands-on tasks that follow the initial course outline.

Illustrations by Storyset

Creating Datasets

1. Data Collection and Curation

The most critical starting point is the data. Algorithms learn from the data they're trained on, so any biases present in that data will be learned and amplified by the model. To address this, you should:

• Examine Your Data Sources: Identify if the data is representative of the entire population the model will serve. For example, is there an overrepresentation of one demographic group?

• Balance and Augment Datasets: Actively work to fill gaps in underrepresented data. This might involve collecting more data or using techniques to augment existing data to ensure fairness.

2. Model Design and Development

The choices made during a model's development can also introduce or reinforce bias. A focus on fairness must be built into the model itself. This includes:

• Defining Fairness Metrics: Decide what "fairness" means for your specific application before you start. For example, does it mean the model has equal accuracy for different demographic groups?

• Bias Mitigation Techniques: Use specific techniques designed to reduce bias during training. These can include algorithmic adjustments that re-weight data or modify the model's objective to optimize for fairness in addition to performance.

3. Evaluation and Monitoring

Bias isn't a one-time problem; it can emerge or change over time. Therefore, a continuous process of evaluation is essential. You should:

• Disaggregate Performance Metrics: Don't just look at a model's overall accuracy. Break down its performance by different subgroups (e.g., age, gender, race) to identify where it might be underperforming.

• Continuous Monitoring: Implement systems to monitor the model's behavior in the real world. As new data comes in, the model's performance on different groups can shift, requiring a new round of data collection or model retraining.

Using Datasets

When addressing algorithmic bias as a user, you have a critical role to play by being a conscious and active participant, especially in how you interact with and understand the data that fuels these systems.

1. Be Critical of the Data's Source and Scope

As a user, you should always question the outputs you receive. When a search result or LLM response seems limited or biased, consider the possible sources of the data it was trained on. Ask yourself:

• Is the information dominated by a single perspective?

• Are certain groups or ideas underrepresented?

• Does the output reflect a historical bias that might be present in the data?

By doing this, you're not just passively consuming information but actively evaluating its integrity and pushing for greater transparency.

2. Proactively Contribute Diverse and Correct Data

Your own online behavior and contributions are a form of data. You can help improve future datasets by being a source of diverse and accurate information.

• Add Missing Information: If you notice a platform has incorrect or incomplete information about a topic or group, contribute accurate data.

• Correct Inaccuracies: When an algorithm provides a biased or false result, use feedback mechanisms (if available) to flag it. Your feedback helps developers retrain models with better data.

3. Use Bias Detection Tools

There are a growing number of tools and browser extensions designed to help users identify potential bias. These tools can analyze search results or social media feeds to highlight potential echo chambers or demonstrate how content might be filtered based on your online profile. Using these tools can make you more aware of how algorithms are shaping your online experience, empowering you to seek out more balanced information on your own.

Resources

Open-Source Libraries

These are Python libraries that data scientists and engineers can integrate directly into their machine learning workflows. They provide a range of metrics and algorithms for both detecting and mitigating bias.

• AI Fairness 360 (IBM): A comprehensive open-source toolkit with over 70 fairness metrics and 10 bias mitigation algorithms. It's designed to help you not only detect bias but also understand and remove it.

• Fairlearn (Microsoft): This toolkit helps you assess the fairness of your model's predictions and provides tools to mitigate unfairness. It allows for a more hands-on, interactive approach to bias detection.

• What-If Tool (Google): This tool, often used within Google's TensorFlow platform, is a visual and interactive way to explore your data and model performance. It allows you to analyze your model's behavior across different subgroups to check for fairness.

Methods and Metrics

Beyond specific tools, the core of bias detection relies on a set of standardized methods and metrics used to evaluate a model's fairness.

• Disaggregated Performance Metrics: Instead of looking at a model's overall accuracy, you break down its performance by different subgroups (e.g., race, gender, age). This reveals if the model is performing better for some groups than others.

• Demographic Parity: This method checks if the model's positive outcomes are distributed equally across different groups. For example, does a loan-approval model approve the same percentage of loans for both men and women?

• Equalized Odds: This is a more advanced metric that ensures a model has equal false-positive and false-negative rates across different groups. This is especially useful in critical applications like healthcare or criminal justice, where a false positive can have serious consequences.

• Explainable AI (XAI) Audits: Using methods like SHAP or LIME, you can understand which features are most influential in a model's decisions. By auditing these explanations, you can detect if the model is disproportionately relying on a sensitive attribute or a proxy for one (e.g., using ZIP codes to infer race).

The principles behind bias detection—such as proactively auditing for representativeness, defining fairness, and continuously monitoring outcomes—are critical for the entire development lifecycle of a digital learning structure. These methods are essential for ensuring that any automated or predictive systems used for pedagogical purposes do not perpetuate bias.