Mostly it just finishes the work, unless you change how you use it. Stanford SCALE's November 2026 review of 52 causal studies found AI supports learning when people use it to explain material and can reduce learning when it produces the work for them. Naval Ravikant's advice to master foundations rather than summaries sets the rule: ask ChatGPT's study mode to explain, then rebuild the idea yourself.

Every founder I know is learning something new on a deadline right now: a pricing model, a regulation, a technical stack they are about to hire for. And almost all of them are doing it the same way, by asking an assistant to produce the memo, the summary or the slide, and reading it. The question worth asking is whether anything sticks. This month the best evidence base we have on that question got more than twice as large.

Stanford's SCALE Initiative published its November 2026 update of The Evidence Base on AI in K-12, by Seth Walker, Lily Fesler, JP Martinez Claeys, Chris Agnew and Susanna Loeb. It counts only strong causal studies, randomized trials and quasi-experiments, and the pool grew from 20 papers in the March 2026 review to 52, after 32 new papers published between October 2025 and June 2026 passed human review. Despite the title, most of the new student-facing studies were run with university students and adults, which makes them closer to a professional's situation than the headline suggests.

The central finding is a split, not a verdict. In the review's words, "AI appears more likely to support learning when students use it to understand new material than when they use it to produce work for them, although neither type of use consistently improves learning." The examples are uncomfortably familiar. Brazilian undergraduates who used AI finished a presentation faster and then did worse on a surprise conceptual test. Canadian programming beginners who generated complete solutions with ChatGPT performed well on the task and showed limited understanding of it. On the other side, a trial with 211 US undergraduates (Contractor and Reyes, 2026) found AI improved knowledge scores immediately and one week later, and the gains held up more for people who used it to explain than for those who used it to draft the essay.

Two more findings matter for anyone who thinks the answer is simply to struggle first. Two US studies found no significant learning benefit either from making students attempt a task before they got AI access or from offering help proactively when they seemed to struggle. And one found that students given the correct answer with an explanation straight away did better than those whose chatbot withheld answers and offered hints. So the line is not suffering versus convenience. It is whether the explanation ends up in your head or the deliverable ends up in your folder. The review also notes that people often report a better learning experience with AI even when their results do not improve, which is the trap: feeling informed is not evidence of learning.

Naval Ravikant has a line in The Almanack of Naval Ravikant that maps onto this almost exactly: read the originals before the interpretations, because foundations matter more than advanced concepts. An AI-written briefing is an interpretation by definition. It is useful for deciding whether something deserves your attention, and a poor substitute for owning the idea. Naval's point about specific knowledge, the knowledge that cannot be trained into someone else, makes the stakes clear for a founder: if the assistant produced your understanding of the market, that understanding is not specific to you.

The explain-then-rebuild rule (four steps)

  1. Name the job before you open the assistant. If you only need the output, a summary for a meeting or a draft you will not have to defend, let the AI produce it and do not tell yourself you learned the topic.
  2. If you need to own it, ask for explanation, not a deliverable. ChatGPT's study mode, which OpenAI built around guiding questions and knowledge checks and offers on Free, Plus, Pro and Team plans, does this by default; in any assistant, "explain this so I can write it myself" works.
  3. Close the chat and rebuild it. Write the one-page memo or sketch the model yourself. This is the step the 211-person trial points to: its gains held up better for people who used AI to explain than for those who used it to draft.
  4. Test yourself a week later without AI. Explain it out loud or on paper. If you cannot, you got the experience of learning without the learning.

The limits are real. These are students, not executives, and most of the papers are arXiv preprints, which the authors flag themselves. The review says plainly that the evidence "does not yet show whether learning gains persist over longer periods." The same week, Harvard's Compass Collaborative launched an AI Guidance by Age matrix built on more than 5,000 pieces of research, much of it older than generative AI, and its advice for young adults lands in the same place: let AI widen what you read and reflect your work back, and "leave the composing and the puzzling-out" with the learner. Treat both as the best current evidence, not settled science. For choosing a tutor tool, see the best AI tutor for a busy professional.

Sources: The Evidence Base on AI in K-12: A November 2026 Review, Stanford SCALE Initiative, November 2026; Z. Contractor and G. Reyes, Experimental Evidence on the Learning Impact of Generative AI, arXiv preprint, 2026; New AI Learning Matrix Brings Nuance to "Wild West" Moment, Inside Higher Ed, November 12, 2026; Introducing study mode in ChatGPT, OpenAI; Eric Jorgenson, The Almanack of Naval Ravikant (2020), chapters on reading and specific knowledge.


Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours