Here we are again!

Since ChatGPT was first launched towards the end of 2022, there has been much alarm expressed in schools and colleges, in discussion forums, blogs and other social media platforms, in the educational press and in the general press too. There has also been calmer discussion; we shall come to that.

ChatGPT is an artificial intelligence (AI) text-generator, developed by OpenAI.  Its appearance marks a huge step forward in the evolution of AI.  To now, text-based AI has been uninspiring and flawed: think of the chatbots used by many support centres which provide stock answers which never quite answer your specific problem, think of translation software and synonymisers and rewriting software which give word-for-word equivalents which are often nonsensical.

At first glance, ChatGPT is different.  ChatGPT responds to prompts to provide coherent and unique responses to often-complex questions; it is able to hold conversations, understand what is said and to make appropriate – if not always correct – responses.  Artificial it is and it may “know” a lot, but its intelligence and indeed what it “knows” may be limited.

AI is capable of basic writing tasks; the Washington Post for instance has been using AI to write reports for at least six years; the story The Washington Post leverages automated storytelling to cover high school football was published in the newspaper on 1 September 2017, and was supplemented a fortnight later by a Digiday article which declared The Washington Post’s robot reporter has published 850 articles in the past year.  Text-generating AI has been with us for some time, ChatGPT is a natural development.

Much of the alarm has been voiced by teachers who fear that students will use AI, and ChatGPT in particular, to cheat, to submit work generated by ChatGPT instead of doing the work themselves.   Because its output is fresh each time, its responses to the same prompt are never quite the same; what is more, it can be asked to regenerate its responses, and will churn out much the same information in different forms.  Here, for instance, are three responses to the question:

ChatGPT – would you pass the Turing Test?

The Turing Test is a test used to determine whether a machine is capable of exhibiting intelligent behavior indistinguishable from a human. As an artificial intelligence, I am designed to mimic human conversation and behavior as closely as possible, but ultimately the decision of whether or not I pass the Turing Test would be up to the person evaluating me.

As an artificial intelligence, I am not capable of taking the Turing Test because I do not have feelings or consciousness. The Turing Test is a measure of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. It is intended to determine whether a machine is capable of thinking and behaving like a human. I am just a program designed to process and generate text based on the input that I receive.

The Turing Test is a measure of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. While I have been designed to have a wide range of capabilities and to be able to carry out conversations with people in a natural and human-like manner, I am still just a machine and my responses are generated based on the input that I receive. Whether or not I would be able to pass the Turing Test would depend on the specific circumstances of the test and the abilities of the human evaluator.

Each of these responses is different.  Text-similarity software such as Turnitin will find no similarity to any other text.  Any student submitting work produced by ChatGPT will not be “caught” by a “plagiarism detector”, it will be difficult to prove conclusively that the work is not the student’s own.  The same is true, of course, for any plagiarism-free work produced by a friend, a tutor or a custom-essay writing service and submitted as the work of the student when it is not – and ChatGPT has the advantage of both exceedingly quick response and, at least for the time being, being free.

So it was that in the days immediately following ChartGPT’s debut. articles with alarmist headlines were quick to appear:

The College Essay Is Dead (Stepehn Marche, The Atlantic, 6 Decmber 2022)

ChatGPT Will End High- School English (Daniel Herman, The Atlantic, 9 December 2022)

Teachers Fear ChatGPT Will Make Cheating Easier Than Ever (Rashi Shrivastava, Forbes, 12 December 2022)

Schools and school districts are already blocking use of ChatGPT, a move which seems particularly futile, not least because the block can be anabled on education department devices; students can still use ChatGPT on their cell-phones and other devices both inside and outside school.

NYC bans access to ChatGPT on school computers, networks (Michael Elsen-Rooney, Chalkbeat (New York), 3 January 2023).

It is almost as if there is an assumption that students are ever looking for unethical shortcuts and ways to cheat; I believe that this is questionable.

Some students will never cheat. They have personal integrity, encouraged and supported by home background and a school ethos and culture which promotes and celebrates honesty and authentic work.

I accept that there may be some students who, for whatever reason, will cheat any time they can.

And there is, I think, a group in the middle who might be tempted to take unethical short-cuts, especially if uninterested in the assignment set or whose poor time-management has led to a race against the clock to complete on time, and if they think they stand a decent chance of getting away with it.  It is this middle group who need the most protection, both from a perception (justified or not) that “everyone else is cheating so why shouldn’t they?” and also from predators who provide shortcuts and temptations of various kinds, often dressing them up as acceptable practices when they are not.

It must also be said that content filters rarely work as intended, nor is ChatGPT the only text-generating AI available, just the best of its kind at the moment.  Better, I think, to teach responsible use of ChatGPT, awareness of its shortcomings (and there are many), and how to use it as a tool to enhance academic work and learning.

Not far removed from the alarmist stance are those who would fight technology with technology, in a never-ending game of catch-up.

Startup says it can reliably detect AI-generated content (Paul Gillin, SiliconANGLE, 16 December 2022),

Can Anti-Plagiarism Tools Detect When AI Chatbots Write Student Essays? (Daniel Mollenkamp, EdSurge, 22 December 2022).

A Princeton student built an app which can detect if ChatGPT wrote an essay to combat AI-based plagiarism (Pete Syme, Business Insider, 4 January 2023).

Although test results suggest that these apps have high degree of success in distinguishing genuine work from AI-produced text, there may be issues in “proving” that an assignment was completed by AI rather than by the student themselves – which could well be the case if a student insists that the work is their own.  With text-matching software, if the instructor can produce text written earlier which matches word-for-word text produced by a student, there may be suspicions of copy-paste plagiarism, and the more text there is which matches, the more certain the probability of plagiarism becomes. With original text generated by AI, there is nothing to match against.  Suspicions can not be proven.

And there is a growing body of opinion which holds that we cannot hold back the tide, education and educators must adapt.

New AI tools that can write student essays require educators to rethink teaching and assessment (Mike Sharples, Impact of Social Sciences, 17 May 2022).

What does writing AI mean for education? (Hanna Seariac, Deseret News, 11 December 2022).
(Later retitled: Does AI mean the death of the college essay?)

What Would Plato Say About ChatGPT? (Zeynep Tufekci, New York Times (Opinion), 15 December 2022) 

Teachers v ChatGPT: Schools face new challenge in fight against plagiarism (Osmond Chia, The Straits Times, 3 January 2023)

The notion here is that AI will develop ever further, but we can use AI, including ChatGPT, as learning tools to enhance our teaching and students’ learning. Our pedagogy, the ways we teach and the methods we use, may need to change, but this has always been the way, we have always had to adapt our teaching, both content and methods, as new technology has become available. Tufekci’s article is particularly apposite here, reminding us that Plato feared that the written word would affect our abilities to memorise.  What goes around, comes around.  Much more recently than Plato, Clive Thompson admitted

I’ve almost given up making an effort to remember anything because I can instantly retrieve the information online” (“Your Outboard Brain Knows All,” Wired, October 2007) while around the same time, David Brooks declared “I had thought the magic of the information age was that it allowed us to know more, but then I realized the magic of the information age is that it allows us to know less (“The Outsourced Brain,”  New York Times October 26, 2007).

Change is always a threat, a challenge.  We have learned to with writing, books, paper, typewriters and computers, learned as educators to use to our own and our students’ benefit Google and Wikipedia, Facebook and Twitter.

A line in the NYC bans access to ChatGPT article is worth commenting on:

The chatbot’s ability to churn out pitch perfect essay responses to prompts spanning a wide range of subjects has sparked fears among some schools and educators that their writing assignments could soon become obsolete.

Yes, those writing assignments as they stand may become obsolete – or they can evolve and be adapted, make them rely less on stock responses but instead require thinking, critical and personal thinking, responses which cannot be artificially generated.  Several decades ago, the mantra was “If you can find the answer on Google, then you have not asked a (research) question”.  The need for inquiry approaches is more important than ever.

There is also the questionable thought that AI can “churn out pitch perfect essay responses”.  There is more and more realisation that ChatGPT’s is bland, repetitive, formulaic. The spelling and grammar are generally correct but the output sounds robotic and is often wrong. ChatGPT’s abilities are shallow, limited, far from pitch-perfect when asked to “think” more deeply.  This AI can work only with the information with which it has been provided, and while it has been provided with a lot of information, it cannot (as yet) add to it, and perhaps not even to synthesise it.

I asked ChatGPT to comment on recent events, it responded that it could not comment as these had happened after its “knowledge cutoff”, so naturally I asked about its knowledge cutoff and got the response:

My knowledge cutoff is the date after which I was not trained on any new information. My knowledge cutoff is 2021, so I do not have information about events or developments that have occurred after that time. I am only able to provide information and answer questions based on what was known at the time of my training. It is important to note that I am not able to browse the internet or access any new information, and I can only provide information and responses based on my programming and the information that I was trained on.

Particularly telling may be that ChatGPT does not always “know” from where its information comes, and will “happily” invent references to support its statements, complete with URLs to non-existent pages.

As instance, I asked ChatGPT several questions about Hamilton Naki’s role in pioneer heart transplant surgery.  The responses showed no awareness of any controversy about his contributions, and when I asked for references to support ChatGPT’s responses, the AI gave me:

None of these pages exist, they all produce error messages.  Click on the links, search for the article titles, check these authors’ writings, trawl the archives of the sites (and the Internet Archive too); the facts might (or might not) be complete and accurate but the sources claimed are not. The day may well dawn when a truly intelligent, knowledgeable, creative and critically thinking computer or artificial intelligence is developed (or develops itself). For the moment, it seems that the tools of information literacy are beyond AI’s ken.

This weakness seems worth adding to our toolboxes. We already ask students to state the source of any information which they obtain elsewhere and therefore needs a reference.  Especially in the early stages of a piece of writing, it can be helpful  quickly to check several of these references to be sure that they have given us the right source and the right information from that source.  This helps us be more certain that the student knows how to cite and reference sources and is writing ethically. If some of those references turn out to be fictitious, it could indicate that AI might be behind much else of the work submitted.

There is an uneasy sense of déjà vu here, we have been here before. Checking the accuracy of citations, both for content and source, is a tool in our plagiarism toolbox.  Another tactic is to ask students to provide screenshots, printouts, photocopies or other evidence of their searches and finds, perhaps routinely as part of the assignment.  As with any work which we suspect may not be that of the student, careful questionning regarding the research process, perhaps discussion of sources not used as well as those used and reasons for those decisions, and a check that the student understands what has been submitted, words, concepts and ideas, can help us decide on the balance of probabilities.

But, as with plagiarism detection, better still surely is discussion as to the shortcomings of AI, what it can and cannot do and how to use it ethically. When Wikipedia first appeared, the knee-jerk of many teachers and schools was to ban it completely, forbid its use for school assessments. Students were forbidden to use Wikipedia even while many of the teachers forbidding them to use it happily made use of it themselves.  Gradually a more nuanced approach emerged, the notion of using Wikipedia for basic information, teaching the use of the references cited as a springboard to the finding and use of more authoritative sources, teaching how to use the padlock and star features to find locked and moderated (and therefore more accurate if not quite authoritative) articles, and so on.

This may be the way to go with ChatGPT in education: use it as a springboard – with the added twist that students themselves still need to find and cite the sources which provide the information it gives and which they use.  They need also to be sure that any AI output that they use is accurate, and that it addresses the question asked.  Teach them how to make best use of AI – and when to avoid it.  With regard to ChatGPT, teach them that they need to check responses, add to them, find the gaps and fill them, use their own voice.  Use ChatGPT as a tool, not as an enemy…

A footnote on bland, formulaic output

Early in this article, I presented three different responses, one original and the other two “regenerated”,  which ChatGPT gave to the question “ChatGPT – would you pass the Turing Test?”  Each response is unique, but I did wonder if the AI might give over-similar responses to someone else – so I set up an account on a different computer and using a different email and asked the same question.  The three original responses are listed first, the second set of three follow.

ChatGPT – would you pass the Turing Test?

Computer 1

The Turing Test is a test used to determine whether a machine is capable of exhibiting intelligent behavior indistinguishable from a human. As an artificial intelligence, I am designed to mimic human conversation and behavior as closely as possible, but ultimately the decision of whether or not I pass the Turing Test would be up to the person evaluating me.

As an artificial intelligence, I am not capable of taking the Turing Test because I do not have feelings or consciousness. The Turing Test is a measure of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. It is intended to determine whether a machine is capable of thinking and behaving like a human. I am just a program designed to process and generate text based on the input that I receive.

The Turing Test is a measure of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. While I have been designed to have a wide range of capabilities and to be able to carry out conversations with people in a natural and human-like manner, I am still just a machine and my responses are generated based on the input that I receive. Whether or not I would be able to pass the Turing Test would depend on the specific circumstances of the test and the abilities of the human evaluator.

Computer 2

As an AI language model, I am designed to generate human-like text and I can hold conversations on a wide range of topics, but the Turing Test is a measure of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. The test is named after Alan Turing, who proposed it in 1950 as a way to determine whether a machine could think. To pass the Turing Test, a machine would need to be able to engage in a conversation with a human evaluator and convince the evaluator that it is a human, rather than a machine. While I am capable of generating human-like text, I am not a conscious being and do not have the ability to think or exhibit intelligent behavior in the same way that a human does. Therefore, I would not be able to pass the Turing Test.

As an AI language model, I was not designed to specifically pass the Turing Test. The Turing Test is a measure of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. While I have been designed to generate human-like text, my primary purpose is to assist users by providing accurate and reliable information, rather than to deceive or mislead them into thinking that I am a human.

As an AI language model, I was not designed to pass the Turing Test. The Turing Test is a measure of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. While I have been designed to generate human-like text, my ultimate goal is not to deceive people into thinking I am a human. Instead, my goal is to assist users by providing helpful and accurate information to the best of my ability.

It seems there may be a limited number of ways to say the same thing, at least in this short example, at least when asked to regenerate a response. Although each response is almost (?) unique (?), similarities are evident, the same phrases repeat, not necessarily in the same order, within the paragraph. This could alert teachers when the class is set the same assignment.

That said, the two sets of responses are different; the first set declares that whether or not ChatGPT would pass the Turing Test depends on the evaluation of the human interacting with the AI, the second set suggests ChatGPT has not been designed to pass the Turing Test so this is a non-question.

The blandness of these responses, of much ChatGPT output, should raise red flags, suggest further investigation is necessary. We might not be able conclusively to “prove” that a student has used AI to respond to a prompt, but we can question closely to make sure that the student at least understands what has been presented, that some form of learning and understanding has taken place.  The danger is that some teachers accept ChatGPT output without question while others question it most closely, and may even require students to re-write their responses, in their own words and in real time, in front of the teacher.

This is another déjà vu scenario, a hark-back (I hope it is just a “hark-back”) to the days when some teachers accepted possibly plagiarised material without question while others in the same school would not accept work which was not the student’s own.  Education of teachers as well as students supported by adoption of and school-wide adherence to academic honesty/ integrity policies have been effective strategies in raising awareness of the problems of plagiarism. It may be necessary, it probably will be necessary, for schools to revise their honesty and integrity policies to include considerations of ethical and unethical use of AI – and of course to make sure that all members of the school community understand the rationales, and acceptable use of AI too.

For sure, we cannot ignore ChatGPT and banning it is not the way to go.

10 thoughts on “Here we are again!

  1. Great article, thank you for the research too! I have started to warn our administration, and hope we can have this discussion soon with all our teaching staff. GPT Chat may change the way we educate our students, and it’s about time. Now we are forced to act. I look forward to how this will develop.

  2. Thanks for this thoughtful post John. I agree we have to find ways to work with AI and sticking our head in the sand or banning it is not the way to go. I look forward to seeing how this develops over the next few years… if it takes that long 😊.

  3. Thank you, John, this was a really helpful blog post and on the cutting edge of what we’re facing in schools.
    Would you ever consider ChatGPT a source itself? What if students cite the ChatGPT as their source?

    I don’t necessarily see ChatGPT as a source. I’m just trying to wrap my head around how all this works and am very interested to hear your take.

    • Thank you for the comment Kendra, and an interesting question you raise.

      I do not think that I would cite ChatGPT as a source (unless I was writing about ChatGPT, of course). But then I would avoid citing Wikipedia as a source (unless writing about Wikipedia); if I used information from Wikipedia, indeed if I use information from Wikipedia, I verify the information in an authoritative source and that is what I cite. I might say in passing that I was led to this train of thought by Wikipedia (or by ChatGPT), and that is a citation, but I would not reference it.

      There is also the thought that Wikipedia is verifiable – you can check that Wikipedia said what is claimed (if necessary, using the history feature) – but you just cannot do that with ChatGPT, every answer is different, even to the same question, the ever-growing number of articles attests that ChatGPT may even contradict itself, might at times “invent” its responses – and it certainly invents its sources, when asked to reference the sources used, supposedly used; there is evidence that it plagiarises, is at times hugely inaccurate (AI-written articles in US publication fail fact- check, plagiarism tests <https://www.techcircle.in/2023/01/26/ai-written-articles-in-us-publication-fail-fact-check-plagiarism-tests/>) and the CNet experience is another reason for avoiding it, at least for the moment (CNET’s AI Plagiarism Debacle <https://www.plagiarismtoday.com/2023/01/25/cnets-ai-plagiarism-debacle/>)

      I think APA might support that notion; a basic principle is that, if the source used is unpublished and so unverifiable unless you can gain sight of the unique document or overhear the actual conversation, then you can cite your source in the text but do not include it in the list of References – and your readership can judge how much weight they wish to place on what you claim was said.

      I have seen too that reputable journals are refusing to accept ChatGPT as authors (Science journals ban listing of ChatGPT as co- author on papers <https://www.theguardian.com/science/2023/jan/26/science-journals-ban-listing-of-chatgpt-as-co-author-on-papers>).

      The next generation of AI may be very different, but for serious research writing, I would avoid ChatGPT.

      The question for education is, is (sometimes) good enough good enough?

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.