In 2019, the United Nations Educational, Scientific, and Cultural Organization (UNESCO) formulated a plan for the integration of artificial intelligence with education, and suggested that AI has the potential to create a more equitable, efficient, and innovative educational system (1). Although this statement was made prior to Covid-19’s rapid proliferation, it warrants renewed consideration in light of the calamitous effects the pandemic has had on the education system. The abrupt shift to remote learning and subsequent return to in-person instruction have revealed significant decreases in learning outcomes and readiness to meet academic standards (2); decreases that a return to normalcy has little hope of addressing. Now, if ever, is the time to look towards innovative means of enhancing education, and artificial intelligence could prove to be the change required to bring about meaningful academic growth.
Recent studies indicate that in the United States, school closures and remote learning have resulted in significant decreases in educational outcomes for K-5 students that roughly equate to a five month deficit in math skills, and a four month deficit in reading comprehension (2). Furthermore, this data assumes equity in education, which is not inherently accurate, as learning outcomes have been further stunted in low-income groups, and in students with intellectual disabilities (3). On the global scale, pandemic related budget cuts have hit historically poorer countries hard, and widened the gap in educational quality (measured by reading level) between low to lower-middle income countries, and upper-middle to high income countries (4). The dire state of the current educational system caused the United Nations to announce that
“Covid-19 has wiped out 20 years of educational gains. (5)”
While not a panacea for all problems in education, AI offers a plethora of solutions to help mitigate the damage done in recent years, and is currently reshaping the way students learn in classrooms all around the world. Below are six ways that artificial intelligence can be leveraged to enhance the ever changing educational landscape for students and teachers alike.
1. AI Enhanced Personalized Learning
“Teach to the middle” is an old, outdated adage. The practice of giving lessons to the modal classroom population may once have been necessitated by large classroom sizes and pressures to complete yearly curriculum, but it is entirely unequipped to address the individual needs of today’s students. When lessons are aimed at the average student, those above and below the line may feel that they are not being challenged, or that they are falling behind respectively.
Every student is different, therefore the most effective pedagogical model must differ between students. As it is unlikely that teachers have the time or resources to enact large scale differentiated instruction, AI backed personal learning systems have stepped in to give students greater control over their academic trajectory. Century Tech, one of the more promising personalized learning systems, is based on the idea that “one size fits one” (6). The two core processes of Century Tech are diagnostic assessments, and the recommendation of “micro lessons”. After a student completes an ungraded diagnostic in math, English, or science, Century Tech categorizes student strengths and weaknesses, and recommends supplementary materials (6). Each student receives a unique pathway of course materials, assessments, and memory boosting exercises that updates to accommodate their performance (6).
The entire process is made possible by the ample store of user data - in the form of diagnostic assessments - that Century Tech has accumulated since the program’s launch in 2016, and continues to collect from new students every day. This increasingly diverse store of student data enables Century Tech’s natural language processing algorithms to account for a wide variety of student responses during the diagnostic phase, and facilitates the use of a clustering framework to group students based on their limitations (6). Once a student’s needs are identified, Century Tech assembles an individualized lesson plan that sporadically introduces additional diagnostic assessments in order to monitor student progress. Century data scientist Michael Ma likens Century Tech’s autonomous recommendation system to the algorithm used by streaming services to suggest movies and TV shows, albeit more focused on giving students “more of what they need, not just more of what they want” (6).
While Century Tech is already tried and tested - it was recently adopted by over 700 schools in Belgium (7) - the potential scalability is quite impressive, as the machine learning algorithms responsible for its success can only improve and offer more nuanced personalization as more student data is collected.
2. Improving Accessibility for Students with Disabilities
Approximately 14.4% of students in the United States are classified as having a form of disability (8). Of the 7.3 million students with disabilities, the most prevalent categories are students with specific learning disabilities, students with speech or language impairments, students with autism spectrum disorder (ASD), and students with intellectual disabilities (8). Unfortunately, many of these students do not get the help they need, or perhaps even more disconcerting, may not receive the proper diagnosis at all. Forty nine states report a shortage of specialized instructional support personnel, or otherwise under qualified personnel, resulting in a disparity in academic accessibility (9). While artificial intelligence can not foreseeably replace in-person aid, it does offer a variety of tools for both diagnosing and empowering students with disabilities.
The key to the diagnostic process are machine learning algorithms, and their exceptional ability to identify patterns, as students may exhibit patterns of behavior that are consistent with specific disabilities. Take for example the Australia based startup Dystech and their titular app, which uses machine learning to screen children for dyslexia and dysgraphia (10). Dystech is trained off a dataset consisting of audio recordings and writing samples from children with and without dyslexia or dysgraphia, and screens children by comparing their ability to read aloud and write to those in the dataset (10). The primary metrics used by the AI in the diagnosis of dyslexia are reading reaction time (the time it takes for a child to begin reading once presented with a word), and overall reading time (10). The primary metrics used in the diagnosis of dysgraphia include written word spacing, consistency of letter size, regularity of letter slant, and the amount of pressure applied on the paper by the child (10). Similar processes have been proposed for the diagnosis of autism spectrum disorder, but the binary classification model currently lacks the nuance to account for the vast range of behaviors potentially exhibited by those on the autism spectrum (11).
Effective education starts with effective communication, but simply being understood can present a challenge for millions of students with disabilities. One recurring problem with present voice recognition technologies is that the algorithm is not trained to recognize “atypical speech”, making the technology all but unusable for people with speech impairments. Luckily, this obvious gap in the market is beginning to be filled by a variety of specially trained speech-to-speech AIs, such as Voiceitt, an app that recognizes individualized speech patterns and repeats user phrases with clear diction (12). Voiceitt users train the app themselves by repeating phrases aloud until the machine learning model begins to recognize their unique speech patterns, pronunciation, and cadence (12). Once Voiceitt learns how the user pronounces specific phrases it can predict what the user intends to say with very little prompting (12). Although currently limited to pre-trained phrases, Voiceitt has recently partnered with like-minded organizations to spearhead the EU funded Nuvoic Project, with the goal of advancing Voiceitt’s core technology to accommodate continuous speech recognition - that is, the recognition of sentences and phrases that are not pre-trained (13).
Nonverbal or “augmentative alternative communication” devices have similarly been improved with AI, allowing for a greater range of expression with pictogram based communication systems. Livox, a pictogram based communication app geared towards people with limited mobility and speech capabilities, utilizes machine learning algorithms to analyze user data and provide context specific pictogram recommendations (14). Livox takes into account a wide range of factors including frequently used pictograms, approximate time of usage, and GPS location, in order to recommend pictograms most likely to be used in a given scenario (14). For example, if Livox detects that the user is at home in the morning, it may display pictograms apt for communicating that the user wants breakfast, needs to use the restroom, or wishes to go outside. Additionally, Livox employs natural language processing to recognize spoken trigger words and recommend pictograms best equipped to answer questions, enabling more fluent and efficient conversation (14).
3. AI Assisted Tutoring
The concept of the Intelligent Tutoring System (ITS) can be traced back to the early 1970s, with its maturation coinciding with the advent of the personal computer nearly a decade later (15). While modern ITSs resemble early electronic tutoring systems about as much as modern ultrathin laptops resemble the first home computers, they still share the core principles of problem diagnosis and tailored remediation. Broken down into its basic components, an ITS is programmed to receive input from a student in the form of a written or mathematical problem, offer a solution or correction of the student's work, and present a unique progression path for the student (15). This sort of bespoke learning experience involves a laborious programming progress, as the programmer must teach the ITS to identify and perform every possible action a student may take at every step of a given problem (16). Programming by demonstration in this way can take as long as 50-100 hours per 1 hour of tutoring (16).
However, artificial intelligence has the potential to drastically reduce the time and skill necessary to create a ready to use ITS. Researchers out of Carnegie Mellon’s Human-Computer Interaction Institute employed a novel process that utilizes apprentice learners, or AI agents that act as simulated students and can be taught to solve various types of problems (16). In the Apprentice Learner system, a teacher demonstrates how to solve a problem to an AI - in the case of the study, multi column addition -, and then the AI attempts to solve problems in a similar style. If any inaccuracies arise, the AI is corrected by the teacher, thereby forcing it to learn the full scope of possible inputs, and enabling the AI to make generalizations about future problems (16).
More technically, the Apprentice Learner system is designed to exhibit model-tracing completeness, which is defined as “recognizing all intended correct actions as correct, and no incorrect action as correct for all possible states in a problem” (17). In the example of multi column addition, the AI draws on different mathematical rules depending on the carry pattern (i.e. when and where a 1 must be carried to a different column) of a problem, and the teacher uses an interface to affirm or deny the correctness of the rule used (17). If an incorrect rule is used, the AI may implement a different rule, or request a demonstration from the teacher; in either case, the AI learns as it is corrected (17).
Once the AI understands a concept, it in turn can teach and correct students. Teachers need only be familiar with the content they are teaching, as apprentice learner powered ITS do not need to be programmed in a traditional sense. Although not quite within reach, the ideal ITS could be trained by a teacher in just a few hours per 1 hour of tutoring, and operate in a variety of subjects including algebra, chemistry, and English (16).
4. Monitoring and Predicting Student Progress Using AI
What if a teacher could identify students at risk before they fell behind? In their role as educators, teachers are privy to large quantities of student data, and AI has the ability to consolidate this data into meaningful insights about student progress. Taken as a whole, student progress can be a nebulous concept, and one that is better organized into specific metrics: student engagement, likelihood of a student to meet academic standards, and student retention.
Numerous techniques have been proposed for monitoring student engagement. In a traditional classroom setting, the most common method is by way of student surveys. In online courses, unique data such as the number of mouse clicks a student makes, or duration of inactivity can be analyzed. However, the availability of facial detection AI, has led to a new way of monitoring student engagement. By recording students in either online or offline classrooms, convolutional neural networks are able to recognize student expressions, and categorize them based on their consistency with low, medium, or high levels of engagement (18). Assuming that an educational institution received consent from students to be monitored, and could sufficiently address privacy and security concerns, this information would allow educators to adapt their lesson plans to better maintain students' attention throughout a course.
Predicting student academic performance has distinct limitations. For a machine learning model to make a prediction, it must first receive student data in the form of academic records; the more data it receives, the more accurate it becomes. The challenge is getting enough student data early on so that the prediction can be acted upon. The breadth of data collected from online learning management systems (LMS) has provided the best solution for identifying early indicators of student performance (19). In a study produced by Stockholm University’s Department of Computer and System Sciences, researchers concluded that the most accurate variables for predicting student performance were active participation in the LMS, initial quiz and assignment scores, forum discussion, and previous knowledge of featured concepts (19). Using a random-forest based prediction model, researchers found that the aforementioned metrics were positively correlated with student performance later in the course (19).
Additionally, the machine learning algorithm assigned “influence values” to individual variables, which revealed the level at which specific factors influenced student performance (19). For example, participating in online discussion had a significantly greater influence on student performance than did reading articles from external sources (19). In this approach, a wide range of factors collected over a short period of time substitutes a large quantity of a single factor (i.e. assignment or test scores) collected over a long period of time, enabling earlier prediction of student performance. If detected early enough, AI prediction could be used to identify students who would benefit from academic intervention.
No prediction can be 100% accurate, especially when the prediction is ultimately dependent on the volatility of human choice. That being said, student retention is being predicted with surprising accuracy by the combined usage of data mining and machine learning techniques. One of the most successful studies of student retention utilized predictive analytics to predict dropout rates in Chilean universities (20). Drawing on numerous student attributes including university entrance exam scores, economic quintile, high school and university performance, machine learning algorithms were able to predict dropout rates of first, second, and third year students with greater than 80% accuracy in all cases (20).
5. Grading and Assessing Students Using AI
Students are not the only ones who have fallen victim to the educational crisis. Many teachers report feeling underappreciated and overworked. A survey conducted by the UK’s National Education Union in 2021 found that one in three teachers plan on leaving their profession in the next five years, with over 50% citing workload as a determining factor (21). Similarly, the Education International Report on the Status of Teachers in 2021 reported that 55% of respondents felt that teacher workload was unmanageable (22). Moreover, empirical evidence suggests that overworked and overcommitted teachers are more likely to exhibit signs of emotional and physical exhaustion, and presenteeism, and conversely, experience reduced job satisfaction (23).
One way of alleviating teacher workload is to automate administrative tasks like grading. AI has shown proficiency in helping to grade repetitive assignments with little variation, namely simple math, multiple choice, and fill-in-the-blank questions. In fact, students and teachers may already be familiar with AI-assisted grading, as popular grading programs like Gradescope have been employing AI-assisted grading tools for years (24). While Gradescope’s AI does not directly grade assignments, it does drastically reduce the time required to mark individual submissions by automatically analyzing handwritten student responses and grouping similar answers (24). A human grader can then assign a grade to each answer group rather than grade individual submissions, a process that Gradescope co-founder Arjun Singh is confident will reduce grading time by 80% (25).
“AI-assisted grading, which uses deep neural network methods for few-shot learning and is a type of supervised machine learning, can help faculty and teaching assistants cut grading time by 80% or more.”
Less widely employed are AIs designed to grade long form written work (i.e. essays). By analyzing human graded essays, AI graders are increasingly capable of imitating the grading style and metrics used by human graders, and applying them to student essays (26). AI graders can recognize numerous features of student writing including vocabulary, syntax, sentence structure, and content focus (26). That being said, AI graders are ultimately operating on an advanced form of pattern recognition which, though effective in some applications, can lead to the overvaluation of specific metrics. For instance, the AI might overvalue certain word choices, but fail to recognize the lack of cogency in a sentence, leading to a high scoring but unintelligible essay (27). In their current form, automated essay graders work best in tandem with human graders to circumvent errors and bias.
All things considered, AI can still play an important role in aiding teachers with the essay grading process. For example, AI plagiarism checkers have access to unrivaled quantities of published and submitted work, often making them superior to their human counterparts in scrutinizing student work for copied material. Modern plagiarism checkers like Grammarly can even detect paraphrased content, and incorrect or missing citations (28). Furthermore, students can use plagiarism checkers themselves to avoid submitting work that would compromise their academic integrity.
6. Leveraging AI to Deconstruct the Language Barrier
Pursuing an education can be challenging in itself, but the challenge can appear nearly insurmountable when compounded with the extra stress and time required to navigate an education system in a language different to one's own. This section will primarily focus on the United States, as the US has both the largest number of English as a second language learners (ESL), and university level international students (29).
The average ESL learner requires 4-7 years to reach English proficiency, however it would be incorrect to assume that proficiency results in ideal academic outcomes, as ESL students historically graduate at lower rates and receive lower test scores than non-ESL students (30). Accordingly, the use of effective, engaging English lessons are essential for ESL students, especially during their most formative years. The recent merger between Edwin.ai and MyBuddy.ai (31), two effective English lesson apps in their own right, aims to provide a solution for millions of children worldwide. Buddy.ai in its current iteration combines Edwin.ai’s ESL curriculum and adaptive learning algorithm - not dissimilar to previously discussed personalized learning systems - with Mubuddy.ai’s virtual AI assistant (31). Natural language processing allows students to converse directly with Buddy.ai, an animated robot, and practice their spoken English. Buddy.ai’s gamification model is designed to engage students aged 4-10, whilst correcting their pronunciation, and expanding their vocabulary (32).
2019 and 2020 saw a large portion of academic institutions switch to distance learning, and as a result, many university students were exposed to massive open online courses (MOOC). With the necessity of in-person attendance removed, MOOCs grant students the freedom to expand their curriculum to include courses from international institutions. While it is not uncommon for universities to include courses specifically geared towards international students, the emergence of AI translators capable of live, multilingual translation make it possible for students to take any course regardless of language. This dismantling of the language barrier is facilitated by neural machine translation (NMT), a strategy of machine translation that relies on neural networks to accurately receive input in one language, and deliver output in another (33). Whereas other machine translation systems focus on translating individual words or phrases separately, NMT takes into account entire sentences, or sequences of sentences before making a translation (33).
NMT has its uses in the context of text-to-text, and speech-to-text translation, but arguably the most intriguing implementation of NMT is in live speech-to-speech translation, such as is found in Waverly Labs’ Ambassador Interpreter system (34). The Ambassador is an over the ear device capable of translating 42 different languages and dialects directly from the speaker into the preferred language of the user (34). The applications of NMT technology are multifarious, but its use in education could enable the direct translation of lectures and coursework for any student, regardless of spoken language.
The education system is invaluable, but that is not to say that it can’t be improved. The goal of this article is not to suggest that AI can or should replace human instruction, but rather that it can be implemented effectively to forward educational goals, and expedite both the learning and teaching process. Admittedly, it would take time for students and teachers to adapt to AI in the classroom, but AI’s efficiency can more than make up for any lost time. The pandemic disrupted nearly every aspect of the educational system; might it be possible that further change is necessary to get it back on track?
Sources:
https://en.unesco.org/artificial-intelligence/education
https://www.mckinsey.com/industries/education/our-insights/covid-19-and-education-the-lingering-effects-of-unfinished-learning
https://www.hrw.org/news/2020/06/15/school-closures-particularly-hard-children-disabilities
https://www.worldbank.org/en/news/press-release/2021/02/22/two-thirds-of-poorer-countries-are-cutting-education-budgets-due-to-covid-19
https://sdgs.un.org/goals/goal4
https://www.century.tech/news/how-does-centurys-ai-work
https://qz.com/1577451/century-tech-signs-deal-to-put-ai-in-700-classrooms-in-belgium
https://nces.ed.gov/programs/coe/indicator/cgg
https://specialedshortages.org/about-the-shortage
https://dystech.com.au/about-us/dyslexia-dysgraphia-screening-explained
https://onlinelibrary.wiley.com/doi/full/10.1111/jep.13527
https://www.forbes.com/sites/gusalexiou/2021/06/30/voiceitt-app-for-atypical-speech---a-triumph-in-disability-co-design
https://voiceitt.com/nuvoic-project
https://www.researchgate.net/publication/333609740_Using_Artificial_Intelligence_for_Augmentative_Alternative_Communication_for_Children_with_Disabilities
https://medium.com/@roybirobot/how-intelligent-tutoring-systems-are-changing-education-d60327e54dfb
https://www.cmu.edu/news/stories/archives/2020/may/intelligent-tutors.html
https://dl.acm.org/doi/pdf/10.1145/3313831.3376226
https://onlinelibrary.wiley.com/doi/epdf/10.1111/exsy.12839
https://www.frontiersin.org/articles/10.3389/frai.2021.723447/full
https://pubmed.ncbi.nlm.nih.gov/33923879
https://www.theguardian.com/uk-news/2021/apr/08/one-in-three-uk-teachers-plan-to-quit-says-national-education-union-survey
https://www.ei-ie.org/en/item/25366:new-global-report-points-to-overworked-underpaid-and-undervalued-teaching-profession
https://onlinelibrary.wiley.com/doi/epdf/10.1002/smi.2820
https://help.gradescope.com/article/51lz101z28-instructor-assignment-ai-grading-format#math_fill_in_the_blank_and_text_fill_in_the_blank
https://www.disruptordaily.com/ai-in-education-use-case-7-gradescope
https://elearningindustry.com/artificial-intelligence-new-role-in-education-automated-paper-grading
https://www.npr.org/2018/06/30/624373367/more-states-opting-to-robo-grade-student-essays-by-computer
https://www.grammarly.com/plagiarism-checker
https://www.pewresearch.org/fact-tank/2018/10/25/6-facts-about-english-language-learners-in-u-s-public-schools
https://www2.ed.gov/datastory/el-outcomes/index.html
https://www.businesswire.com/news/home/20200302005207/en/Y-Combinator-Backed-EdTech-Startup-Edwin-Merges-With-Voice-Assistant-Company-MyBuddy.ai-to-Teach-Children-Spoken-English
https://buddy.ai/en
https://www.memsource.com/blog/neural-machine-translation
https://www.waverlylabs.com/shop/ambassador-interpreter