Artificial Intelligence and the Future of K-12 Schools Part 1: Artificial Intelligence

I have followed the progress of Artificial Intelligence (AI) for many years. In recent years, AI has shown that it is and will continue to be a major change agent in our world. The corporate world is well aware of this, and is investing heavily in developing and implementing changes. For example, I just read the following announcement of an event being scheduled by IBM:

Examine the ways advanced wireless technologies, including 5G and Wifi 6, accelerate and advance innovation through AI, the cloud, and the Internet of Things.

Hear how IBM is moving away from infrastructure services and full speed ahead toward a future built on AI and hybrid cloud, and ultimately, quantum computing. (IBM, 12/19/2020, link).

Such change and its underlying vision of the future is now common in the corporate world. It is much less so in the world of education.

I recently read AI and the Future of Learning: Expert Panel Report (Roschelle, Lester, & Fusco, December 2020, link). The 27-page document is based on the work of 22 carefully selected experts in the field of AI in education. Here is the executive summary from the report:

Artificial intelligence (AI), machine learning, and related computational techniques have the potential to make powerful impacts on the future of learning. Technology’s impact on education is often to amplify impacts, regardless of whether the impacts are intended. Due to the accelerating pace of integration of technology in learning environments, the knob on the amplifier is rapidly going from low to high. Impacts on learning, whether positive or negative, could soon have consequences for many more students. Now is the time to begin planning for how to best develop and use AI in education in ways that are equitable, ethical, and effective and to mitigate weaknesses, risks, and potential harm.

This current newsletter is the first of several motivated by this report, and is designed to provide background information that will be useful if you decide to read the full report. It also offers insights into some of the types of problems that AI is addressing, and the progress it has been making. Subsequent newsletters will address goals of education and changes needed to accommodate our changing world, and also delve more deeply into the report cited above.

Increasingly, experts in the field of AI and education are beginning to understand the potentials for AI to make significant contributions to our ongoing efforts to improve education. They (and I) believe that, over the next decade, we will see very large changes in our formal and informal educational systems due to the increasing capabilities of AI, as well as to the increasing availability and quality of the infrastructure required to make this capability readily available to learners of all ages.

The quote from Marvin Minsky given at the beginning of this newsletter shows considerable foresight into the challenges faced by the emerging AI field. He was a pioneer in this field who made a number of major contributions during his long career.

In 2011, I wrote an IAE Blog about Artificial Intelligence and Artificial Muscle (Moursund, 2011, link). The blog poked fun at the idea that humans and prehumans have been developing tools to enhance their physical capabilities for more than three million years, but never referred to them as artificial muscle. Reading and writing obviously are tools to enhance the human brain, but they were never called artificial brain or artificial intelligence. Tools such as an abacus or calculator were never considered to be artificial brains, although they certainly are aids to our brain’s capabilities.

I first became interested in learning about computers and AI shortly before beginning my doctoral studies in mathematics at the University of Wisconsin in 1963. By then, programmable electronic digital computers had been commercially available for more than ten years, and were already in wide use. The field of AI had been defined and established seven years earlier.

One requirement of my doctoral program was to demonstrate the ability to read mathematics texts written in two foreign languages. I chose French and German because I had two years of high school French and had completed second year German as an undergraduate. I was not particularly good in languages, but I studied and practiced diligently and passed both tests on my first tries.

By that time, progress was already occurring in the development of language translation programs, and I wondered how soon AI would become quite successful in this endeavor. It certainly is fortunate that I persisted on my own, because it has taken more than 50 years of further progress since my graduate school days for language translation programs to become good enough to meet some of the contemporary needs. AI still has a long way to go in language translation before it becomes as good as human experts.

So, when you read about the potentials of AI to substantially improve education during the next five to ten years, be very suspicious. The long history of computer-based language translation will help to increase your insights into the challenges.

Language translation has proven to be a very difficult problem in AI, and there are a number of funny stories about its early failures. One measure of the success of translation programs is to use them to both translate some text into another language, and then to translate the resulting translation back into the original language. The following is an amusing example:

Rumors have it that early modules for English to Russian have mistranslated some idioms with amusing results. Translating the phrase “The spirit was willing, but the flesh was weak” to Russian and back to English resulted in: “The vodka was good, but the meat was rotten.” Likewise “out of sight, out of mind” reportedly yielded the phrase “blind and insane.” (Mikkelson, n.d., link.)

I recently used Google Translate to translate the paragraph quoted above, first into Russian and then back into English. Here is the English language result:

It is rumored that some idioms were incorrectly translated from English to Russian in the early modules, which led to funny results. The translation of the phrase “The spirit was ready, but the flesh was weak” into Russian and back into English led to the following: “The vodka was good, but the meat was rotten.” Likewise, the phrase “out of sight, out of mind” is reported to have morphed into the phrase “blind and insane.”

Wow, not bad! Notice such changes as:

• Rumors have it was translated into It is rumored that.
• The spirit was willing was translated into The spirit was ready.
• Also note that the translation inserted a comma after the word Likewise.

I am impressed! Remember, the computer has no understanding of the meaning of what it was translating.

Here is a personal example. After finishing the writing of my 2018 book, The Fourth R (Second Edition), I used Google Translate to experiment with translating a short summary of the book. I was amused by the results of translating the following from English into Russian and then back into English:

Like Reading, Writing, and Arithmetic (Reading, ‘Riting, and ‘Rithmetic), the 4th R of Reasoning/Computational Thinking is both a discipline of study in its own right as well as being an aid to representing and solving problems throughout the curriculum and at all grade levels.

In the 2018 Google Translate program, the computer decided that the R in the 4th R above was an abbreviation for Ruble, a unit of money in Russian. So, this translation back into English demonstrated that the translation program had no understanding of what I was trying to say.

I used the same paragraph to repeat my language translation experiment on December 12, 2020, and achieved the translation:

Like reading, writing, and arithmetic (reading, drawing and rhyming), Level 4 Reasoning / Computational Thinking is both a discipline in itself and a means of representing and solving problems within the curriculum. and at all levels of learning.

This was an improvement over the 2018 result. However, the computer translates writing and arithmetic as drawing and rhyming. This clearly shows the difficulty of the computer not understanding that the 3 R’s is a phrase commonly used in talking about reading, writing and arithmetic in the elementary school curriculum in the United States.

It is very important to understand that, although Google Translate and other language translation programs are continuing to be improved, they definitely have not achieved the capabilities of good human translators.

In essence, this problem of a lack of understanding runs through all current applications of AI. While AI is indeed a very powerful aid to solving a wide range of problems, it currently is achieving this with no underlying understanding of the problem being solved, nor of the task being accomplished. Humans far exceed computers in this particular aspect of intelligence. Humans doing simultaneous translations by speakers of two different languages are able to receive and understand the input, and then instantly translate the meaning of the input into the second language.

As we explore both the current and possible future uses of AI in education, we need to be aware that the overall field of AI has made substantial progress, and currently still is making rapid progress. This is both because of more powerful computers becoming available, and also because throughout the world a large number of very smart and dedicated researchers are working to improve the capabilities of AI.

AI is a broad and very complex field. A very simple definition is that it is the study of the uses of computers to solve problems and accomplish tasks which—if this were being solved by humans—would require human intelligence, thinking, and judgement.

Here is one part of a three-part description of some current problems being addressed, quoted from AI and the Future of Learning: Expert Panel Report mentioned earlier. One way to think about AI is that it includes a set of specific capabilities that are advancing rapidly today. These include:

• Perception, via multiple sensors and ability to recognize complex sets of features (e.g., use of cameras and motion detectors to recognize particular faces entering a building).
• Representation and Reasoning, building models of people and their behaviors and making inferences based on those models about what might happen next.
• Learning, discovering meaningful patterns in large amounts of data.
• Natural interaction (e.g., interacting through speech or gesture).
• Societal impact, leveraging infrastructures to do all the above at a massive scale and in ways that directly affect people’s lives (Holland, 2020, link).

Each of these statements identifies an emerging capability of AI, and each is an active area of research in this field. Spend a little time comparing the complexity of each of these five areas of study and research with the specific task of language translation of written and/or spoken language. You will see that each is important to the use of AI in education, and each is a major challenge to current and emerging AI capabilities. I find it rather amazing that AI-based systems can do as well as they currently do in language translation and other tasks in which humans need to use their human understanding of what they are doing.

TSuppose that I need to find the product of two 4-digit integers. This problem has exactly one correct answer. I might calculate this answer using pencil and paper, a calculator, or a computer. Or, maybe I don’t really need an exact answer. I could make a mental estimate of an answer, and it might be good enough to meet my needs. In my attempts to obtain an exact answer, I might make a mistake in entering the data. And, although it is highly unlikely, the calculator or computer might make an error. In both cases, a mental estimate might help me to detect the error in these results.

This is a very important idea, and I will discuss it more in a later newsletter of this series. Think about the many decisions you make as you proceed through a day. How many of these decisions are correct because they are based on exactness in the information and thinking used in the decision? My point is, there are some situations when a high level of accuracy or exactness is desirable or necessary, but most of life is not that way. Just because it is a computer that is making the decision, we should not automatically think that every computer-made decision represents perfection.

For example, suppose that I am a physician making a medical diagnosis and prescribing a treatment. If you are my patient, you might be satisfied to hear me say that I have considerable knowledge about the ailment and how to treat it. I might be able to tell you that I had encountered this problem before, that a variety of treatments have been tried, and they have had varying levels of success. Then you and I would carry on a conversation about the next possible steps that might be taken.

Finally, consider the same scenario, but in this case the physician is an AI-based system. It tells you that, while examining and talking to you, it has simultaneously retrieved and read 1,000 recent research papers on this ailment, and its recommended treatment is based on this research. However, it notes that success levels have varied considerably, as have the side effects.

Hmm. I’ll bet you would like to carry on a conversation with your AI-based physician before deciding on your next step. Unfortunately, today’s AI systems are not up to this conversational task. This illustrates an absolutely fundamental idea about AI uses in education. The AI system of the future will be making decisions that affect students, and likely neither the students nor their teachers will really know the basis for these decisions or the likelihood of possible outcomes.

You undoubtedly are familiar with the fact that, many years ago, a computer program defeated the world’s reigning chess champion. More recently, another computer program became far better at playing the game of Go than any human players. There is no claim that each move the computer makes in playing chess or Go is a perfect move. However, in total, the computer’s moves are good enough to defeat its human opponent. Moreover, such an AI-based computer program can learn to play still better by playing games against itself.

However, the computer is unable to provide humans with a good explanation for why it makes each move it makes. The computer can defeat expert human opponents in many other different games. In all of these games, there is the concept of winning or losing. But, what does winning or losing mean in the education of a child? A child is not an opponent, and I certainly do not know what it means for a computer to win in the game of educating a child. With the help of AI, we can develop computer programs that are quite good at some aspects of teaching. If we decide that our goal is for large groups of students to do well on certain types of tests, then we can develop computer programs that are better than the humans who are teaching classes of 20 to 30 or more students. But, education is far more than obtaining high scores on tests.

It seems to me that researchers and implementors of AI as aids to teaching and learning face a daunting task. Users need assurance that, for each individual student, the decisions made by the computerized teaching machines will meet standards that are clear, understandable, and readily available.

First, consider a very simple case. A company develops AI-based Computer-Assisted Learning (CAL) materials that are designed specifically to help students prepare for and do well on a quite specific test, such as a widely-used college entrance exam. The company employs one or more highly reputable educational research organizations to do research on the effectiveness of their company’s materials versus those of other companies, and also in comparison with other approaches students take in preparing for the exam. These research studies meet high standards for reliability and validity. The results of the studies provide solid evidence that the company’s materials are effective, both from a statistical point of view and from other points of view. The other points may include such things as the fact that the test score increases are of practical significance—large enough to make a meaningful difference to the students and to the people who make use of the test scores—and are student time-use efficient.

Next, consider larger aspects of education that are not measured by widely used standardized tests. This covers all of the materials currently used in the day-to-day education of students across a number of curriculum areas and in schools across the country. This presents a huge challenge in measuring the effectiveness of CAL, one that so far has not been addressed effectively in any of the curriculum areas that commonly are taught in schools in the United States. Countries that have a national curriculum have a distinct advantage in this regard, since their national standards are to a certain extent measurable ,and can be accessed uniformly across all of their school systems.

In a recent Newsletter, I proposed that the U.S. Federal Government pay for the development and regular updating of a large number of AI-based CAL precollege courses that then would be made available free, both in the U.S. and worldwide.

A key idea in this proposal is a requirement that there be ongoing research on the effectiveness and impacts of the courses. There also would be major ongoing research on ways to improve these courses, including ways that make effective use of continuing progress in AI and other aspects of ICT.

The issue is not whether computers can be used effectively to help teach students. Rather, the question is about the quality of teaching, and the overall nature of learning that occurs in computer-student interaction versus human teacher-student interaction, interactions between and among students in the classroom, student-parent interaction, and so on. It appears obvious to me that, both now and for quite some time to come, education can be improved by an appropriate balance among these types of interactions. It would be a major mistake to greatly decrease the human elements that are of major importance in education today.

Information and Communication Technology (ICT) has made amazing progress during my lifetime. The pace of this progress has increased over the years, and continues to increase ever more rapidly. Our schooling systems certainly have made substantial progress since the first schools were developed nearly 5,500 years ago. They now are attempting to determine the changes needed in curriculum content, teaching processes, and assessment in order to make more effective use of steadily improving ICT. At the same time, schools need to help prepare students for adult life in this changing world. These are daunting tasks!

The next IAE Newsletter will take a closer look at the goals of education. I will address many of the ongoing changes in our world, together with a discussion of the ways that progress in AI and other aspects of ICT can and should be affecting our current schooling goals.

IBM (12/19/2020). A curated exploration of the computing landscape. technologyreview.com. Retrieved 12/20/2020 from https://event.technologyreview.com/future-compute-2021/?utm_source=event_email&utm_medium=email&utm_campaign=future_compute_
2021.unpaid.acquisition&discount=EMAIL121950&mc_cid=7aeceb1270&mc_eid=
4c700e292b#register.

Mikkelson, D. (n.d.). Computer mistranslation. Snopes. Retrieved 12/14/2020 from https://www.snopes.com/fact-check/quality-is-the-first-occupation/.

Roschelle, J., Lester, J., & Fusco, J. (eds.) (2020). AI and the future of learning: Expert panel report. Digital Promise. Retrieved 12/13/2020 from https://circls.org/reports/ai-report.