Almost as soon as ChatGPT burst on the scene and stoked fears of widespread cheating, support for teachers also arrived in the form of detectors promising to sniff out writing generated by the artificial intelligence tool.
But just as ChatGPT sparked big questions around the purpose and different ways of teaching writing or what it means to communicate or be creative, these tools come with their own potential problems.
The online cheating or plagiarism detectors make mistakes. Teachers need training to understand and cope with their limitations. Too much reliance on them may leave schools poorly positioned to teach writing in a post-ChatGPT world. And AI writing tools are almost certain to get better at eluding these digital whistleblowers.
鈥淭hese types of detectors could be maybe one tool in an arsenal,鈥� said Christopher Doss, a quantitative researcher at the Rand Corporation, a research organization. 鈥淚 don鈥檛 think they would ever be the only tool, so that a teacher can just create a file of assignments, run it through the system, [and] get a 100 percent accurate yes or no, and then they move on with their life. I just don鈥檛 think it鈥檚 ever going to be that simple.鈥�
There are already several programs that help identify AI-crafted writing, and many more could become available soon. The Watson AI lab at the Massachusetts Institute of Technology developed a detector called GTLR. Packback, a learning platform, added an AI detection tool to its existing program. Even OpenAI, the developer of ChatGPT, has one. And Turnitin, a prominent plagarism detector, is developing one. It鈥檚 generally unclear what the error rates are for the products currently available.
Detectors could launch an 鈥楢I Arms Race鈥�
Even if these programs can accurately pinpoint whether work was produced by the current version of ChatGPT, a fresh iteration of the AI writing tool is due out later this year, potentially sending educators back to square one.
鈥淚t鈥檚 a little bit of an arms race,鈥� said Andreas Oranje, the vice president of Assessment and Learning Technology for the Educational Testing Service鈥檚 research and development division. 鈥淓ventually, these models [like ChatGPT] will incorporate more human behavior, get smarter. And so, then they become a little bit better and the tools that were made to detect [AI writing] are no longer working.鈥� He likened the process to bacteria evolving, thwarting antibiotics.
But Annie Chechitelli, the chief product officer for Turnitin, a company that offers plagiarism detection software widely used in K-12 schools, believes it will be possible to spot ChatGPT鈥檚 writing for quite some time.
The bot tends to use hackneyed phrases鈥攖hink 鈥渋t was a dark and stormy night鈥濃�攁nd constantly repeats the same wording and ideas.
鈥淲e鈥檙e seeing other traces come out that I鈥檓 confident, for the near future, will remain,鈥� Chechitelli said. Turnitin expects to release its own AI detection software in time for the 2023-24 school year.
鈥業 can鈥檛 radically transform my classroom just yet鈥�
Educators will ultimately need to figure out how to teach writing in a way that incorporates tools like ChatGPT, said Joshua Rosenberg, an assistant professor of STEM education at the University of Tennessee in Knoxville.
At some point, asking students to write without consulting AI will become 鈥渁lmost like requiring students not use the calculator when completing math problems,鈥� he said. 鈥淭hat鈥檚 where this is going.鈥�
But most educators aren鈥檛 prepared for such an abrupt transition, smack in the middle of the school year, Rosenberg said.
鈥淚 totally empathize with teachers who are like, 鈥榃hat the heck? It鈥檚 January! It鈥檚 been a crazy three years,鈥欌�� Rosenberg said. 鈥溾�楢nd I want to make sure that my students are understanding writing or English language arts concepts that I want them to learn and that [they] are expected to learn based on our state standards. I can鈥檛 radically transform my classroom just yet.鈥欌��
Some of the detection tools, though, aren鈥檛 user-friendly yet. Teachers that want to deploy one free detection program鈥擥PT2 Output Detector, an open-source tool created with code from ChatGPT creator OpenAI鈥攃ould be in for a frustrating process. It often crashes, according to some users. And in explaining its error rates, the tool uses technical jargon most educators won鈥檛 easily grasp.
That鈥檚 why two literacy-focused education technology nonprofit organizations, Quill.org and CommonLit.org, created an AI detector platform designed with teachers in mind called . It is essentially a teacher-tailored version of the GPT2 Output Detector, which the organizations say has an accuracy rate of detecting plagiarism roughly 80 to 90 percent of the time.
The move comes in response to a survey of more than 750 educators who use Quill鈥檚 literacy platform. More than 80 percent said they are concerned about students using ChatGPT to complete their writing assignments. And even though the latest, headline-grabbing version of ChatGPT has only been around since late last year, 17 percent of educators surveyed said they had already seen students try to pass off the bot鈥檚 work as their own original writing.
In that environment, teachers need to have some sort of mechanism to detect whether an assignment has been outsourced to AI, even if 鈥渢hey鈥檙e not perfectly accurate,鈥� said Peter Gault, Quill鈥檚 founder and executive director.
鈥淲e don鈥檛 think that this is a be all, end all solution,鈥� Gault said of his organization鈥檚 platform. 鈥淭here will be false positives and false negatives, and that鈥檚 something that needs to be taken really seriously. But we think this is a helpful stopgap for teachers to give them more data and information than they otherwise would have access to.鈥�
Other types of tools would be helpful too, he added. For instance, developers could create one that analyzes keystrokes or various versions of a draft to decide whether a particular piece of writing was produced by a human or a robot, said Gault and Michelle Brown, CommonLit鈥檚 founder and chief executive officer.
That type of technology might help educators get to a middle ground, where they can use ChatGPT to inform some writing instruction, but also expect students to do their own original work, Gault said.
鈥淩ight now, [ChatGPT is] either banned or it鈥檚 totally embraced,鈥� Gault said. 鈥淗ow can we use the tool while still ensuring academic authenticity in the classroom?鈥�
Teachers must 鈥榬ecognize the tool may be wrong鈥�
Teachers who rely on these detectors need to be aware of their limitations, Rosenberg said.
It would be unfortunate for a detector to erroneously conclude that a bot-crafted essay was human-produced. But it could be even worse for a student who completed an assignment honestly to be accused of using tech to cheat, Rosenberg said.
That 鈥渃ould put a student through a really negative, possibly humiliating process,鈥� Rosenberg said.
Teachers who don鈥檛 want their students using ChatGPT as a writing tool need to make that expectation clear from the outset, Rosenberg said.
If a student鈥檚 essay is flagged by an AI detector, teachers should see that as a 鈥渟tarting point for a conversation鈥濃�攏ot a final verdict, Rosenberg said. Teachers could ask students to tell them about their writing process and 鈥渞ecognize that the tool might be wrong,鈥� he said.
The truth could also be complicated. Students may have used AI as a starting point for generating ideas, or employed a tool like Grammarly, which may rewrite sentences to make them more coherent, Rosenberg said.
There are other clues that teachers could look for to figure out if a student relied, at least in part, on AI to complete an assignment, Doss said.
鈥淚n a really kind of black and white case, if you have a student who has struggled to write, and then they give you a really, really nicely written [piece], you might be suspicious as to whether or not they were actually the ones that created the product,鈥� Doss said.
While Brown believes ChatGPT and tools like it will eventually be part of writing instruction, she said students will miss out if they rely too heavily on AI to do their work for them.
鈥淲riting and learning how to write helps us learn how to think and it helps you organize your thoughts and helps you generate language,鈥� she said. 鈥淚 am not ready to say that all of education should just throw in the towel on the way we鈥檝e taught writing and thinking and organization just because [developers] made cool AI.鈥�
