Credit: Getty Images

The rise of large language models like ChatGPT has led to widespread concern that "everyone is cheating their way through college," as a recent New York magazine article memorably put it. Now, OpenAI is rolling out a new "Study Mode" that it claims is less about providing answers or doing the work for students and more about helping them "build [a] deep understanding" of complex topics.

Study Mode isn't a new ChatGPT model but a series of "custom system instructions" written for the LLM "in collaboration with teachers, scientists, and pedagogy experts to reflect a core set of behaviors that support deeper learning," OpenAI said. Instead of the usual summary of a subject that stock ChatGPT might give—which one OpenAI employee likened to "a mini textbook chapter"—Study Mode slowly rolls out new information in a "scaffolded" structure. The mode is designed to ask "guiding questions" in the Socratic style and to pause for periodic "knowledge checks" and personalized feedback to make sure the user understands before moving on.

It's unknown how many students will use this guided learning tool instead of just asking ChatGPT to generate answers from the start.

In an early hands-off demo attended by Ars Technica, Study Mode responded to a request to "teach me about game theory" by first asking about the user's overall familiarity with the subject and what they'll be using the information for. ChatGPT introduced a short overview of some core game theory concepts, then paused to ask a question before providing a relevant real-world example.

In another example involving a classic "train traveling at speed" math problem, Study Mode resisted multiple simulated attempts by the frustrated "student" to simply ask for the answer and instead tried to gently redirect the conversation to how the available information could be used to generate that answer. An OpenAI representative told Ars that Study Mode will eventually provide direct solutions if asked repeatedly, but the default behavior is more tuned to a Socratic tutoring style.

A 24/7 tutorbot

In building Study Mode, OpenAI said it was inspired by "power users" who were already trying to adapt ChatGPT into a personal tutor or test prep tool via specific sets of prompts. Through Study Mode, OpenAI says it hopes to give less technically minded users the ability to access "a personal tutor that never gets tired of their questions."

OpenAI says it worked with pedagogy experts who evaluated the mode's behaviors and gave the model "golden examples" of how ideal tutors would respond in certain situations. It also consulted with groups of college students who were given advance access to the new feature.

In a press conference attended by Ars, some of those students waxed rhapsodical about Study Mode's ability to push them toward the next important bit of knowledge and the confidence they gained in their ability to learn when using the tool. One student talked about the vulnerability and embarrassment of having to go to a TA's office hours for help and the flexibility of having multi-hour tutoring sessions available via ChatGPT at any time of the day.

Can you trust it?

People who are familiar with LLMs' long-standing tendency to confabulate completely false information might be reluctant to use these kinds of models as a study aid. In a press release, OpenAI allowed that the current Study Mode prompt "results in some inconsistent behavior and mistakes across conversations."

That said, a company spokesperson told Ars that the risk of hallucination was lower with Study Mode because the model goes through information in smaller chunks, calibrating along the way.

While Study Mode wasn't designed explicitly to address concerns about ChatGPT being used to cheat on assignments, a spokesperson told Ars that it could help allay educators' fears that students are simply using LLMs to get out of doing their own work. Study Mode will be added to its ChatGPT Edu product "in a few weeks" for subscribing schools that want to offer a different kind of AI experience to students.

For now, though, OpenAI says the specialized Study Mode system prompts are a "first step" to training similar behaviors "directly into our main models" in the future.