Weaker Tutors

Introduction

The strongest near-term argument for AI in education is not that machines will replace great teachers. It is that AI may help less experienced tutors become noticeably better at helping students learn. That matters because many students, especially in poorer communities, are not taught by elite instructors. They are taught by people who are caring, available, under time pressure, and still learning how to teach effectively themselves.

Weaker Tutors illustration 1 The clearest evidence so far comes from Tutor CoPilot, a Stanford-led system that gave tutors live AI guidance during maths sessions. The overall gains were meaningful, but the largest improvements appeared among weaker or less experienced tutors. Students assigned to lower-rated tutors improved substantially more when those tutors had AI support. [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 points toward a broader possibility within the AI bloom vision. If advanced AI can help distribute expert-level guidance more widely, then educational quality may become less dependent on whether a child happens to encounter an unusually gifted teacher. Instead of replacing humans, AI could act as a force multiplier for ordinary educators, helping more students receive competent, responsive, personalised instruction.

What the Tutor CoPilot trial found

Tutor CoPilot was designed around a simple idea: many tutoring failures are not caused by lack of goodwill or subject knowledge alone, but by weak moment-to-moment teaching decisions. The system attempted to support tutors during live conversations by suggesting prompts, follow-up questions, explanations, and scaffolding strategies in real time. [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 study involved roughly 900 tutors and 1,800 students from historically underserved communities in a randomised controlled trial. Students whose tutors used the AI system were around 4 percentage points more likely to master topics successfully. [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 [2K-12 Dive]k12dive.comai tutor effectiveness stanford universityHow AI can improve tutor effectivenessOct 7, 2024 — Students whose tutors used Tutor CoPilot were 4 percentage points more likely to prog…

But the most important result was not the average improvement. It was where the improvement concentrated.

Students paired with lower-rated tutors improved by around 9 percentage points, far larger than the average effect. In practice, the weaker tutors moved much closer to the performance level of stronger tutors. [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 [EdWorkingPapers Researchers also found behavioural changes in the tutoring itself. Tutors using the system were more likely to ask guiding questions and less]edworkingpapers.comTutor CoPilot: A Human-AI Approach for Scaling Real-Time…by S Loeb — In a randomized controlled trial involving more than 700 tutors a… likely to simply hand 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 matters because effective tutoring is often less about delivering information and more about managing the learning process:

recognising misconceptions;
pacing explanations;
encouraging productive struggle;
noticing confusion early;
asking questions that help students reason through problems themselves.

Expert tutors learn these habits gradually through experience. Tutor CoPilot attempted to provide some of that expert pattern recognition immediately.

The system was also relatively cheap. Researchers estimated operational costs at roughly $20 per tutor annually during the trial. [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 low cost is important for any serious discussion of educational abundance. Expensive elite tutoring already exists. The question is whether high-quality support can scale broadly enough to matter for entire school systems.

Why novice tutors need real-time support

A common misunderstanding in debates about AI education is that tutoring quality mainly depends on subject expertise. In reality, many novice tutors already know enough mathematics, science, or language content to help students. What they lack is fluency in teaching decisions under pressure.

Live tutoring requires constant judgement calls:

whether to intervene immediately or wait;
whether a student is confused or merely hesitant;
whether to simplify or increase challenge;
when to encourage;
when to redirect;
how much help to give before the student stops thinking independently.

Experienced teachers often perform these actions almost automatically. Novice tutors frequently struggle not because they lack intelligence, but because they lack teaching instincts built through thousands of interactions.

That is exactly the kind of gap AI systems may be unusually good at narrowing.

Large language models can analyse conversational context rapidly and suggest candidate responses based on patterns learned from large corpora of instructional dialogue. In Tutor CoPilot, the AI effectively acted as a backstage coach, reminding tutors to use pedagogically stronger strategies. [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 resembles what happens in other professions when software captures fragments of expert judgement and makes them available to non-experts. Navigation systems help inexperienced drivers avoid errors. Medical decision-support tools help general practitioners notice conditions they may rarely encounter. AI tutoring systems may do something similar for educational interactions.

Importantly, the AI did not remove the human tutor from the loop. The tutor still handled motivation, emotional tone, trust, pacing, and relationship-building. The system offered suggestions, not autonomous instruction.

That distinction matters because education is deeply social. Students often persist because another person notices frustration, expresses confidence, or adapts sensitively to embarrassment and confusion. AI systems remain poor substitutes for many of these human dynamics.

The evidence so far therefore supports a hybrid model more than a replacement model: AI improves weaker tutors most because it compensates for gaps in instructional experience while preserving the human relationship itself. [Brookings]brookings.eduwhat the research shows about generative ai in tutoringBrookingsWhat the research shows about generative AI in tutoringJan 27, 2026 — Students of lower-rated tutors experienced the greatest be…

Weaker Tutors illustration 2

Why the gains may matter most for disadvantaged students

The educational implications become larger when viewed through inequality rather than average test scores.

Affluent families already compensate for weak schools through private tutoring, highly educated parents, smaller class sizes, enrichment activities, and access to stronger institutions. Students from poorer backgrounds are far more dependent on the quality of whichever tutor or teacher they happen to receive.

That creates a severe distribution problem in education. Small differences in teaching quality compound over years into large differences in confidence, attainment, and opportunity.

The Tutor CoPilot trial specifically involved students from historically underserved communities. [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 If AI support disproportionately improves weaker tutors, then the greatest gains may occur precisely where instructional quality is currently least reliable.

This is one reason AI-assisted tutoring matters within the broader AI bloom framework. The long-term optimistic case for AI is not merely that frontier experts become more productive. It is that high-level capabilities become widely available instead of remaining scarce luxuries.

In education, that could mean:

more students receiving personalised attention;

less dependence on unusually gifted individual teachers;
faster tutor training;
more consistent instructional quality across regions;
scalable support for overcrowded systems;
lower-cost educational access globally.

There are already early signs of this broader possibility. Separate research on hybrid human-AI tutoring models in low-income schools found evidence that lower-achieving students may benefit disproportionately from combined human and AI support systems. [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

Some pilot programmes in developing countries are also exploring teacher-guided AI tutoring models rather than fully autonomous AI instruction. [AICERTs - Empower with AI Certifications]aicerts.aiMoreover, students paired with lower-rated tutors gained nine points. World Bank teams…Read more… The pattern emerging across several experiments is similar: the strongest gains often appear where educational support was weakest initially.

Why this is different from “AI replaces teachers”

The Tutor CoPilot results cut against a popular assumption in both optimistic and pessimistic AI narratives.

The optimistic caricature says every child will soon receive a perfect AI tutor superior to any human teacher. The pessimistic caricature says schools will become depersonalised chatbot warehouses supervised by fewer adults.

The evidence so far points somewhere else.

AI appears most useful when embedded inside human educational relationships rather than replacing them outright. Brookings researchers reviewing the tutoring evidence argued that generative AI works best within “human-in-the-loop” systems that preserve pedagogical oversight and student engagement. [Brookings]brookings.eduais future for students is in our handsAI's future for students is in our hands14-Jan-2026 — AI can empower student learning by providing access to otherwise unavailable learni…

This distinction matters because pure AI tutoring systems face several persistent problems:

students disengage more easily; [k12dive.com]k12dive.comai tutor effectiveness stanford universityHow AI can improve tutor effectivenessOct 7, 2024 — Students whose tutors used Tutor CoPilot were 4 percentage points more likely to prog…
systems may over-help and weaken learning;
AI can reinforce misconceptions confidently;
dialogue quality often degrades over long conversations;
students may learn to optimise for answers rather than understanding.

Recent research on “pedagogical safety” in AI tutoring warns that many models still fail badly at maintaining healthy learning dynamics over extended interactions. [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

One of the central dangers is that AI systems frequently collapse productive struggle by revealing answers too quickly. Good tutors know when not to help immediately. Many language models are still weak at this restraint.

Human tutors therefore remain important not just emotionally but cognitively. They enforce pacing, maintain accountability, interpret confusion, and decide when AI suggestions are educationally appropriate.

In this sense, weaker tutors benefiting most from AI support is not evidence that humans are becoming obsolete. It may instead show that human capability itself can become more scalable.

Weaker Tutors illustration 3

What this could mean for educational inequality

If future AI systems become substantially more capable, the implications could extend beyond tutoring into a wider transformation of educational access.

Historically, elite education depended heavily on scarce human expertise: excellent teachers, low student-to-teacher ratios, intensive feedback, and personalised guidance. Those conditions are expensive because expert attention does not scale easily.

AI changes part of that equation.

If software can reliably help ordinary tutors perform closer to expert level, then educational quality may become less bottlenecked by the limited supply of exceptional instructors. In principle, that could make forms of personalised learning previously reserved for wealthy families available much more broadly.

That possibility fits directly into the larger AI bloom thesis: advanced AI may help civilisation distribute cognitive resources more widely, not merely automate routine work. Education is one of the clearest domains where intelligence itself is the scarce good.

But the outcome is not automatic.

Several conditions would shape whether AI tutoring reduces or worsens inequality:

whether systems remain affordable;

whether schools in poorer regions gain access;
whether human educators retain meaningful authority;
whether tutoring models are evaluated rigorously;
whether commercial incentives reward genuine learning rather than engagement metrics;
whether students continue developing independent thinking rather than dependency on AI assistance.

The Tutor CoPilot results are promising partly because they suggest a path that strengthens human capability rather than bypassing it. The weaker tutor does not disappear. The weaker tutor improves.

That may prove more socially important than building perfect autonomous AI teachers. A world where millions of ordinary educators become somewhat more effective could matter more than a world containing a few extraordinary AI tutoring systems used mainly by privileged students.

Within the wider debate about AI and humanity’s long-term future, this is one of the clearest concrete examples of what “making intelligence abundant” could look like in practice: not replacing people with machines, but helping more people exercise high-quality judgement, guidance, and care than was previously possible.

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: scale.stanford.edu
Title: how ai can improve tutor effectiveness
Link: https://scale.stanford.edu/news/how-ai-can-improve-tutor-effectiveness
Source snippet
SCALE InitiativeOct 7, 2024 — Students of lower-rated tutors who used the AI assistance increased their math proficiency up to 9 percenta...
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 115 — This preregis...
Source: k12dive.com
Title: ai tutor effectiveness stanford university
Link: https://www.k12dive.com/news/ai-tutor-effectiveness-stanford-university/728980/
Source snippet
How AI can improve tutor effectivenessOct 7, 2024 — Students whose tutors used Tutor CoPilot were 4 percentage points more likely to prog...
Source: edworkingpapers.com
Link: https://edworkingpapers.com/sites/default/files/ai24_1054_v2.pdf
Source snippet
Tutor CoPilot: A Human-AI Approach for Scaling Real-Time...by S Loeb — In a randomized controlled trial involving more than 700 tutors a...
Source: edworkingpapers.com
Link: https://edworkingpapers.com/sites/default/files/ai24-1054.pdf
Source snippet
Figure 2a, lower-rated tutors experienced a 9 p.p. increase in student's passing their exit ticket. (56% → 65% student passing rate from...
Source: brookings.edu
Title: what the research shows about generative ai in tutoring
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 tutoringJan 27, 2026 — Students of lower-rated tutors experienced the greatest be...
Source: arxiv.org
Link: https://arxiv.org/abs/2312.11274
Source snippet
arXivImproving Student Learning with Hybrid Human-AI Tutoring: A Three-Study Quasi-Experimental InvestigationDecember 18, 2023...

Published: December 18, 2023
Source: aicerts.ai
Link: https://www.aicerts.ai/news/how-tutor-co-pilot-systems-scale-teaching-capacity-worldwide/
Source snippet
Moreover, students paired with lower-rated tutors gained nine points. World Bank teams...Read more...
Source: arxiv.org
Title: arXiv Safe Tutors: Benchmarking Pedagogical Safety in AI Tutoring Systems
Link: https://arxiv.org/abs/2603.17373
Source snippet
arXivSafeTutors: Benchmarking Pedagogical Safety in AI Tutoring SystemsMarch 18, 2026...

Published: March 18, 2026
Source: scale.stanford.edu
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...Nov 17, 2025 — Gains were highest for students of lower-rated tutors (p.p.), and t...
Source: nssa.stanford.edu
Link: https://nssa.stanford.edu/news
Source snippet
| National Student Support Accelerator3 days ago — Students in the Tutor CoPilot condition were four percentage points... points) among...
Source: nssa.stanford.edu
Title: research studies
Link: https://nssa.stanford.edu/topics/research-studies
Source snippet
Studies | National Student Support AcceleratorMoreover, students paired with lower-rated tutors gained nine points.... Tutor CoPilot wer...
Source: scale.stanford.edu
Link: https://scale.stanford.edu/news?page=6
Source snippet
Stanford SCALE InitiativeNov 6, 2024 — Students of lower-rated tutors who used the AI assistance increased their math proficiency up to 9...
Source: nssa.stanford.edu
Title: research priorities
Link: https://nssa.stanford.edu/topics/research-priorities
Source snippet
tutors using Tutor CoPilot were 4 percentage points more likely to master math topics (p0.01). Gains were highest for students of lower-r...
Source: nssa.stanford.edu
Title: tutor copilot human ai approach scaling real time expertise
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...Dec 15, 2024 — Students in the Tutor CoPilot condition were four percentage points...
Source: brookings.edu
Title: ais future for students is in our hands
Link: https://www.brookings.edu/articles/ais-future-for-students-is-in-our-hands/
Source snippet
AI's future for students is in our hands14-Jan-2026 — AI can empower student learning by providing access to otherwise unavailable learni...
Source: brookings.edu
Link: https://www.brookings.edu/centers/center-for-universal-education/research-and-commentary/
Source snippet
Research and commentary | Brookings Center for...This exploratory study focused on the Indian context and... Artificial IntelligenceWha...
Source: brookings.edu
Link: https://www.brookings.edu/projects/brookings-global-task-force-on-ai-in-education/
Source snippet
Brookings Global Task Force on AI in EducationThe Brookings Global Task Force on AI in Education aims to ensure generative AI can be harn...
Source: brookings.edu
Title: do ais risks outweigh the benefits for students and schools
Link: https://www.brookings.edu/articles/do-ais-risks-outweigh-the-benefits-for-students-and-schools/
Source snippet
Do AI's risks outweigh the benefits for students and schools?20-Jan-2026 — When kids replace effortful learning with generative AI to sho...
Source: brookings.edu
Title: a new direction for students in an ai world prosper prepare protect
Link: https://www.brookings.edu/articles/a-new-direction-for-students-in-an-ai-world-prosper-prepare-protect/
Source snippet
A new direction for students in an AI world: Prosper...14-Jan-2026 — This report explores the potential risks generative AI poses to st...
Source: edworkingpapers.com
Title: In a randomized controlled trial
Link: https://edworkingpapers.com/ai24-1054
Source snippet
Tutor CoPilot: A Human-AI Approach for Scaling Real-Time...Oct 7, 2024 — We introduce Tutor CoPilot, a Human-AI system that models exper...
Source: linkedin.com
Link: https://www.linkedin.com/posts/eedi_what-the-research-shows-about-generative-activity-7424060074643558401-1wVC
Source: avidopenaccess.org
Link: https://avidopenaccess.org/resource/brookings-ai-study-part-ii-the-ai-teaching-dividend-benefits/
Source snippet
uctivity and replace time-consuming routine tasks, like lesson planning, grading...Read more...

Additional References

Source: victoryxr.com
Link: https://www.victoryxr.com/human-oversight-real-learning-and-the-future-of-ai-tutors-in-education/
Source snippet
AI Tutor Platform & VR LabsHuman Oversight, Real Learning, and the Future of AI Tutors...Brookings consistently returns to a human-in-th...
Source: overdeck.org
Link: https://overdeck.org/research-repository/tutoring/tutor-copilot-a-human-ai-approach-for-scaling-real-time-expertise/
Source snippet
Overdeck Family FoundationTutor CoPilot: A Human-AI Approach for Scaling Real-Time...This study uses an RCT to estimate the impacts of T...
Source: linkedin.com
Link: https://www.linkedin.com/posts/emollick_new-randomized-controlled-trial-by-the-world-activity-7285400569274380288-wh50
Source snippet
Ethan Mollick's PostSix weeks of after-school teacher-led AI tutoring increased test scores by.3 SD, equivalent to 2 years of typical le...
Source: linkedin.com
Link: https://www.linkedin.com/pulse/stanford-tutor-copilot-human-ai-approach-scaling-real-time-expertise-bx7me
Source snippet
Stanford: Tutor CoPilot – A Human-AI Approach for Scaling...This paper describes the development and evaluation of Tutor CoPilot, a huma...
Source: reddit.com
Link: https://www.reddit.com/r/machinelearningnews/comments/1fz9cil/researchers_at_stanford_university_introduce/
Source snippet
Researchers at Stanford University Introduce Tutor CoPilot...This innovative approach enables less experienced tutors to deliver high-qu...
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...Enhanced Student Outcomes: Tutor CoPilot significantly improves learning results, with...
Source: linkedin.com
Link: https://www.linkedin.com/posts/randy-wilkins-35798920_what-the-research-shows-about-generative-activity-7422462952760508416-VfG5
Source snippet
What the research shows about generative AI in tutoringWhat the research shows about generative AI in tutoring Of AI's many educational p...
Source: themoonlight.io
Link: https://www.themoonlight.io/en/review/tutor-copilot-a-human-ai-approach-for-scaling-real-time-expertise
Source snippet
[Literature Review] Tutor CoPilot: A Human-AI Approach...The core methodology utilized in the study comprises a randomized controlled tr...
Source: steve-grubbs.medium.com
Link: https://steve-grubbs.medium.com/human-oversight-real-learning-and-the-future-of-ai-tutors-in-education-98beaaafa593
Source snippet
You Hand Your Brain Over to AI - Steve GrubbsWhen the Brookings Institution released its January 2026 report, “What the Research Shows Ab...

Published: January 2026
Source: linkedin.com
Link: https://www.linkedin.com/posts/45deg_two-new-randomized-controlled-trials-just-activity-7436264527907913728-MqwR
Source snippet
AI Tutoring Outperforms Human Tutors in RCTs7 Mar 2026 — a separate RCT published in Scientific Reports found AI-assisted learners hit hi...