ChatGPT is a large-language model – a powerful algorithm that can ‘chat’ with a user in eerily human fashion, drawing on word arrangements and patterns it has ‘learnt’ from the millions of pieces of text, including books and articles on the internet, it has been fed. It was built and released by the American company OpenAI.
It has already cleared an MBA examination set by Wharton Business School faculty, and did reasonably well in law examinations. It wrote a piece of law whose objective is to regulate the use of such chatbots. Students have also been using it to write essays and produce code while lobbyists have started using it to draft petitions.
Academic institutions are rightly concerned about its impact on academic activities that are currently used to evaluate and grade academic work, such as essays and computer programs.
The author asked ChatGPT to pen a critique of the ideas of Ivan Illich, the Austrian philosopher who wrote, among other things, the famous book ‘Deschooling Society’ (1971). This is a typical essay assignment to students. ChatGPT responded with a piece that would have bagged a B+ or an A-.
The primary problem for educationists is that ChatGPT can produce coherent, well-written text that is hard to distinguish from something a human would write. At the moment, there is some scope to develop tools to detect signs of robotic authorship, but it is likely to diminish as the bots become better. The rise of Google Search augured the rise of common plagiarism; today, the use of plagiarism detection software is mandatory in most academic settings.
ChatGPT could upend this state of affairs by combining the activities of searching for source material, collating and synthesising it, and finally producing human-like text into one straightforward activity. That is, all examinations that involve these steps can now be automated. It may still struggle with sourcing and attribution, especially when offering a claim that draws from multiple sources, but solving this could just as well be a matter of time, as the impending integration of such bots with search engines indicates.
This in turn would constrain the activities that teachers currently use to test and grade students’ work to in-person interactive sessions, such as oral examinations and in-class proctored exams. Another hack may be to ask students to write about what they have written in a more personal vein: what they learned, what they found difficult, etc. This may not always be possible, and as ChatGPT and its peers become more sophisticated, evaluation exercises are likely to become more complex, and perhaps less indicative of the material being examined.
Ironically, this particular challenge is an instance of machine intelligence (such as it is) driving us directly towards a situation in which more, instead of fewer, people will need to be involved to monitor and guide its use. But there are some deeper consequences with more fundamental implications.
First, the effect on reading and writing skills: The use of ChatGPT could render the ability to write well, or better, redundant, which could in turn affect reading and comprehension skills. The advent of internet search together with digitisation, online publishing and their attendant business models frayed the once more-common habit of visiting libraries and reading books. Search eased the process of locating an item of interest but weakened the notion that it must be read.
Yes, if today’s students are interested in a topic, they have more textual material available to read at their literal fingertips – but only a minority reads it. Most students (in the author’s experience) simply locate a paragraph of interest and skip the rest – a trend that has mostly been concurrent with decreasing attention spans and a general distractedness.
ChatGPT et al. will exacerbate these tendencies: a student could just say, “Write a critique of Ivan Illich and mail a PDF of the result to my professor from my email ID” – and voila!
Second, the outsourcing of thinking and analysis: Conscious processes of imbibing external inputs, filtering, interrogation, assimilating, etc. are how we develop new points of view, opinions, and critique. Reading is a thoughtful way to process recorded information because it allows for rumination and analytical filtering. In response to a search query, internet search engines output a list of links ordered by a ranking algorithm. The onus of inspecting the contents at each link is still on us; these platforms don’t produce finished articles.
On the other hand, ChatGPT and its successors could mark the shift from ‘why read’ to ‘why think’. For example, we could go from “Alexa, what are today’s headlines?” to “ChatGPT, tell me about the Russia-Ukraine war” – and the answer is likely to be an output ‘averaged’ over a mass of aggregated information. Perhaps this will be the deepest cut of all.
Third: A basic limitation of ChatGPT is that the ‘wisdom’ it produces in an article or conversation depends on the inputs on which it has been trained and the parameters of the training model. Imagine ChatGPT’s responses if it is trained on conspiracy theories and the claims of India’s infamous WhatsApp-based misinformation mill (a.k.a. ‘WhatsApp university’). Similarly, Google Bard – the search giant’s response to ChatGPT – could provide a different answer to the same query if its ‘learning’ algorithms are built differently.
A related issue is that extant search engines only output a list of links, without any guidance on how their contents can be interpreted. ChatGPT et al. take one more step, raising the problem of verifying the accuracy of correlative truths. Put simply, aggregate views, ‘alternate’ facts, and half-truths can subtly ‘leak’ into a chatbot’s output without its user being able to detect it. A bot-written essay can thus contain false nuances embedded in an overall largely reasonable write-up.
Universities and institutions have taken some of these issues in earnest. The New York City department of education, for example, banned the use of ChatGPT in public schools, as have some universities. Several scientific journals have disallowed ChatGPT from the authorship of papers. However, these measures are likely to be short-lived: once unleashed, technologies rarely allow room to live in exclusion of them. This is why many educators have also pushed the more optimistic, if also simplistic, view that we ought to “integrate” ChatGPT et al. into our education system.
Last year, Google engineer Blake Lemoine was so impressed by his interaction with a chatbot that he insisted it had become sentient. The truth was simpler, albeit still remarkable: what these bots profess to know is a sophisticated regurgitation of what they have already been fed. That this deceptively simple fact could threaten the education system as we know it is, however, more fascinating.
Anurag Mehra is currently the Science and Society Faculty Fellow at Vassar College, New York. He teaches at IIT Bombay.
- The primary problem for educationists is that ChatGPT can produce coherent, well-written text that is hard to distinguish from something a human would write.
- ChatGPT could upend this state of affairs by combining the activities of searching for source material, collating and synthesising it, and finally producing human-like text into one straightforward activity.
- This in turn would constrain the activities that teachers currently use to test and grade students’ work to in-person interactive sessions, such as oral examinations and in-class proctored exams.