Off to detention |
When I first tried ChatGPT, the almost magically sophisticated artificial intelligence tool made by OpenAI, my initial reaction was to marvel at its crisp answers and conversational lucidity. My second thought, as the father of high-school kids, was to say a small prayer for the teachers of the world.
Because ChatGPT knows things. It knows the metaphorical significance of Boo Radley in To Kill a Mocking Bird. It can opine with relative clarity on the merits of autobiography as a literary form in relation to Richard Wright’s Black Boy. It can wax on about morality in Shakespeare’s final play, The Tempest.
It can’t yet spin perfect answers right away. But that’s the charm of this remarkable AI chatbot; it riffs off user questions and then hones its answers, based on a large corpus of internet content that OpenAI’s language processing model has ingested over the past few years.
To put it rather bluntly: ChatGPT can do students’ homework for them. “The undergraduate essay” wrote The Atlantic, “is about to be disrupted from the ground up.” When the service went offline this week due to high demand, my Bloomberg Opinion colleague Trung Phan tweeted, “There must be a biblical amount of semester-end essay cheating going on right now.”
This isn’t just the usual Sisyphean cycle of panic that accompanies all new technologies. Even experts in the field of safeguarding the originality of student work are marveling at the advances represented by ChatGPT. “What caught me off guard was how much of a leap forward it was,” said Eric Wang, vice president of artificial intelligence for Turnitin, a plagiarism detection service owned by Advance Publications Inc., the parent company of Conde Nast.
Founded in 1999, Turnitin works with about 16,000 school systems around the world to detect student plagiarism. Its software scours the web to see if a student ripped off an essay from, say, Wikipedia, and its algorithms can identify problems like an overreliance on citations or even spot examples of contract cheating — hiring third parties, sometimes from developing countries, to complete their homework for them.
But ChatGPT represents a more significant challenge: an AI-powered chatbot from a Silicon Valley startup, funded to the tune of $1 billion from Microsoft Corp. alone. Wang said for now, ChatGPT’s answers should be easily identifiable both by teachers and Turnitin software. The service makes lots of factual errors, and its language model tends to generate linear sentences and pick broad, obvious words, instead of the occasionally narrower vocabulary that a student would select. This creates signals that could be detectable by Turnitin and other anti-plagiarism tools.
But Wang also acknowledges that AI is only going to get more sophisticated — perhaps quickly. OpenAI, for example, is currently showing partners the fourth iteration of its language model, GPT-4; ChatGPT is based on a variation of GPT-3. “It’s early days, the halcyon days of the field,” Wang said. “We’ll look back in a year or two and think about these things with wonder and recognize how primitive they were.”
OpenAI appears to recognize that it bears some of the responsibility for guarding against deceitful uses of its service. According to TechCrunch, Scott Aaronson, a computer science professor at the University of Texas, Austin, and a guest researcher at OpenAI, recently said in a lecture that the company is studying hiding cryptographic signals, called watermarks, in ChatGPT results, so that they’ll be more easily identifiable by companies like Turnitin. Independent experts worry that it will be easy for committed adversaries — say, a technically sophisticated student — to detect and strip out these signals.
Another option is for schools and colleges to acknowledge the new reality posed by AI and change how they evaluate students. “The days of asking students to summarize the themes in The Odyssey – no, that’s an easy question one would get from an AI model,” said Annie Chechitelli, Turnitin’s chief product officer. Instead, teachers could ask the class to explore themes of hospitality in the epic poem and then put them in the context of current events, which would defeat AI models like ChatGPT that don’t ingest current news sources. She also envisions a time when teachers ask students to record a video explaining or defending what they wrote.
But Chechitelli is also pretty solemn about the challenges schools will face over next few years. “The glacial speed with which change happens in higher education is going to make things hard,” she said. “The conversations need to start on campus today. Students are already well aware of what tools they can use.” —Brad Stone