BEACON Foundation

Building AI Mentors Worth Learning From

Every child, whatever their family can afford, deserves a first-class education: patient, personal and expert. For most of history that meant a private tutor, and only the wealthy had one. AI is changing that. A good AI tutor now costs less than a cent a minute to run, which means the best teaching in the world could reach anyone with a phone. Whether it does depends on how these tools are built.

This guide sets out what BEACON believes a great AI mentor should do, offers a standard for doing it honestly, and points to tools worth trying today. It grew out of ILUMINA, a concept BEACON developed for Socratic dialogue with the great minds of history. Much of what ILUMINA set out to build now exists, so rather than build one more product, we are sharing the thinking openly. Use it, adapt it and build on it.

What the evidence says

One-to-one tutoring is among the best-supported ideas in education, and early AI tutors are showing real promise. In a Harvard physics course, students learned about twice as much from a carefully designed AI tutor as from an active-learning class. Studies in Nigeria and Ghana found large gains in a matter of weeks.

The common thread is design. The Harvard tutor was built to give hints, not answers. An AI that simply hands over answers can let a learner look like they are learning while doing none of the thinking.

What a great AI mentor does

• Asks before it tells. It guides the learner toward the answer with questions and hints, and gives the answer only when that serves the learning.
• Moves at the learner’s pace. It moves on only when understanding is clear, and never treats a question as too basic.
• Keeps its turns short. A good reply is usually a question, not a lecture. Short replies also cost less, which helps keep good tools affordable.
• Connects ideas across subjects, so learning builds into understanding rather than a pile of separate facts.
• Is honest about its limits. When it doesn’t know, it says so rather than guessing.

Historical figures, done honestly

A conversation with a great mind of the past can be a powerful way to learn. It can also go badly wrong. Early historical-figure chatbots invented quotes, got facts wrong, and even had real people express regrets they never expressed. A mentor that speaks as a real person should meet a higher bar:

1. Grounded in sources. What the figure says rests on what they actually wrote, said or did, drawn from material reviewed by historians.
2. No invented words. The figure never presents a made-up quote as real, and makes clear when it is paraphrasing or interpreting.
3. No rewritten history. Figures are not made kinder, wiser or more modern than they were. Their blind spots and failings are part of the lesson.
4. Speculation stays fictional. “What would they think of today?” belongs to clearly invented characters, so no speculation is ever put in a real person’s mouth.
5. Clear about what it is. The learner always knows they are talking with a simulation built from the historical record, not the person.

Measuring real thinking: the Curiosity Index

Time spent and words typed say little about learning. The kinds of questions a learner asks say much more. The Curiosity Index sorts them into three levels:

• Level 1, Clarifying. “What does that word mean?” Building vocabulary and facts.
• Level 2, Hypothetical. “What if…?” Testing ideas and exploring.
• Level 3, Cross-disciplinary. Connecting ideas across subjects or eras. Real synthesis.

A learner whose questions climb from level 1 toward levels 2 and 3 is thinking more deeply. Any tool builder, teacher or parent can use it.

Choosing a tool today

Five questions worth asking before a learner settles on a tool:

• Does it guide, or just answer? Try a homework problem and see whether it asks questions first.
• Is it built for the learner’s age? Many general AI assistants require users to be 13 or 18. Check before signing up a child.
• What does it cost? Several good options are free.
• What happens to the learner’s data? Look for clear privacy terms, especially for children.
• Can it be wrong? All of them can. Encourage learners to check anything important.

Where to start

These are well-established options as of October 2026, chosen from public reviews and reporting. Listing a tool here is not an endorsement, and prices and age rules change, so check each one’s current terms.

• ChatGPT Study Mode. A guided, step-by-step mode inside ChatGPT. Teens 13 to 17 get a dedicated teen experience with Study Mode and parental controls.
• Claude Learning mode. Choose “Learning” in Claude’s style menu. For adults 18 and over.
• Gemini Guided Learning. A step-by-step mode inside Google’s Gemini, also offered through school accounts, with age-based limits.
• Khanmigo. Khan Academy’s AI tutor for school-age learners. About $4 a month for families, with free tools for teachers.
• Socra. A Socratic tutor for students and self-directed learners. Free to start.
• Hello History. Conversations with historical figures. Worth using with the honesty standard above in mind.

Use it, adapt it, improve it

Everything on this page is free to use, share and build on. If you build learning tools, teach, or study how people learn, we would welcome your corrections and ideas: contact@beaconfound.org.