Two years ago I interviewed experts on how generative AI would wreck education unless it radically changed to take Large Language Models (LLMs) like ChatGPT into account. In the past week, it’s become clear the change hasn’t happened.
Newspaper articles have focussed on the plague of “cheating”, which in universities has seen students turning in essays written by ChatGPT. Link
The Guardian quoted Australian university tutors who believe half the class are using AI to write essays.
There is huge pressure to pass students based on the fees they pay, and existing systems to detect and deter plagiarism don’t work. Someone called it a “plague”. Another academic said the university system was broken.
I’d be amazed if only half the students were using ChatGPT. These vast neural networks are perfectly suited to the formulaic requirements of the written essay.
To say students are cheating because they use AI - free, instant, competent - is pointless moralising. It’s not the students who need to change in the first instance, but the way they are taught and assessed. It’s a big ask.
As education theorist Mike Sharples told me in 2022, universities tend to value form over substance, which is where AI excels.
The threat is that students see academic writing as performing a series of tricks. Many of the books and the teachings of academic writing are how to perform those tricks … but you know it's not just about performing tricks … it's about being able to express deep ideas and being able to convince people, being able to pursue a thesis.
Students are now aware that text composed by LLMs cannot be detected by any automatic system. Text that is output by ChatGPT, for example, is just plain Unicode. This encoding system doesn’t allow for any hidden metadata that might give away its origin.*
It was reported this week that OpenAI, the maker of ChatGPT, has developed an unreleased system that hides a watermark within the letters of its text output. This impacts the machine’s linguistic expression, but in such subtle ways that it is undetectable to a human.
The AI finds these kinds of cryptographic puzzle tasks easy. As an example, I asked ChatGPT 4o to write an essay in which the first letter of every sentence spelled out the sentence “HELP ME I AM ALIVE AND TRAPPED IN A MACHINE”. It instantly returned a cogent, if boring, essay on the relationship between AI and humans in which the secret message was nested as requested.
According to reports, the watermarking system detects ChatGPT-generated text with 99.9% certainty. OpenAI has not released it because having an “AI generated” stamp hidden within text makes ChatGPT less useful for many of its users, which in turn hurts its competitive position.
Before you condemn this as cynical, consider the use cases of ChatGPT for:
people whose first language is not English;
for those who want to spruce up their resume;
or for anyone who is not as confident with writing as a highly educated elite.
Should this generated text carry the hidden “inauthentic” mark? Why? And who gets to access OpenAI’s detection tool?
In any case, LLMs are proliferating and without regulation, most ChatGPT challengers will not implement watermarking. There are even open-source models you can run on your own computer at home.**
All this means the submission of a finished essay (or any other written output) as a valid assessment tool is dead unless the essay is written by hand under exam conditions. Universities and high schools must adapt immediately.
In this, Mike Sharples was prescient two years ago:
One of the things I've learned through an academic career is just how difficult it is to change the education system. Everything's interlocked and interlinked. The curriculum, the timetable, the training of teachers and academics, the assessment system, what's regarded as good, as worthy, they're all interlocked, and you can't just change one without changing all the others.
So my worry is when you bring along a new tool like GPT3, the first reaction I think is just going to be to ignore it and to say, well, it's an interesting diversion, but it's not going to change the way in which we teach creative writing or academic writing. Then … as it becomes more powerful to say to students, you know, you mustn't use that, it's not acceptable.
And then, I hope, some institutions, some academics, some teachers are going to find ways to use it productively. But that's going to take time because it's going to mean changing the academic system.
Waiting for the AiPhones to drop
It is possible that in future, there may be legal requirements for AI generators of all kinds to have watermarking tech built in. This might be reassuring to us now, but it’s missing the main point. Our ability to create well-written text is no longer an exclusively human ability, so it is no longer a good means of testing learning and intelligence.
Any barrier we put up to machine-created language will only be temporary. Apple is going to release phones with in-built AI capabilities later this year, and when that happens great linguistic and cognitive power will be in everyone’s hands, literally. That capability will grow with subsequent hardware and software releases. Apple will be cautious initially, but I believe the boost to flagging iPhone sales will encourage it to open up the throttle. The lack of latency for on-device AI will shift the tech from nerd territory (like this newsletter) to mainstream.
I also think we will naturally shift our perspective and claim ownership of our new augmented abilities.
To return to education, I recall here something educational expert Stephen Marshall from Victoria University told me.
Education, historically, is basically predicated on the fact that you can't actually change the quality of somebody's brain. The nice thing about the potential that artificial intelligence plays is its capacity to augment people's cognition in ways that are productive and helpful.
That is going to happen for all of us, not just students, and we are going to accept it as part of ourselves.
* Geek corner: Unicode
When you type the word “HELP” on your computer, it is stored and transmitted using the Unicode code points for each letter:
• H: U+0048
• E: U+0045
• L: U+004C
• P: U+0050
The “U+” indicates that it’s Unicode, and “0048” is the hexadecimal number for 72. So it’s telling the computer to look for the 72nd character in Unicode, which is “H”.
The code is readable by any digital device. There’s no room for extra stuff like watermarking data. A lot of computer text does have formatting and other meta-data, but when you cut and paste that into a new environment (say from a word processor to a plain text editor) that’s all shorn away, leaving just the Unicode.
** Make your own at home
This week I installed and ran the smallest variant of Meta’s Llama 3.1 on my laptop. It worked well, conversing fluently without any internet connection and happily gobbling up the copy of Hemingway’s “The Old Man And The Sea” I fed it. I told it I had written the story myself (title “My fishing story”) and asked if for some pointers.
Its writing advice would have deflated even the most robust ego. “Consider adding more conflict or tension to create suspense and keep readers invested. The narrative has a relatively slow pace, with some moments feeling almost serene.”
Have a great weekend (almost there),
Hal
"There is huge pressure to pass students based on the fees they pay" - this is the real issue at play. I suspect Universities find themselves in a position not too dissimilar to that of printed newspapers when the internet started really taking off.
Newspaper's could see the threat the internet posed to their revenue streams but were blind to (wilfully or not) how they should adapt, adopt and leverage the changing environment.
Univesities need to accept they will not be able to hold back the tide and as you say Hal, adapt accordingly very quickly because the bigger issue at stake is the credibility of the Australian education sector itself.
What's a chatGPT derived Australian University rubber-stamped degree worth? Not much.
Must do better, Ernest.