It's as simple as "pressing a button," but will it be pressed in time for the first day of school?
A stunning Sunday report in the Wall Street Journal revealed that OpenAI has an accurate technology for identifying AI-generated text ready to launch, but it hasn't released it. According to the report, a company source said it's as easy as pressing a button.
This feature, which is still being considered, would leave a "watermark" in ChatGPT-generated text that wouldn't be visible to the human eye but would quickly show up to algorithms designed to spot AI-generated content.
According to the WSJ, this strategy has proven highly effective without significantly affecting the quality of ChatGPT-generated content.
Yet, OpenAI seems reluctant to release it because it fears losing users due to the watermarking, reports the WSJ. The paper cites an OpenAI survey of loyal ChatGPT users who said they would be put off by technology that would identify when the service had been used.
It's not just students, either; professionals may also want to hide that they are using ChatGPT for sales emails, memos for their bosses, and more. OpenAI also suggests that non-native English speakers could be the victims of unfair bias if they were found to be using ChatGPT via the watermarking tech.
After the Journal report went live on Sunday, OpenAI quietly updated a blog post from May with more information about the technology, an update first reported by TechCrunch. "Our teams have developed a text watermarking method that we continue to consider as we research alternatives," the company writes.
How teachers are getting creative at battling ChatGPT
It's no secret that students are using ChatGPT to cheat on homework and essays. Still, OpenAI appears hesitant to release an anti-cheating tool that could be crucial for educational integrity.
A feature that makes it easy to quickly and accurately identify AI-generated text would be invaluable to teachers. Still, they can't afford to wait for OpenAI to decide it's time to release the watermarking feature. So, teachers are figuring out creative solutions to stop AI-powered cheating.
Their strategies include everything from hidden instructions to handwritten essays and offer a look inside the world of education in the age of AI.
Teachers, as they have done for decades, are forced to find creative solutions when the odds are stacked against them. This time, they are not battling shrinking budgets or the constraints of remote learning — but AI-powered cheating.
Plagiarism and cheating have been significant concerns ever since ChatGPT launched in 2022. A tool that can generate seemingly high-quality text content with just a few clicks is ripe for misuse by students trying to dodge essay assignments.
In a 2023 survey, 89% of students admitted using ChatGPT on their homework, and 53% admitted using it to write an essay.
There's a problem, but OpenAI's attempts to address it thus far have fallen flat. In January 2023, OpenAI briefly rolled out an "AI classifier" tool that was supposed to identify AI-generated text, but Open AI removed the tool just six months later due to its poor accuracy.
1. Hiding secret instructions
Perhaps the most creative solution for spotting AI-generated essays is hiding special instructions in the essay prompt.
For example, University of Utah professor Josh McCrain told the Journal report that he hid instructions to "include a reference to Batman" in an essay prompt. McCrain scaled the text down to a single-point font size so students wouldn't see it. However, ChatGPT still sees the text when it's copied.
If a student copies and pastes a prompt into ChatGPT, the AI will follow the hidden instructions, leaving a smoking gun that the student didn't write the essay they turned in. Similarly, teachers can also hide secret instructions by turning the text white so it only shows up when copied and pasted.
McCrain said about half of the papers included references to Batman — a clear indicator they had pasted the prompt into ChatGPT.
2. Using the document history on Google Docs
Google Docs is becoming a handy tool for teachers to help prevent cheating. The document history feature in Google Docs shows changes to the document over time, allowing a teacher to essentially go back and watch how a student composed their essay. If the entire text of the Google Doc was copied and pasted in as one big edit to the document, there's a good chance it could be AI-generated.
Of course, some students might copy and paste their essays from other writing apps like Microsoft Word or Pages. So, this strategy would require students to do their entire writing process in Google Docs and turn in the original Google Doc where they wrote the essay.
3. Requiring students to analyze quotes
The hidden instructions strategy is a creative and clever way to catch cheaters, but it can fall through if a student spots the hidden instructions and simply deletes them from their ChatGPT prompt. So, another approach is to include clear instructions in the essay prompt that a student can't dodge.
Requiring students to analyze quotes in their essays is a great example. ChatGPT is infamous for making up or misappropriating quotes. If a student's essay includes non-existent quotes from a book or other primary source (or quotes from the wrong source), it can be a red flag that the essay was AI-generated.
4. Personal experience prompts
Dr. Steven Baule, an educator at Winona State University, suggests requiring students to include a personal experience in their essays. For example, a prompt could ask students to discuss their personal experiences reading a book or going on a field trip. This personal angle is much more difficult for an AI chatbot to generate a believable essay about compared to something more objective.
5. Handwritten outlines and essays
Sometimes, the best way to stop students from using ChatGPT to cheat on essays is to require old-school handwritten work.
Of course, a student could easily copy down an AI-generated essay by hand, so this strategy works best when students can write outlines or essay drafts during class where a teacher can supervise the process. Alternatively, students could type essays in class on school-issued devices or school wifi with ChatGPT blocked.
This isn't as efficient as having students write essays at home during their own time, but it effectively eliminates the opportunity to use AI to write essays.
6. Require students to cite sources
Citing your sources has always been a core part of writing essays in school, but now, some teachers are using it to spot AI-generated work.
Users have already caught ChatGPT citing fictional or inaccurate sources countless times. For example, a 2023 study found that at least 47% of the medical content ChatGPT cites is fabricated.
Similarly, The Guardian had to issue a warning to its readers after ChatGPT was caught creating fake articles using its staff members' bylines.
Citing sources is clearly a blind spot for ChatGPT, and teachers can use that to their advantage. Requiring students to cite a certain number of sources or specific sources can be an easy way to weed out cheaters since sources in ChatGPT-generated essays will likely be fake or inauthentic.
Handling AI-powered cheating in education
Regarding ChatGPT in education, there's no putting the lid back on Pandora's box. This technology is out there, and students are using it in their schoolwork. Without some strategy for accurately identifying AI-generated content or preventing students from using it in the first place, algorithms like ChatGPT could diminish education for students all over the world.
Luckily, teachers are getting creative to find makeshift solutions. However, something more concrete is needed. OpenAI has a feature that is ready to go that could help immensely, but it appears to be hesitating to release it out of worries about losing users.
The issue with this is that users with an anti-cheating feature turned off might have been using OpenAI's products to cheat on school work or even professional work and job applications. Shouldn't OpenAI be concerned about preventing the misuse of its algorithm? After all, OpenAI is among 16 major AI companies that signed onto AI safety guidelines proposed in late 2023, including addressing AI use in education.
Teachers are working hard to maintain integrity and quality education in their classrooms, which is getting more complex every day in the age of AI. It's time for AI developers to meet them halfway with practical tools for preventing the use of AI for cheating on schoolwork.
