Korva is a scroll of the internet’s best ideas.
AI tutors shouldn't hand over answers
Too Helpful
That is fine for someone who only wants the answer. It fails a student who wants to learn, and a teacher who wants to assess understanding. CS50's syllabus calls using those tools unreasonable. The course cannot prevent it technologically, so it communicates the rule through ethics and policy, and it builds its own duck to lead students to solutions instead.
A Tutor That Asks
Picture a student who types that their code is not working and asks for ideas. The duck, seeing the code, notices the program is adding two integers and asks what data type input returns and how that might affect the addition. It behaves like a tutor in a conversation.
The Office Hours Problem
Office hours show why that matters. Six students in an hour is 10 minutes each, which has never been enough time for a struggling student. Ratios of 1 to 6, 1 to 10, or 1 to 20 are common even where teaching fellows are plentiful. A duck available at any hour lets comfortable students work on their own while the human staff focus on the students who need them most.
Why does one-on-one tutoring work so well?
In 1984, educational psychologist Benjamin Bloom reported that students taught one-on-one with feedback performed about two standard deviations better than students in a conventional classroom. That meant the average tutored student outscored about 98 percent of the classroom group. Researchers call the gap Bloom's 2 sigma problem. It has driven decades of effort to find methods that come close to tutoring at scale.
Do other AI tutors withhold answers too?
Khan Academy's Khanmigo, launched in 2023, is built on a large language model and is designed to respond with guiding questions rather than finished answers. Tutoring programs of this kind use the same approach as a human Socratic teacher. The method has long roots: Socrates used questions to lead students to conclusions, as recorded in Plato's dialogues.
Can software detect AI-written work?
Detection tools have proven unreliable. OpenAI released an AI text classifier in January 2023 and withdrew it in July 2023, citing its low rate of accuracy. A 2023 Stanford study led by James Zou found that several detectors wrongly flagged writing by non-native English speakers as AI-generated. Many schools now set policy on use of AI instead of relying on detection.
More from CS50
