Thought Leadership Thought Leadership

Towards a More Equitable Learning Management System Experience with Large Language Models*

Written by:

Sambhavi Chandrashekar, Global Accessibility Lead, and Rajesh Talpade, SVP Product Management

Fueling our mission with AI

“The amount of intelligence in the universe doubles every 18 months.”

- Sam Altman, CEO, OpenAI

This statement signals the shift in the focus of technology from automation of processes to automation of thinking. Sam intended it as a sequel to Moore’s Law, which states that “the number of transistors in an integrated circuit (IC) doubles about every two years.”

Figure 1: ChatGPT depicted as an integrated circuit (Photo by BoliviaInteligente on Unsplash)

OpenAI is a private research laboratory that developed ChatGPT, the fastest-growing app of all time. It had 100 million active users two months after its launch while TikTok took nine months to reach that number. ChatGPT is based on a Large Language Model (LLM) called Generative Pretrained Transformer (GPT). LLMs are artificial neural networks that are trained to look for patterns in large amounts of text data and then use those patterns to predict what the next word in a string of words should be. Advances in LLMs are taking the world by storm. New versions of GPTs with disruptive features are emerging with hyperbolic speed. GPT-4 released in March by OpenAI helps users generate more accurate and effective queries using natural language prompts.

D2L stays on top of emerging technologies and leverages them to transform the way the world learns. Central to our mission is our relentless focus on reaching every learner regardless of their abilities. This requires that our learning management system (LMS) remains accessible to instructors and learners with disabilities.

This post explores common challenges faced by instructors and learners with disabilities online and how we can leverage the capabilities offered by LLM-based technologies to make online teaching and learning experiences more accessible for them.

Challenges Faced by Instructors and Learners with Disabilities

Vision, hearing, dexterity, and cognition are abilities necessary for engaging in teaching and learning in an online environment. These abilities help with perceiving and understanding digital content and with operating the controls on the digital interface.

Perceiving is primarily done visually, and sometimes auditorily.
Operating primarily requires vision, dexterity (or use of hands), and sometimes hearing.
Understanding requires some level of cognitive ability. The UNESCO’s International Standard Classification of Education specifies at or under the lower secondary education level.

Instructors and learners with disabilities face challenges in teaching and learning in an online learning environment due to complete or partial loss of vision, hearing, dexterity, or cognition.

People with vision loss use an assistive tool called a screen reader that converts the text on the screen into audio or tactile braille. Images and videos need text descriptions for a screen reader to parse them.
People with hearing loss require a text version of audio content in audio and video materials.
People with dexterity challenges benefit from systems that allow voice-driven interaction. Thus, offering multimodal access is key to accessibility for people with sensory and physical needs on the LMS.
People with cognitive needs require simplicity, predictability, and consistency in design and in communication. They also find multimodal access and reinforcement helpful; and communicating in their primary language is easier for them.

LLMs have capabilities for modal and language translation, natural language processing (NLP), and personalization, all of which could be leveraged to make online teaching and learning using an LMS more accessible.

Leveraging the Capabilities of LLMs for Accessibility

About LLMs

At the root of all LLMs is Generative AI, which is a set of algorithms that can be used to create new content, including audio, code, images, text, simulations, and videos. DALL-E2 and Midjourney are models that generate images using text as input to create AI-generated art. GPT is an LLM that generates text using text or image as input.

There are several interesting applications based on LLMs. ChatGPT and Bard are conversational tools based on GPT 3.5. AI-driven code-accessibility checkers like Equally.ai’s Flowy help with accessible coding. Voiceitt and Whisper are tools for speech-to-text conversion that can discern even muffled audio or heavy accents.

OpenAI released GPT-4 in March 2023. It is a multimodal model trained on both images and text that can generate text by taking images as input. Unlike with the earlier versions, OpenAI is not revealing much about the data used or the architecture of GPT-4. However, the benefits offered by LLMs are fascinating regardless of the potential risks.

What LLMs offer for LMS Accessibility

LLMs can automatically generate supplemental materials and materials compatible with a person's assistive tools as detailed below:

Image to text: LLMs can parse through images and create a text description. This can be used to create alternative text for images and videos for those with vision loss to read with their screen reader.

Text to speech: LLMs can convert text to speech. Converting text content into audio files and audiobooks can help learners with vision loss. These could also be useful to sighted learners while they are commuting or doing other visual tasks.

Text to braille: LLMs can convert visual text to electronic braille content. This can be useful for those with vision loss or vision and hearing loss using electronic braille readers, or printed braille on paper.

Speech to text: LLMs can accurately recognize speech and understand the context of conversations. This helps in creating voice-driven interaction and communication interfaces for those with vision or dexterity loss and automatic video captioning and translation services for those with hearing loss and language needs. It helps learners with speech impediments by picking up their unique speech patterns, recognizing any mispronunciations, and normalizing speech before creating an output of audio or text.

Language translation: LLMs can provide multilingual support for the LMS. This can help learners whose first language is not the same as the language of instruction by translating text or audio material to the language they need it in.

Simplification of text: LLMs can identify and suggest changes to content that may be difficult for some users to understand. It can also summarize text. This can help users who have learning disabilities and cognitive constraints by helping them obtain simpler and more concise learning material.

Natural Language Processing: LLMs can understand the queries and responses of users and provide them with more accurate and personalized responses. This can help in providing instructors and learners with a learning environment that is easier to navigate and more user-friendly. Creating conversational user interfaces makes interaction easier for users with cognitive constraints.

Personalization: Given inputs about user behavior and preferences, LLMs can suggest content and activities that are tailored to each individual user. This can make the learning experience cognitively more engaging and effective. Personalization makes learning platforms more adaptive, adjusting to the individual needs of each learner to help them learn more effectively.

In summary, LLMs can improve the online learning experience on an LMS for instructors and learners with disabilities by providing support for generating content in alternative formats, improving communication on the LMS, and enabling adaptive learning.

Risks with adopting LLM-based solutions

GPT-4 raises great concerns about lack of accuracy, biases, and reasoning errors in the text generated. While considering accessibility solutions based on LLMs, the following risks must be kept in mind and handled carefully:

Inaccuracy and bias: LLMs derive all their knowledge from accumulated text and have no discretion or access to real-world, embodied referents. The facts generated could sometimes be inaccurate or biased, leading to learners receiving misleading, unfair, or discriminatory information.

Plagiarism: Using the information generated by LLMs could result in learners submitting content that is considered plagiarized due to lack of proper citation of the original creators, resulting in Intellectual Property and Copyright issues.

Privacy concerns: Personal information being revealed in the data being sent out for conversion and through the prompts being provided for chats could lead to privacy concerns. These need to be carefully thought through with appropriate privacy assessment of tools.

We can use LLM-based tools to generate content, but their accuracy and relevance within the context cannot be fully relied upon. Ironically, there are AI-based tools like GPTZero and ZeroGPT that claim to detect if a piece of content is generated by a human or AI. Commercial solutions for detection of originality in content such as Turnitin Originality and Copyleaks could be of interest to the education community.

While the onslaught of AI is unstoppable, we can respond positively by teaching learners to apply critical thinking to come up with ways to add value to their writing and avoid plagiarism. In this context, communities built and nurtured on the LMS using the LLMs’ natural language processing capabilities can provide the much-needed scaffolding and support for managing limitations relating to inaccuracy, bias, privacy concerns, and plagiarism by:

providing validation and verification through discussion,
helping the development of critical thinking, and
providing emotional support for overwhelmed users.

Cheerful multiethnic students having high five with teacher.

Figure 2: Cheerful multiethnic students having a high five with teacher (Photo by Kampus Production)

Community can also play a key role in supporting learners across modalities, like a member with vision loss can ask a sighted member to verify and confirm the accuracy of a generated alternative description for an image.

LLMs are here to stay

LLMs can be used to create more inclusive teaching and learning experiences in an LMS by improving access to content, enabling adaptive learning, and enhancing communication to promote community and collaboration. Embracing such technological advances while managing the associated risks can help learners with disabilities achieve their learning goals with improved efficiency and greater autonomy.

The solution is not for educators to prohibit the use of ChatGPT and other AI tools but they must redesign pedagogy to enhance critical thinking and creativity. Policing education might not be desirable or feasible. In the words of Dr. Jutta Treviranus, leader in AI ethics and Inclusive Design:

“If we need to police education, we are doing something wrong. If a machine can do what we are teaching our students to do, we are teaching our students to be machines.”

LLMs are here to stay and grow. Leveraging them to promote inclusive online learning facilitates heutagogy, or self-directed learning, where learners are in-charge of their own learning. We can use AI to promote effective online learning for all, regardless of their abilities by supporting education that nurtures critical and creative processes.

*Technologies and software/service providers are mentioned in the blog post for purely illustrative purposes. Their mention does not imply that D2L endorses or in any way recommends their usage.