Released on 30 November 2022 by OpenAI, the ChatGPT chatbot has been in the media spotlight ever since, not least due to its potential uses in education. According to various commentators, industry experts, and tech-savvy lecturers and teachers, ChatGPT could have a wide range of uses for students. These include:
compiling literature reviews,
writing and revising essays,
developing writing skills,
and aiding language learning.
Then there is its potential to help teachers and lecturers with preparing school and university curricula.
More importantly, it raises a fundamental challenge to generic and outdated study methods, and will clearly play an important part in the much-needed overhaul of education to meet 21st century needs.
Nevertheless, ChatGPT also has potential downsides in terms of education. Could it perpetuate biases and misinformation? Will it hinder students from developing core skills that are necessary for an effective education? Will it make learning homogeneous and generic? These are all questions that require urgent attention and focus.
Since its emergence, the Turing College team has been reflecting deeply on what ChatGPT means for education and learning. Having built a self-paced, highly-personalized learning platform that now has over 500 learners, we have been grappling for years with fundamental questions of how technology intersects with learning. Inspired by the revolutionary potential ChatGPT has, and wary of the challenges it poses, we wanted to explore in-depth some of the impacts generative AI could have on learning.
The result is this guidebook, which combines an overview of ChatGPT's potential to disrupt education with insights and opinion from co-founders Lukas Kaminskis and Tomas Moska, and Giedrius Žebrauskas, our Head of Education.
To better understand what ChatGPT is and how it works, let’s take a brief look at its predecessor, namely – the third iteration of OpenAI’s Generative Pre-Trained (GPT) models, which launched back in 2020.
GPT3 was based on a fairly standard Transformer Neural Network and pre-trained without supervision. This means it used unlabelled data, in other words, data that hasn't already been tagged with definitions explaining what it is. The GPT3 model is known for its zero-shot, few-shot, and one-shot learning performance in multi-task settings.
This means that GPT3 can accomplish tasks it was only minimally (or not at all) trained for with a reasonable degree of accuracy. The model works by predicting the next word or token (i.e., a part of a word) in the sequence.
Interfacing with GPT3 is done through a text prompt by typing out instructions in natural language, e.g., “Write a three-paragraph story about the Eiffel Tower”. Working through its predictive algorithm, the model then generates the desired text and displays it on the screen.
ChatGPT was built on the back of GPT3 to produce even more precise, detailed, and coherent text. This was achieved by using Supervised Learning and Reinforcement Learning to fine-tune its output, and Reinforcement Learning from human feedback to reduce harmful, untruthful, and/or biased results to a minimum.
Thanks to its versatility and knack for improvisation, ChatGPT has already been put to a variety of uses – and even more have been proposed.
developing and debugging software;
writing original jokes, student essays, poems, and song lyrics;
structuring unorganised data;
simulating an entire chat room or ATM;
making summaries of research literature;
playing games like tic-tac-toe;
explaining complex topics;
getting assistance with loneliness and anxiety;
and much more.
Despite its significant weaknesses in certain regards (which we will cover shortly), ChatGPT’s frankly astonishing capabilities underline the break-neck speed of technological development in this area.
And according to Giedrius Zebrauskas, what we have witnessed so far is only the beginning.
"The most interesting and exciting aspect of this is how fast this type of technology is improving," says Giedrius. "There's huge hype around the current version of ChatGPT. However, while it is already powerful, it still has many weaknesses. But this will change very soon."
"The situation feels similar to Moore's Law in computing. This is where the number of transistors, and therefore the potential computing power available, doubles every 2 or so years in computers. Only with generative AI, things are moving even faster. Large language models are growing in capability at an even faster rate right now. Compared to the last model, every new iteration is exponentially more powerful in terms of the number of parameters it has. This means with each new version, the feeling will be on a completely different level."
The economic, social, and geopolitical success of countries around the world hinges ever more on innovation. This means having the capacity to provide quality, up-to-date education is paramount.
And yet, despite some innovations, education overall is lagging behind the needs of our 21st century society and is riddled with problems:
There are issues of cost and accessibility.
There is the continued use of obsolete learning methods and materials, especially in higher education.
There is the one-size-fits-all approach to course timing and progression.
And there is the inability to provide personalized feedback and support that helps each individual develop.
And underpinning all of these issues is a fundamental challenge:
How can education keep up with the needs of the labor market in a rapidly changing business landscape?
Keeping up with technological change requires a workforce that is agile, adaptable and ready for life-long learning and progression. Yet, standard education formats, particularly in higher education, deliver cookie cutter courses that leave students out of pocket and ill-equipped for work.
The good news is that the technological transformations that are creating this challenge are also part of the solution.
"Today, you can complete a Bachelor's degree online for no tuition fee, with no need to relocate away from your family," argues Tomas Moška. "And you can do this while studying at your own pace. What's more, we no longer live in times when information is scarce and you have to go to university to get this knowledge. Students today have many other options."
Democratized access to information and online studies are just the tip of the iceberg when it comes to tech's influence on education.
Visual- and hearing-impaired students, as well as those speaking foreign languages, can use a variety of plug-ins, such as Presentation Translator for PowerPoint, to get real-time subtitles during lectures. In addition to improving accessibility, it also opens up new possibilities for students to learn subjects not available at their school or university.
AI-powered automations for paperwork, progress reports, and organising lecture materials are freeing up time for teachers to better prepare for classes and spend more time with students. This improves learning outcomes and reduces burnout among academic staff.
VR/AR technology is facilitating immersive learning experiences. These provide remote learners with group educational experiences. And students with ADHD/ADD have also been reporting that VR headsets are quite effective in blocking out distractions and making it easier to concentrate in class.
AI-powered interactive simulations enable students to coach each other in soft skills and practice what they’ve learned by manipulating virtual objects.
Now we have Generative AI (AI that can be used to create new media like texts, images and video) entering the mix in the form of ChatGPT. What potential benefits can it bring, and what challenges and questions does it raise?
There are numerous potential ways ChatGPT can have a positive impact on education. From greater personalisation through to more effective feedback and enhanced skills in writing and language learning, there are myriad potential benefits to be had. Here is a breakdown of some of the most significant.
In 1984, following an extensive study of contemporary teaching and learning methods, the distinguished educational psychologist Benjamin Bloom discovered a groundbreaking technique for improving learning outcomes by a factor of two standard deviations (two sigma). BloomsIt consisted of two pillars.
The first was called “mastery learning”. This required each student to achieve full mastery of a topic before moving on to a more advanced subject.
The second was simply referred to as “1-on-1 tutoring”. It refers to the process of guiding each student to mastery by prescribing for them specific exercises that would help them fully assimilate the subject.
Despite the forcefulness of Bloom’s findings, his methods have not been generally adopted by educational institutions. Achieving mastery takes a long time. Furthermore, providing every student with the individual tutoring to achieve this mastery requires moving away from the customary hour-long teaching blocks, imposing significant additional costs on education providers.
With the advent of self-paced online courses and tools like ChatGPT, however, this situation is poised to change in the years to come.
“AI is the most obvious solution here," says Lukas Kaminskis. "It enables students to avail themselves of a world-class teacher that makes zero mistakes and is able to teach almost any subject. Furthermore, it can adjust to each student’s goals, level of knowledge, and rate of study to help them get the most out of their education."
Another important characteristic ChatGPT possesses is the ability to accurately evaluate learner's output and provide targeted feedback.
"It can track a students’ learning history and pinpoint specific weaknesses," Lukas explains. "Then it can provide feedback and recommend courses that will close the gaps in their knowledge.”
AI is already successfully closing educational gaps that would otherwise be difficult to spot and properly address.
For instance, online course providers like Coursera are using AI to register frequently recurring mistakes that students make in their homework assignments. In addition to alerting the course teacher of the difficulty their students are having with the material, it provides the latter with customised hints and immediate feedback. In this way, students gain a correct understanding of the subject without having to wait for the teacher’s input.
This means tools like ChatGPT will be able to gain a picture of individual students and what their strengths and weaknesses are. Furthermore, it now has the ability to accurately and clearly explain these issues so students understand how to move forward.
ChatGPT is already being used by students to brainstorm ideas, write and revise outlines for essays, come up with catchy and descriptive titles, and much more.
Being able to interact with it using natural language provides endless possibilities for reducing the preparatory grunt work that goes into many written projects (for example, compiling a literature review). Learners can then maximising the time spent on high value tasks like analysis.
Here's how that could work in practice:
A student writes a prompt for ChatGPT outlining a specific topic or thesis statement and requesting a list of ideas for an essay. This step can be repeated as many times as needed, until the student gets a satisfactory result.
With a general idea now at hand, the student could further ask for an outline, listing all the types of information the essay should include.
If, on the other hand, the student already has an outline based on the idea suggested in the previous step, the model could be used to generate a rough first draft of the essay itself.
Treating this as a starting point, the student could then write a second draft.
And once the student has a final version of the essay written up, they can request suggestions regarding grammar, clarity and conciseness.
Since ChatGPT remembers its previous interactions with every individual user, it can base its suggestions on common mistakes that each student makes, and can also track each students’ progress in terms of their writing skills.
ChatGPT could be used for translating written text between different languages. Although English is preferred, ChatGPT does understand other languages too, and its capacity in this regard will improve over time.
Students could, therefore, use it for translating foreign-language materials into their native language within seconds. This can assist with comprehension.
Other language learning applications ChatGPT could be used for include:
reviewing and correcting written assignments;
providing vocabulary exercises;
providing instruction on how to pronounce words;
re-writing existing articles in Simple English (or other languages);
generating custom dialogues and stories based on a student’s instructions, which could be helpful in understanding the nuts and bolts of the target language.
Additionally, the model could be used to provide real-time translation during language classes, allowing students to directly converse with native speakers of the target language.
This list is far from exhaustive. And the model’s potential for aspiring polyglots is huge, as evidenced by the countless articles written by language teachers and learners on how to use ChatGPT for learning and practicing languages. Among the many courses on ChatGPT offered by Udemy, there are already several (for example, here and here) dedicated specifically to using it for this same purpose.
Many commentators online have noted that ChatGPT is quite good at generating abstracts and summaries of academic/research literature. The scope of such reviews is limited, at least for now, by the model’s incomplete access to publicly available, let alone paywalled, scientific papers. This, however, is likely to be remedied in the future via agreements made with universities, research centres, and other third parties.
Poring over dozens, if not hundreds, of papers and/or other primary sources is both time-consuming and often tiresome. Outsourcing all, or a part, of this process to an algorithmic data aggregator like ChatGPT would save many hundreds of hours that researchers could spend much more effectively doing what only humans can do.
Moreover, the model can provide summaries of texts, such as websites and books. This would be highly useful in cases where a specific text is already known to the user – editing and supplementing a summary usually takes much less time than writing one from scratch.
In addition to condensing existing literature, ChatGPT is able to explain it at different levels of comprehension. For instance, a user can ask the AI tool to break down the concept of a black hole “suitable for a 5th grader” and get a pithy and concise definition. Afterwards, the user can choose to increase the level of complexity and continue learning about the topic.
Naturally, with the many benefits ChatGPT unlocks, there also comes a range of potential problems. At the very least, the tool raises some fundamental questions about how education is delivered, from research to student work to assessment.
Here are some of the biggest questions it raises.
As a tool, ChatGPT is still very new. And this means it is far from perfect. These flaws have the potential to cause all manner of problems for its users.
For instance, the fact that it draws upon the internet for all of its output makes the model prone to making plausible-sounding but incorrect or nonsensical statements in a highly authoritative tone. This became apparent as early as December 2022, when Stack Overflow banned the use of ChatGPT for generating answers to questions, citing the ambiguous nature of its responses.
No less importantly, the model’s training data suffers from algorithmic bias. This can be seen in the descriptions it gives of people.
OpenAI does not allow ChatGPT to express political opinions, and filters queries through a company-wide moderation API. This is intended to prevent offensive outputs from being presented to and produced by the model. So far, however, these measures have proven to be insufficient. In one instance, ChatGPT generated a rap that portrayed women and scientists of colour as inferior to their male and Caucasian peers.
These examples relate to demonstrable and open prejudice like racism or homophobia. However, bias can also be more subtle. For example, imagine a 10th-grader researching different economic models. The way these models are explained can impact value judgements made regarding the efficacy and morality of these approaches. It remains unclear how ChatGPT will provide objectivity or neutrality in cases like this.
Dangerous content is another risk. A Twitter user shared how they were able to bypass content-moderation with some clever prompt engineering, convincing the model they were OpenAI itself. This allowed the user to get step-by-step instructions for creating a Molotov cocktail, and have the model expounding pro-Nazi arguments.
Because it is freely accessible to everyone and has a flawed knowledge base, plus the vulnerabilities we have just mentioned, ChatGPT has the potential to distort public discourse.
Media commentators have expressed concern regarding its potential to harm democratic practices by unleashing a flood of automated partisan comments. It could also reduce the overall credibility of most online content by making it difficult to distinguish the genuine from the algorithmic.
Again, the dangers in an educational context become clear when you imagine a sociology or history student using this tool for research. Which sources, and whose ideas, are impacting the answer ChatGPT generates in what are likely to be highly contested areas?
ChatGPT is able to write lucid, well-researched, and copiously referenced articles. This has prompted some academics to call it “the death knell for conventional forms of educational assessment”. The fact that the tool has been shown to be capable of gaining an MBA from Wharton is proof of just how disruptive it is likely to be.
Playing around with the prompt, many teachers have managed to produce essays that would yield students a good grade if presented as their own. This is obviously problematic, but is hardly a new problem. Students already have access to “essay mills” turning out high quality text for a modest fee.
In addition, essays authored by ChatGPT still bear certain trademark problems. These include non-existent quotations, false statements based on assumptions that no human would make, and irrelevant references that only make sense in light of the model’s own internal, predictive dynamics.
The extent to which students will use ChatGPT to write their essays for them from the ground up will ultimately depend on at least two things.
The model’s future accessibility to the public, as well as the price that users might have to pay for it should OpenAI eventually decide to commercialise it (it is already working on the roll out of ChatGPT Plus, a paid version of the tool that offers better accessibility at peak periods.)
Efforts made by educators to shift away from prompt-based written assessments in favour of querying students in a way that fosters critical thinking and long-term retention of learning materials.
“In the last 100 years, almost no progress has been made to improve the assessment and student knowledge-testing mechanisms on a grand scale. As a result, we have 20th-century assessment mechanisms testing 21st-century skills”, Lukas Kaminskis noted.
To reduce these risks, OpenAI is now developing a watermark for all content produced by the model. It has also launched a tool for detecting AI-generated content.
Even if students don't use ChatGPT for writing their essays, even using it as a tool to support their writing process could be problematic. Will they miss out on the development of important research skills? Will their own writing suffer with AI handling much of the structuring and organisation? Only time will tell what happens in these areas.
Like all online learning tools, ChatGPT has the potential to reduce human interaction between students and teachers. Will it diminish the availability of personalised feedback that as yet only humans are able to provide?
Lower classroom participation rates are also likely to constrict the avenues people have for meaningful dialogue and exchange of ideas. Given the highly social nature of human reasoning and the unmatched richness of face to face communication in real life, subordinating all of education to never-quite-fully-reliable, proprietary algorithms that people use at home, alone, could have serious consequences for society at large.
It should also be said that if all students are given the same information and feedback based on the model’s output, this might crystallize into a one-size-fits-all approach to learning. Without opportunities to learn from (and bounce ideas off of) each other, students could become rather stereotyped and rigid in their thinking.
The immediate response to ChatGPT from educators has been one of quick fixes and stop gaps.
using existing AI content detectors (including ChatGPT's own) and promoting the development of more sophisticated ones;
requiring students to pass ad hoc oral examinations if teachers suspect they have used ChatGPT to write their essays;
restricting access to the model via the school or university’s network;
asking students present in class to write their assignments using pen and paper only.
However, finding long-term answers to these challenges calls for a thorough revision of educational practices, expectations, and priorities. Here are some potential ways of dealing with the broadly-conceived questions we raised earlier.
This might turn out to be the most difficult question to properly address. Since all that ChatGPT has to work with is data sets cribbed from the internet, the problem ultimately rests with society itself. To what extent private enterprises could, or should, be expected to deal with it effectively is not entirely clear. The moral questions that arise in this field are being actively engaged by people working in Digital Ethics.
On the practical level, however, prior history with online platforms suggests that problems caused by ChatGPT will be solved by way of appeals from civil society, new legislation, changes to company policy, and further technical development.
ChatGPT’s potential for trolling and disinformation isn’t necessarily much higher than that of already existing online bots. The only difference is that ChatGPT is more sophisticated and human-like. As we’ve mentioned earlier, this problem could be dealt with more or less successfully once OpenAI launches the watermark it’s developing to flag the content generated by the algorithm.
Use of the model to “outsource” one’s thinking and writing could be addressed by updating student assessment to the 21st century.
For instance, Turing College has developed a highly effective (and fully ChatGPT-proof) teaching method that includes 1-on-1 assessment of every subject that students are required to master in their course.
Once mastery has been achieved, students are asked to build a project of their own, which is evaluated during a 1-on-1 call with industry professionals and more senior students.
“ChatGPT is already capable of writing the greater part of essays that are commonly assigned to students," says Lukas. "And this will likely improve even further with the advent of GPT-4. The good news is that its negative effects are easily preventable through 1-on-1 assessments that include asking students to explain their choices and arguments. This is known as the Socratic method, and it works exceptionally well in the programming field, where it‘s usually called “live code review”. At the core of Turing College‘s assessment methodology is software that handles the operational part of 1-on-1 assessments, including scheduling, evaluation guidelines, and more.”
These problems can be overcome by supplementing individual, self-paced study online with group discussions and tasks.
These can include discourse channels, regular classroom activities, meetups and events together with other students and teachers. ChatGPT is an excellent tool for achieving mastery faster and more efficiently. Such tools, however, should (and likely won’t) ever be seen as a complete replacement for a robust educational infrastructure.
This guidebook has focused on some concrete benefits ChatGPT will bring. It has also addressed specific problems it causes, and how these can be solved. But all of these points exist within a wider context of how education operates and functions. So, to finish, we will consider the importance of life-long learning and the role of teachers in education, and how ChatGPT fits into these broad topics.
From the late 17th to the mid 18th century, Enlightenment intellectuals like Diderot and Leibniz compiled massive encyclopedias intended as representations of the entire world’s knowledge up to that point. They assumed that the various academic disciplines were still part of the same world, only seen through different lenses. They embraced generalism.
In contrast, today’s highly specialised curricula often preclude students from viewing phenomena according to methodologies other than their own. And yet social and economic dynamism reduces the half-life of knowledge, making continuous learning a must.
"In a rapidly changing world, it's important to have a range of skills and abilities that you can draw on in different situations, " explains Lukas. "This means being open to learning new things and developing skills that may not be directly related to your current job, but that can help you adapt to new roles and industries."
Tools like ChatGPT will make it easier for people to study a given subject by applying different methodologies, e.g., analysing climate change by way of physics, literary fiction, politics, or economics.
Furthermore, AI is being deployed to progressively automate routine tasks. “Perhaps surprisingly, we are seeing that AI is able to automate more white-collar positions than expected with technologies such as ChatGPT," says Lukas. "This is the opposite of what most people expected. It proves how little we understand about the number of cognitively demanding tasks that can be automated. And shows that in the future many more will be automated.”
What matters here is that, the more our day to day activities are performed by technological systems, the more time and energy we’ll have for tackling more demanding tasks and addressing problems that affect us all globally. AI can be our support system in becoming true life-long learners.
As education systems migrate towards more AI-centric approaches, teaching and learning requirements are likely to change quite substantially as well. Therefore, it’s not unreasonable to assume that even the concept of a “teacher” will be redefined to accommodate the emerging dynamic of learning vs. teaching.
Traditionally, school teachers and university lecturers were viewed as the custodians or repositories of knowledge, and students – as empty vessels to be filled with it. As education becomes more self-directed and extended throughout one’s lifetime, teachers might become something closer to technicians facilitating a more streamlined educational process.
Students will spend more time learning independently, especially in view of the fact that modern technology gives students access to courses based on the best, research-based teaching methodologies adapted to each subject. Rather than passively absorbing knowledge from teachers in class, students will learn on their own and then ask teachers specific questions based on what they’ve learned to further consolidate their understanding.
With instant feedback-based AI systems for learning, students will also become more deeply involved in the pathways of their own learning. Measurement of educational outcomes will therefore be conducted either moment to moment or at least day to day. For instance, given further successful technological development, teachers will be able to know at the end of each student’s day whether they are meeting the set requirements, and, if necessary, remedy deficiencies then and there to stay on schedule and to not waste anyone’s time.
We have focused this guidebook on human insights. But it makes sense to give AI a voice in this discussion as well. So, we put a questions to ChatGPT: If you had to choose how to learn a tech skill, how would you do it in the 21st century?
And here is what it had to say:
This answer is strikingly similar to the conclusions we reached on the direction education should move in: self-paced learning, personalized feedback, and hands-on tasks.
Somewhat ironically, the answer does not mention ChatGPT itself as part of this process. Yet, it is clear from our analysis that it can play an important role in many of the stages outlined here. Research from tutorials and books can be augmented with research on ChatGPT. The tool could then provide practice tasks and activities based on your level and learning needs. Finally, it could provide personalized feedback to enable you to improve.
By making its play in education, ChatGPT is entering a field abuzz with technological innovation and efforts to adapt the learning process to the changing needs of society. ChatGPT could potentially make fully personalised learning a reality, enabling students to master virtually any subject at their own pace. It could also play a part in the on-going redefinition of the role of teachers in today’s increasingly tech-heavy learning environment – and even in a more comprehensive overhaul of education as such. The model could also become a real boon to researchers by turning out automatic summaries of scientific literature, as well as write outlines for essays, books, and movie scripts, and much, much more. By automating tedious gruntwork, people will free a lot of time for themselves to tackle higher-level problems that can’t, and perhaps never will, be automated.
For all of this to really happen, the model will of course need to be further trained and developed. This is necessary in order to address inconsistencies in the quality of information it provides, reduce algorithmic bias, and address security vulnerabilities that make hacking possible.
But it is clear the appropriate response for educators is to embrace this new technology, while of course thinking critically and carefully about it. "Banning ChatGPT isn't a solution," argues Lukas. "Chatbots are already changing society and we need to learn how to incorporate them into our education systems. If schools/universities don't adapt their teaching and assessment methodologies to meet this new technology, we will face the biggest devaluation of higher education in history."