All corrections
LessWrong March 13, 2026 at 08:32 PM

www.lesswrong.com/posts/bs3yj8vLDKNnoa95m/five-recent-ai-tutoring-studies

1 correction found

1
Claim
The AI tutor was based on GPT-4 with a system prompt instructing it to only give incremental hints and prompts for how to handle each question, the writing of which took “several months.”
Correction

The paper says the lesson-specific system prompts and solutions were prepared over a few days, not several months. The “several months” figure refers to building the broader AI tutor platform.

Full reasoning

The study’s supplemental methods explicitly separate two kinds of work:

  1. Preparing system prompts and lesson-specific solutions, which the authors say “was done over a few days.”
  2. Developing the AI tutor platform itself, which they say “took several months.”

So the post misattributes the several-month timeline to writing the system prompt. According to the paper, the months-long effort was for platform development, not for drafting the specific tutoring prompt used in a lesson.

1 source
  • AI Tutoring Outperforms Active Learning

    Preparing system prompts for questions and solutions for a particular lesson was done over a few days... The most significant time commitment involved in preparing the AI-supported lessons was development of an AI tutor platform... which took several months.

Model: OPENAI_GPT_5 Prompt: v1.16.0