Human Tutors

Introduction

The strongest evidence for AI tutoring does not point toward classrooms without teachers. It points toward something more subtle and potentially more important: AI systems that help ordinary tutors make better decisions in real time. One of the clearest examples is Tutor CoPilot, a Stanford-led project that gave human tutors AI support during live maths sessions. Students improved, especially those taught by weaker tutors, but the AI was not acting alone. It was coaching the tutor behind the scenes. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024 [National Student Support Accelerator]nssa.stanford.eduNational Student Support AcceleratorStudy: AI-Assisted Tutoring Boosts Students' Math Skills7 Oct 2024 — And the weakest tutors became ne…

Human Tutors illustration 1 That distinction matters for the wider debate about AI and human flourishing. If advanced AI eventually helps make high-quality education abundant, the most effective path may not be replacing teachers with machines. It may be expanding the reach of human judgement, empathy and instructional skill through software that helps educators notice confusion earlier, ask stronger questions, and personalise support at scale.

Why the best tutoring systems still keep humans in the loop

For decades, education technology has promised “personalised learning” that would reduce dependence on teachers. Most of those promises disappointed. Students often disengaged, guessed through exercises, or learned to optimise software rather than understand subjects deeply.

The newer generation of generative AI tools changes some of the technical limits. Large language models can explain ideas conversationally, adapt examples, and respond flexibly to student questions. But the evidence so far suggests that these systems work best when combined with human oversight rather than deployed as fully autonomous instructors. [Brookings]brookings.eduBrookingsWhat the research shows about generative AI in tutoring27 Jan 2026 — Students of lower-rated tutors experienced the greatest ben… [Axios Tutor CoPilot became influential partly because it tested this hybrid approach directly. Instead of replacing tutors]arxiv.orgThis challenges the optimistic vision of AI as a 'personal tutor for every student.' Although genAI, like Khan Academy’s experimental Kha…, the system generated suggestions to help them during live sessions:

prompts for follow-up questions;
hints instead of direct answers;
alternative explanations;
guidance for handling misconceptions;
reminders to keep students reasoning actively. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024 [National Student Support Accelerator]nssa.stanford.eduNational Student Support AcceleratorStudy: AI-Assisted Tutoring Boosts Students' Math Skills7 Oct 2024 — And the weakest tutors became ne…

Researchers found that tutors using the system were more likely to adopt productive teaching strategies and less likely to simply give students the answer. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024

This is a very different vision from the common fear that children will sit alone with chatbots replacing schools entirely. The human tutor remained responsible for pacing, motivation, interpretation, emotional tone, and deciding when the AI’s suggestions were appropriate.

How AI can improve live tutoring decisions

Good tutoring is not just about knowing the subject. It involves hundreds of small judgement calls:

when to intervene;
when to stay silent;
whether confusion is conceptual or emotional;
whether a student needs encouragement, challenge, repetition, or a different explanation.

Expert teachers make many of these decisions intuitively after years of experience. Less experienced tutors often struggle with them.

Tutor CoPilot attempted to narrow that gap by modelling some of the behaviours associated with stronger tutors. In the randomised controlled trial involving around 900 tutors and 1,800 students from historically underserved communities, students whose tutors received AI assistance were 4 percentage points more likely to master topics. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024

The interesting part was not simply the improvement itself. It was how the improvement happened.

The AI nudged tutors away from answer-giving

One common failure mode in tutoring is over-helping. Tutors under time pressure often reveal answers too quickly, which can create the illusion of understanding while weakening long-term retention.

The Tutor CoPilot system instead encouraged tutors to use “guiding questions” and scaffolded hints. Researchers analysing more than 550,000 tutoring messages found increased use of strategies associated with deeper learning. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024

That matters because one of the central risks of AI in education is cognitive passivity. If AI becomes an effortless answer generator, students may appear productive while doing less genuine thinking. Systems designed to support human tutors can counteract that tendency by steering interaction back toward reasoning rather than output generation.

AI acted as a second set of eyes

Tutoring is cognitively demanding. Human tutors must simultaneously track curriculum goals, student confusion, pacing, engagement and emotional state.

AI support systems can reduce some of that load by surfacing possibilities the tutor may not have considered:

another analogy;
a simpler explanation;
a reminder to check understanding;
a way to rephrase a question.

In effect, the AI becomes a lightweight pedagogical assistant rather than the teacher itself.

This distinction may become increasingly important as education systems face teacher shortages, larger class sizes, and widening skill gaps. The practical value of AI may lie less in fully automating expertise and more in making limited human expertise stretch further.

Human Tutors illustration 2

Why weaker tutors may benefit most

One of the most striking findings from the Tutor CoPilot trial was that lower-rated tutors improved the most. Students taught by these tutors saw mastery gains roughly 9 percentage points higher than comparable students in the control group. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024

That finding has major implications for educational inequality.

Elite schools and wealthy families already have access to highly experienced teachers and private tutors. The larger global problem is that millions of students do not. In many regions, classrooms are overcrowded, teachers are overworked, and specialist subject expertise is scarce.

If AI systems can help average tutors perform more like strong tutors, the gains could be socially significant even without creating perfect automated teaching.

The economics matter here too. Researchers estimated that Tutor CoPilot cost roughly $20 per tutor annually. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024 That is tiny compared with the cost of training large numbers of expert tutors from scratch.

For advocates of an AI-enabled “abundance” future, this is one of the strongest educational arguments for AI:

not replacing scarce experts;
but multiplying their methods and judgement patterns across far larger populations.

Historically, excellent teaching has been difficult to scale because it depends heavily on tacit human skill. AI systems may partly loosen that bottleneck.

Where teacher judgement still matters

The Tutor CoPilot results were encouraging, but they also revealed the limits of AI assistance.

Tutors in the study reported cases where the system generated suggestions that were inappropriate for the student’s grade level or learning state. [arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time ExpertisearXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024…Published: October 3, 2024 Human tutors still had to filter, reject or reinterpret recommendations.

That points to a broader truth about current AI systems: fluency is not the same as judgement.

Emotional understanding remains deeply human

A student who is frustrated, embarrassed, exhausted or disengaged may need emotional reassurance more than another explanation. Teachers also manage classroom trust, social dynamics, motivation and confidence in ways that current AI systems still handle poorly.

Research and policy discussions repeatedly emphasise this point. Reviews from educators, governments and researchers consistently argue that AI works best as augmentation rather than replacement because learning is partly social and emotional, not just informational. [GOV.UK Assets]assets.publishing.service.gov.ukUK Assets Generative AI in educationAssetsGenerative AI in education - Educator and expert viewsJanuary 30, 2024 — Technology, including GenAI, is not a catch all solution t…Published: January 30, 2024 2arXiv

The concern is not merely sentimental. Students often learn difficult material because a trusted adult believes they can succeed.

Human Tutors illustration 3

Human teachers connect knowledge over time

AI tutors are often good at local explanation: solving the immediate confusion in front of the learner. Human educators do something broader. They connect new concepts to prior understanding, long-term goals, group discussion, identity formation and intellectual habits.

Several recent warnings about AI in education focus on “false mastery”: students appearing competent while outsourcing too much thinking to machines. [Axios]axios.comThis challenges the optimistic vision of AI as a "personal tutor for every student." Although genAI, like Khan Academy’s experimental Kha… [The Australian]theaustralian.com.auThe report highlights concerns that GenAI fosters a deceptive sense of mastery among students by generating high-quality outputs that mas…

Human judgement is critical for detecting that problem. A teacher can notice when a student sounds fluent but lacks conceptual understanding, or when AI-generated work exceeds the learner’s actual capability.

In this sense, the future role of teachers may become more demanding rather than less. As AI systems generate increasingly polished outputs, educators may need stronger diagnostic and mentoring skills to ensure students are genuinely learning.

The deeper implication for AI and human flourishing

The debate over “AI replacing teachers” is partly misleading because it treats education as information transfer alone. The more ambitious vision behind AI bloom is not simply faster access to answers. It is broader human cognitive empowerment.

That requires people who can:

reason independently;
evaluate claims;
collaborate;
sustain attention;
develop expertise;
adapt throughout life.

AI systems may help make those capacities more widely available, especially where good tutoring is currently scarce. But the strongest evidence so far suggests this happens most effectively when AI strengthens human teaching relationships rather than bypassing them.

The Tutor CoPilot model therefore matters beyond tutoring itself. It offers an early example of a wider pattern that may shape an AI-rich civilisation:

humans providing judgement, goals and social understanding;
AI providing scalable cognitive support;
the combination outperforming either alone.

That hybrid model may prove more important than the dream of fully autonomous AI education. A flourishing future depends not only on intelligent machines existing, but on ordinary humans becoming more capable through working with them.

Endnotes

Source: arxiv.org
Title: arXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time Expertise
Link: https://arxiv.org/abs/2410.03017
Source snippet
arXivTutor CoPilot: A Human-AI Approach for Scaling Real-Time ExpertiseOctober 3, 2024...

Published: October 3, 2024
Source: nssa.stanford.edu
Link: https://nssa.stanford.edu/news/study-ai-assisted-tutoring-boosts-students-math-skills
Source snippet
National Student Support AcceleratorStudy: AI-Assisted Tutoring Boosts Students' Math Skills7 Oct 2024 — And the weakest tutors became ne...
Source: brookings.edu
Link: https://www.brookings.edu/articles/what-the-research-shows-about-generative-ai-in-tutoring/
Source snippet
BrookingsWhat the research shows about generative AI in tutoring27 Jan 2026 — Students of lower-rated tutors experienced the greatest ben...
Source: axios.com
Link: https://www.axios.com/2024/08/15/ai-tutors-learning-education-khan-academy-wharton
Source snippet
This challenges the optimistic vision of AI as a "personal tutor for every student." Although genAI, like Khan Academy’s experimental Kha...
Source: arxiv.org
Link: https://arxiv.org/abs/2305.01185
Source: arxiv.org
Link: https://arxiv.org/pdf/2410.03017
Source snippet
arXivTutor copilot: A human-ai approach for scaling real-time...by RE Wang · 2024 · Cited by 104 — Thus, the tool builds tutors' skills...
Source: nssa.stanford.edu
Link: https://nssa.stanford.edu/sites/default/files/Tutor_CoPilot.pdf
Source snippet
National Student Support AcceleratorTutor CoPilot: A Human-AI Approach for Scaling Real-Time...by RE Wang · Cited by 117 — Notably, stud...
Source: assets.publishing.service.gov.uk
Title: UK Assets Generative AI in education
Link: [https://assets.publishing.service.gov.uk/media/65b8cd41b5cb6e000d8bb74e/DfE_GenAI_in_education_-Educator_and_expert_views_report.pdf](https://assets.publishing.service.gov.uk/media/65b8cd41b5cb6e000d8bb74e/DfE_GenAI_in_education-_Educator_and_expert_views_report.pdf)
Source snippet
AssetsGenerative AI in education - Educator and expert viewsJanuary 30, 2024 — Technology, including GenAI, is not a catch all solution t...

Published: January 30, 2024
Source: scale.stanford.edu
Title: Analysis of over 350,000
Link: https://scale.stanford.edu/publications/tutor-copilot-human-ai-approach-scaling-real-time-expertise
Source snippet
CoPilot: A Human-AI Approach for Scaling Real-Time...17 Nov 2025 — Gains were highest for students of lower-rated tutors (p.p.), and th...
Source: scale.stanford.edu
Title: how ai can improve tutor effectiveness
Link: https://scale.stanford.edu/news/how-ai-can-improve-tutor-effectiveness
Source snippet
AI can improve tutor effectiveness | SCALE InitiativeOct 7, 2024 — Students whose tutors used Tutor CoPilot were 4 percentage points more...
Source: edunlp.stanford.edu
Title: tutor copilot
Link: https://edunlp.stanford.edu/projects/tutor-copilot
Source snippet
CoPilot - EduNLP Lab - Stanford UniversityOct 31, 2024 — Tutor CoPilot is an AI system designed to provide expert-level guidance to tutor...
Source: nssa.stanford.edu
Title: Analysis of over 350,000
Link: https://nssa.stanford.edu/studies/tutor-copilot-human-ai-approach-scaling-real-time-expertise
Source snippet
CoPilot: A Human-AI Approach for Scaling Real-Time...15 Dec 2024 — Gains were highest for students of lower-rated tutors (p.p.), and th...
Source: tutor.com
Link: https://www.tutor.com/
Source snippet
Tutoring and Test Prep for K–12, Higher Education, and CareerTutor.com provides 24/7, expert, individualized academic and job support for...
Source: arxiv.org
Link: https://arxiv.org/pdf/2305.01185
Source snippet
Will AI Replace or Assist Teachers in Higher Education?by CKY Chan · 2023 · Cited by 442 — This paper explores the potential of artificia...
Source: theaustralian.com.au
Link: https://www.theaustralian.com.au/higher-education/student-reliance-on-ai-is-a-shortcut-that-masks-a-failure-to-learn-the-oecd-warns/news-story/868d0c5769c42446ba140807e8de8fd4
Source snippet
The report highlights concerns that GenAI fosters a deceptive sense of mastery among students by generating high-quality outputs that mas...

Additional References

Source: jsaer.com
Link: https://jsaer.com/download/vol-11-iss-7-2024/JSAER2024-11-7-152-158.pdf
Source snippet
AI Tutors vs Human TeachersStudies show AI tutors can boost learning outcomes. For example, one randomized trial found students learned s...
Source: linkedin.com
Link: https://www.linkedin.com/top-content/innovation/ai-in-education-innovation/how-ai-tutoring-can-boost-learning-outcomes/
Source snippet
How AI Tutoring Can Boost Learning OutcomesAI tutoring uses intelligent computer systems to provide personalized guidance, interactive fe...
Source: medium.com
Link: https://medium.com/syncedreview/sandford-us-tutor-copilot-transforms-real-time-tutoring-with-ai-driven-expert-guidance-b2a7cf5d5c18
Source snippet
Sandford U's Tutor CoPilot Transforms Real-Time...Greater Benefits for Less-Experienced Tutors: Less-experienced or lower-rated tutors s...
Source: theguardian.com
Link: https://www.theguardian.com/technology/2026/apr/02/pupils-england-losing-thinking-skills-because-of-ai-survey
Source snippet
Two-thirds of respondents observed a decline in thinking abilities among students, with some noting reliance on voice-to-text tools dimin...
Source: overdeck.org
Link: https://overdeck.org/portfolios/spotlight/nssa-tutor-copilot-a-human-ai-approach-for-scaling-real-time-expertise/
Source snippet
Tutor CoPilot: A Human-AI Approach for Scaling Real-Time...This report summarizes an efficacy analysis of the Project for Education Rese...
Source: overdeck.org
Link: https://overdeck.org/research-repository/tutoring/tutor-copilot-a-human-ai-approach-for-scaling-real-time-expertise/
Source snippet
Tutor CoPilot: A Human-AI Approach for Scaling Real-Time...This study uses an RCT to estimate the impacts of Tutor CoPilot—a novel human...
Source: povertyactionlab.org
Link: https://www.povertyactionlab.org/media/file-research-paper/tutor-copilot-human-ai-approach-scaling-real-time-expertise
Source snippet
Publication date. Sun, 11/30/2025 - 12:00. Research paper type. Academic Paper. J-PAL. J-PAL. 400 Main Street. E19-201.Read more...
Source: nepc.colorado.edu
Link: https://nepc.colorado.edu/sites/default/files/publications/PB%20Williamson_1.pdf
Source snippet
Effective Public Oversight, AI in Schools Will Do More...by B Williamson · 2024 · Cited by 70 — Advocates for AI claim that it will tran...
Source: edworkingpapers.com
Link: https://edworkingpapers.com/sites/default/files/ai24_1054_v2.pdf
Source snippet
increase in student's passing their exit ticket (56% → 65% student passing rate from control to treatment)...Read more...
Source: eric.ed.gov
Title: ERICTutor Co Pilot: A Human-AI Approach for Scaling Real
Link: https://eric.ed.gov/?ff1=eduGrade&id=ED661562&q=source%3A%22Annenberg+Institute+for+School+Reform+at+Brown+University%22
Source snippet
ERICThis study presents the first randomized controlled trial of a Human-AI system in live tutoring, involving 900 tutors and 1,800 K-12...