No, it won’t. But what about what comes next?
It is hardly surprising that 64% of respondents in a poll think that chatbots, robots or AI can replace teachers in the future.
There is little doubt that this belief has been strengthened considerably with the launch of ChatGPT, a LLM (Large Language Model) that can respond to “prompts”, i.e., input in conversational style and carry on apparently meaningful dialogues that resemble interactions between students and teachers in classrooms.
ChatGPT is remarkably versatile. It can write book reports, translate from French to English, polish resumes, correct codes and write passable poems. It could even pass the law examinations at the University of Minnesota.
According to educationists quoted in a New York Post article ChatGPT can easily teach classes already and is likely to be most effective in middle or high school. But college lecturers and university professors may not be safe either. According to another recent article in the Guardian “Professors, programmers and journalists could all be out of a job in just a few years” because of ChatGPT.
So what would a world where AI has replaced teachers look like?
It was much harder to imagine such a world in 1951, but Isaac Asimov did just that in his wonderful short story “The Fun They Had”.
In the year 2157 (in Asimov’s story) all children are taught at home by “mechanical teachers” who would certainly be described as AI in today’s parlance. Books are of course no longer used. (The story opens with a child, Margie, discovering “a real book”.) Lessons are shown and questions asked on a “big screen”. Children have to write out their homework and test papers “in a punch code”. (Even Asimov’s imagination had limits. He couldn’t imagine computer keyboards in 1951.) Learning is completely personalized with the mechanical teachers programmed to teach each child at the pace best suited for him or her.
With ChatGPT has Asimov’s dream — or nightmare — become real ?
Different This Time?
So will ChatGPT replace teachers (now, or in the very near future)?
People have been talking for more than 100 years about technologies that would revolutionize education. These included motion picture, radio, television, videos and computers. All delivered considerably less than they promised. For an entertaining and informative account see this video.
Is ChatGPT another false alarm? Or is it different this time?
It has been known for sometime that ChatGPT is bad at math. How bad? It performed miserably in Singapore’s notoriously difficult PSLE (Primary School Leaving Examinations) for Grade 6 students, averaging 16% in math and 21% in science on the 2020, 2021 and 2022 papers. Apparently a reboot in mid February has improved the math but I don’t see MOE (Ministry of Education) replacing math or science teachers in Singapore by ChatGPT anytime soon!
What about grammar? Surely it would do well in that? ChatGPT after all writes perfectly grammatical paragraphs and essays.
So I asked ChatGPT:
Me: What is the difference between the ways in which adjectives are used in French, English, Bengali, Hindi and Bahasa Indonesia?
Erudite and impressive at first glance! And why shouldn’t it be? ChatGPT has after all probably imbibed the whole internet. But ChatGPT’s response is riddled with subtle mistakes.
Placement of adjectives usually comes after and not before the noun in Bahasa Indonesia. In fact ChatGPT itself provides an example with “mobil (car) merah (red) ”.
“Ek (one) lal (red) gari (car)” is certainly incorrect usage. Bengalis would either say “lal gari” or “ekta/ekti lal gari”.
It’s quite true that adjectives don’t change their form based on the gender of the noun in English, Bengali and Bahasa Indonesia. However that is not the case for Hindi. For example in Hindi we have “accha (good) larka (boy)” but “acchi (good) larki (girl)”
Of course it is already well known that ChatGPT often produces incorrect answers. It can, for example, “hallucinate” and produce real looking but completely fake biographies of both real and imaginary people as described in my last blog and also here. But subtle mistakes are even more dangerous since they are harder to catch. We can say with high confidence that ChatGPT can’t replace teachers as long as it remains such an unreliable store of information.
One should not overreact to the hype surrounding ChatGPT as some skeptics are doing. For example, the celebrated linguist and left-wing political analyst Noam Chomsky told an interviewer that he does not believe that ChatGPT “has anything to do with education except basically undermining it. ChatGPT is basically high-tech plagiarism.” I disagree. Even though ChatGPT can’t — yet — replace teachers it is already an effective teaching aide that can help teachers plan lessons, discover relevant references and design exercises for students. One needs however to handle it with care and treat it like a somewhat unreliable assistant and check, for example, that it hasn’t made up some of the references .
Teething Troubles?
Surely the current problems with ChatGPT are teething troubles that will be resolved soon?
A meme of uncertain origin recently went viral. It claims that the next version of ChatGPT — GPT-4- will have 100 trillion parameters against the 175 billion of the present version — GPT3. I will come back to parameters, but roughly speaking, the higher the number of parameters the higher the flexibility of the language model. So ChatGPT — 4 is expected to be much more powerful than the current version of ChatGPT.
The claim about 100 trillion parameters of ChatGPT-4 is apparently not correct. But it’s probably true that ChatGPT -4, which has just been released to a limited audience, is more powerful than the current version and the one after that will be still more so. Moreover a horde of LLMs to rival ChatGPT are already being developed by such tech giants as Google, Meta and Baidu.
So surely we will soon have a LLM that is big enough and smart enough to replace teachers? Perhaps by next year? Perhaps in a couple of years? Within this decade? Won’t we?
I am not so sure.
Let’s dig a bit deeper into how ChatGPT (and similar LLMs) work. If you want a comprehensive explanation, please read this article by Stephen Wolfram. I am going to try to paraphrase some key points from Wolfram’s paper in the next few paragraphs, omitting about 90% of the detail.
Let’s begin by discussing autocomplete — a feature of many search engines and messaging applications — that suggests the word or words that should follow a text input. To take an example from Wolfram’s article suppose I begin by typing “The best thing about AI is its ability to”. What reasonable continuation or continuations should autocomplete suggest?
Any autocomplete algorithm is trained on a corpus, i.e. a body of texts collected from sources such as books or the internet. The algorithm determines the most probable next words that could follow the input text. In the above example it determines that the probability of the next word being “learn” is 4.5%, of it being “predict” is 3.5% and so on. It then outputs the most probable word or set of words.
How does it do that? One can imagine that an autocomplete algorithm scans all phrases in its corpus that begin with “The best thing about AI is its ability to” and determines the probabilities of “learn”, “predict” or “make” being the next word by simply calculating the relative frequencies of their occurrence. In principle, this is indeed how autocomplete works. It’s trickier in practice and needs clever tweaking to return its suggestions to user in reasonable time.
One can further imagine autocomplete continuing to extend the original prompt one word at a time. For example after a number of steps we might get “The best thing about AI is its ability to learn and develop over time”. We don’t need to always choose the most probable word as the next in sequence. Instead, we can introduce a degree of randomness in the process of choosing the next word at each step. For some reason this seems to make the output less “flat” and more interesting.
Surprisingly ChatGPT, for all its versatility, works quite similarly to autocomplete and also builds its response one word at a time. It relies on an underlying neural network of the type represented in the diagram below.
In the diagram the nodes of the network represent neurons and the arcs the connections between them. Of course the ChatGPT network is far more complicated than this picture. It has several million nodes and 175 billion arcs each corresponding to a neural connection. Each arc has a weight. These 175 billion weights are the parameters I mentioned earlier.
Each word (or if you wish to be more technical, each token which could be either a word or part of a word) is vectorized or converted into an array of numbers before processing starts. The input therefore becomes a huge array whose elements are themselves arrays. A very complicated series of calculations involving the 175 billion weights, are applied to the input to convert it into an output array. The output array is in turn translated into text output.
Now where do these 175 billion weights come from? They are determined by a process of pre-training the neural net with terabytes of text data. (ChatGPT is an acronym for Chat Generative Pre-trained Transformer).
During the training process ChatGPT is fed with training examples. A training example is simply a piece of text with a bit at the end masked or hidden. ChatGPT’s input is simply the unmasked text. For every input the output generated by ChatGPT is compared with the masked bit to see how closely they match. ChatGPT is trained on millions or perhaps even billions of such examples and the parameters fine tuned to make the matches as good as possible.
The entire purpose of this mammoth training exercise is to teach ChatGPT to recognize patterns in text. So it is not too surprising that ChatGPT learns grammar well, even though as we have seen, it trips up once in a while. Grammar, after all, essentially consists of patterns that we find in language. Of course these patterns can be encapsulated in explicit grammatical rules. But babies learn most of their first grammar through examples rather than rules. Language learning apps like Duolingo also use examples rather than explicit rules to teach languages to adults .
On the other hand the teaching of mathematics always seems to involve a degree of structured instruction. It’s hard to learn mathematics from just examples. So, in hindsight, it’s not surprising that ChatGPT does worse than Singaporean 12-year-olds at math.
ChatGPT’s factual inaccuracies are even less surprising. After all its training is focused on matching text input well and not on achieving factual accuracy.
So LLMs’ weaknesses seem to be fundamental flaws. Features not bugs. And therefore I consider it quite unlikely that any LLM will replace teachers anytime soon.
The Final Question
I suspect that the future will belong to hybrid AI which combine features of LLMs with more traditional powerful software for structured computation. Wolfram has already proposed one such hybrid that would combine the flexibility of input of ChatGPT with the mathematical power of Wolfram|Alpha.
I have no idea how much time it would take to develop such hybrid AI. But with billions of dollars of funding poured in and hundreds of talented scientists working on it I expect that it will get done.
So we come to the final question. Will such hybrids or some other form of AI eventually replace teachers ?
It is very hard to guess what the far future holds. AI will evolve in currently unimaginable directions and undoubtedly impact education significantly.
However I believe that AI will never replace teachers for a reason which Asimov intuited.
This is how the story “The Fun They Had” ends.
“She (Margie) was thinking about the old schools they had when her grandfather’s grandfather was a little boy. All the kids from the whole neighborhood came, laughing and shouting in the schoolyard, sitting together in the schoolroom, going home together at the end of the day. They learned the same things, so they could help one another on the homework and talk about it.
And the teachers were people. …
The mechanical teacher was flashing on the screen: “When we add the fractions 1/2 and 1/4-”
Margie was thinking about how the kids must have loved it in the old days. She was thinking about the fun they had.”
Education is socialized learning for humans and has been so since the time early homo sapiens hunted mammoths. School is much more than a place where children are taught mathematics, science, history and geography. It is a miniature society where children learn to work and play together.
The teacher is the lineal descendant of the tribal elders who initiated the children into the art of living and working together. A good teacher doesn’t merely transmit information. She fires the students’ imagination and helps them to learn to learn. She guides them on the path to becoming good citizens.
And that is what I believe AI can’t replace. Not now. Not soon. Probably never!
Disclosure: I am the Director of Smart Consulting Solutions Pte Ltd, incorporated in Singapore and its subsidiary Radix Analytics Pvt Ltd, incorporated in India. I am also a Visiting Faculty Member at the Indian Institute of Management, Udaipur. The opinions expressed in this article are solely mine and not necessarily shared by any company or institution with which I am affiliated.
I benefited from some helpful discussion with Ali Minai (Professor of Electrical Engineering and Computing Systems at the University of Cincinnati) and Pramit Das (Graduate Student at the University of Michigan, Ann Arbor).