Day 69:03AI-generated essays are a growing concern, so this Canadian student created a free app to detect them
Edward Tian says he fears a world where everyone writes the same.
It was while in a Toronto coffee shop that the 22-year-old tapped into his AI knowledge — having studied it at Princeton — and used software already on his laptop to create an app that unearths AI-generated prose.
Following the fall semester, Tian says he travelled home for the Christmas holiday break dwelling on a recent artificial intelligence breakthrough, ChatGPT, and what it could mean for the future of authentic human writing.
“I think writing can be so beautiful,” said the computer science and journalism student. “There are parts and qualities of human writing that the machines can never do.”
I spent New Years building GPTZero — an app that can quickly and efficiently detect whether an essay is ChatGPT or human written
ChatGPT came out in November, and was released by San Francisco-based OpenAl. Users can ask it questions and assign it to produce things such as essays, poetry or computer code. It then scrapes text from across the internet to formulate a response.
When it surfaced, educational institutions were concerned about it being used for cheating.
Tian’s program, GPTZero, is free and was designed to red flag AI-generated writing. It was released in early January. He spoke with Day 6 host Brent Bambury. Here’s part of that conversation.
You’ve studied AI. When ChatGPT first came out in November, what was your impression?
We were like, wow, this is really good.
We knew this was coming because there’s a lot of these GPT models that were already released, and me and my friends at school were playing around with them, writing poems and raps about each other with these models.
They do a pretty good job. Sometimes it’s really hard to tell if this is human written or computer edited — sometimes it’s better than my own essays.
There’s things that chatbots can’t do, like write about the future, predict the future. But it’s really good at writing your college history essay.– Edward Tian, student and GPTZero app creator
What about when it’s not? Can you see the cracks in the program that was creating the text that you were looking at?
I would say there’s things that chatbots can’t do, like write about the future, predict the future. But it’s really good at writing your college history essay, for example.
It was pretty hard to tell the difference. And I think these models are only going to get better.
WATCH | Students share their thoughts about ChatGPT and AI tools for assignments
What made you think there is an app that can counter some of what this bot is doing?
I’ve been doing AI detection for my thesis research at Princeton this year, so I was pretty interested in looking at implicit bias qualities of these — like what AI writing has and what human writing doesn’t have — and being able to write originally.
AI writing is pretty consistent over time versus human writing that has bursts of originality. It’s almost like our short-term memory makes us have bursts of creativity and differences in the writing styles throughout time.
And those were the primary qualities we were looking for.
Your app, GPTZero, can it detect those biases?
I trained it on basically BBC news articles. And of course it’s not perfect.
I don’t want any teachers making academic decisions with this data. We’re building out a full tool that people can actually use more extensively.
Clearly you feel that there is something about the integrity of human created writing that needs to be preserved. What is that?
This is more the journalism student speaking than computer science student.
I think of a world 10 or 20 years from now where everybody’s writing with ChatGPT, that’s kind of sad because then no one’s writing originally and no one’s writing beautifully and everyone is writing the same.
In that world, maybe being able to write originally will remain a very important skill.
WATCH | Educators, students see challenges, opportunities with new ChatGPT AI software
Isn’t that part of what the new chat models are being able to do? When you tell it to write something in a particular style it seems to be able to do that. I mean, isn’t that beautiful writing?
It feels cheap. These models aren’t coming up with anything original. They’re taking what’s seen, which is entire portions of the internet, and just hitting these patterns.
There are some people out there, presumably students, that are calling you a narc.
Me and my girlfriend were going on Twitter and scrolling through these tweets. A lot of them were a little bit mean.
They’re like, “narc,” — “What a nerd,” and “Go get a life and get a girlfriend.”
We were just laughing.
I don’t want this to be a tool or a secret weapon teachers have at the end and just pull out.– Edward Tian, student and GPTZero app creator
Do you think that students should feel that they’re being spied on, that their work is being spied on?
The balance is there and we want to keep it accessible or a version of it accessible forever.
But no, I don’t want this to be a tool or a secret weapon teachers have at the end and just pull out.
I wanted this to be like a conversation. I want students to use it. And maybe the same technologies here can also detect what’s most original in a piece of writing that can help students.
I do think this should be a two-way street on how teachers and students can both use these technologies responsibly.
Edward, you graduate this spring. You’re also, as I mentioned, a journalism student. When you look at what ChatGPT is capable of today, do you foresee a future when it could actually replace print journalism?
No, I don’t think so in terms of journalism.
I mean, the core values of journalism is looking for the truth. And these technologies are great at regurgitating things they know, but they’re not good at finding the truth and they’re not able to do things like fact checking, reporting — collecting new information that doesn’t exist in their training data.
There’s so much they’re not able to do because, at the end of the day, these general models aren’t coming up with anything original. They’re seeing what they know, which is a lot, and regurgitating it.
Radio segment produced by Laurie Allan. Q&A has been edited for length and clarity.