Why weaker tutors gained most from AI help

Introduction

One of the most striking findings from the Tutor CoPilot study was not that AI improved tutoring on average. It was that the biggest gains came from the weakest tutors.

Students working with lower-rated tutors who received AI assistance were up to nine percentage points more likely to master material than comparable students whose tutors worked without the tool. In effect, Tutor CoPilot narrowed part of the gap between weaker and stronger instructors. Rather than making the best tutors dramatically better, it helped less experienced or less effective tutors behave more like experts.[National Student Support Accelerator]nssa.stanford.edututor copilot human ai approach scaling real time expertiseNational Student Support AcceleratorTutor CoPilot: A Human-AI Approach for Scaling Real-Time…15 Dec 2024 — With the help of the tool…

That result matters far beyond one tutoring platform. It points towards a broader possibility in the AI bloom debate: that advanced AI may not create the largest gains by replacing skilled people, but by making scarce expertise easier to access and apply. If AI can reliably help ordinary workers perform closer to expert level, then high-quality teaching, medicine, engineering, management and many other forms of knowledge work could become substantially more abundant.

What lower-rated tutors lacked before AI support

The Tutor CoPilot researchers did not start from the assumption that weaker tutors lacked intelligence or subject knowledge. The problem was often that they lacked teaching judgement in the moment.

Effective tutoring depends on many small decisions that experienced instructors make almost automatically. They notice confusion early, identify misconceptions, ask productive follow-up questions, and know when not to provide an answer directly. These skills are difficult to learn from manuals alone because they depend on real-time interpretation of a student’s thinking.[arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-TimeTutor CoPilot: A Human-AI Approach for Scaling Real-Time…October 3, 2024 — by RE Wang · 2024 · Cited by 104 — We introduce Tutor…Published: October 3, 2024[EduNLP Without that experience]edunlp.stanford.edututor copilotEduNLP LabTutor CoPilot - EduNLP Lab - Stanford UniversityOct 31, 2024 — Tutor CoPilot is an AI system designed to provide expert-level g…, tutors often fall into predictable patterns:

• Giving away answers too quickly.
• Offering generic encouragement instead of targeted guidance.
• Missing signs that a student has misunderstood a concept.
• Moving through material without checking reasoning.
• Focusing on completion rather than understanding.[hpttreasures.wordpress.com]hpttreasures.wordpress.comTutor Co Pilot: A Human-AI Approach for Scaling Real-TimeTutor CoPilot: A Human-AI Approach for Scaling Real-Time…October 7, 2024 — by RE Wang · 2024 · Cited by 117 — By offering real-time gu…Published: October 7, 2024

These are not necessarily failures of effort. They are often failures of expertise. A novice tutor may genuinely want to help but lack the accumulated experience needed to recognise what kind of intervention is most useful at a particular moment.

This is why the Tutor CoPilot result is interesting. The AI was not primarily adding knowledge. It was supplying judgement cues that weaker tutors had not yet internalised.

How real-time prompts changed tutoring behaviour

Tutor CoPilot worked by analysing the tutoring conversation and generating suggestions for the tutor during the session. Rather than speaking directly to students, it acted as an instructional assistant sitting beside the tutor.[arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-TimeTutor CoPilot: A Human-AI Approach for Scaling Real-Time…October 3, 2024 — by RE Wang · 2024 · Cited by 104 — We introduce Tutor…Published: October 3, 2024 EduNLP The system frequently suggested behaviours associated with effective tutoring research:[edunlp.stanford.edu]edunlp.stanford.edututor copilotEduNLP LabTutor CoPilot - EduNLP Lab - Stanford UniversityOct 31, 2024 — Tutor CoPilot is an AI system designed to provide expert-level g…

What makes this important is that these are exactly the kinds of moves that novice tutors often underuse.

Researchers reported behavioural shifts consistent with this explanation. Tutors receiving AI support became less likely to simply provide answers and more likely to ask productive questions that kept students engaged in the reasoning process.[AICERTs]aicerts.aiEmpower with AI CertificationsHow Tutor Co-Pilot Systems Scale Teaching Capacity…In the Stanford study, weaker tutors' stude… - Empower with AI Certifications

In other words, the AI did not mainly improve outcomes by generating superior mathematical explanations. It improved outcomes by nudging tutors towards better teaching habits.

This resembles a common pattern seen in other professions. Expert performance often depends less on possessing secret information than on repeatedly making good decisions in situations where less experienced people make mediocre ones. If AI can reliably supply those decision cues, performance gaps can shrink.

Why the strongest tutors improved less

The study’s most revealing result may be that the strongest tutors did not improve nearly as much as the weakest.[National Student Support Accelerator]nssa.stanford.edututor copilot human ai approach scaling real time expertiseNational Student Support AcceleratorTutor CoPilot: A Human-AI Approach for Scaling Real-Time…15 Dec 2024 — With the help of the tool…

At first glance, this might seem disappointing. If AI is powerful, why did it not produce large gains everywhere?

But the pattern is exactly what many researchers would expect.

Highly effective tutors were already using many of the strategies the AI recommended. They already knew when to ask guiding questions, when to slow down, and how to uncover misconceptions. The system therefore had less room to improve their behaviour.

This is a familiar phenomenon in education and training. Interventions often generate larger effects at the bottom of a performance distribution because that is where the easiest improvements remain available.

An experienced tutor who already performs at a high level may gain only marginally from an extra suggestion. A novice tutor who routinely misses opportunities for better instruction may gain substantially from the same prompt.

Seen this way, Tutor CoPilot was functioning less like a replacement expert and more like an expert mentor available in real time.

The importance of reducing variance

Many discussions of AI focus on raising peak performance. The Tutor CoPilot results highlight something different: reducing variance.

In education systems, average performance matters, but consistency matters too. A student assigned to an excellent tutor may receive a dramatically different experience from a student assigned to a weak one.

Tutor CoPilot appeared to narrow that gap. Reports on the study described weaker tutors becoming nearly as effective as their higher-rated peers.[The 74 Million]the74million.orgThe 74 MillionStudy: AI-Assisted Tutoring Boosts Students' Math Skills7 Oct 2024 — And the weakest tutors became nearly as effective as t…[National Student Support Accelerator]nssa.stanford.edututor copilot human ai approach scaling real time expertiseNational Student Support AcceleratorTutor CoPilot: A Human-AI Approach for Scaling Real-Time…15 Dec 2024 — With the help of the tool…

That may sound like a modest achievement compared with visions of superintelligence or fully automated education. Yet for large institutions, reducing variance can be enormously valuable.

A system that makes average tutors somewhat better is useful. A system that makes weak tutors substantially less weak can transform service quality across an entire organisation.

This distinction helps explain why AI assistance may matter most in sectors where expertise is unevenly distributed. The greatest gains often come not from improving the best performers but from lifting the long tail of ordinary performance.

What this implies for making expertise more abundant

The broader significance of Tutor CoPilot lies in its model of expertise distribution.

Many valuable skills are difficult to scale because they depend on scarce human experience. The world has limited numbers of exceptional teachers, doctors, engineers, therapists and managers. Training replacements takes years or decades.

The Tutor CoPilot result suggests another path. Instead of waiting for everyone to become an expert, AI may help less experienced people access fragments of expert judgement while they work.[arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-TimeTutor CoPilot: A Human-AI Approach for Scaling Real-Time…October 3, 2024 — by RE Wang · 2024 · Cited by 104 — We introduce Tutor…Published: October 3, 2024

That possibility aligns closely with one of the central ideas behind AI abundance. The key question is not whether AI can outperform the best humans in every task. It is whether AI can make expert-level guidance widely available enough that ordinary people can perform substantially better than they otherwise would.

In education, this could mean:

The same logic extends beyond education. If AI systems can reliably transfer practical expertise in real time, then societies may be able to expand access to high-quality professional judgement without expanding the supply of top experts at the same rate.

Limits and unanswered questions

The Tutor CoPilot findings are promising, but they do not prove that expertise can always be scaled this way.

Several important limitations remain.

First, the gains were meaningful but not revolutionary. Average improvements were measured in percentage points, not orders of magnitude.[arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-TimeTutor CoPilot: A Human-AI Approach for Scaling Real-Time…October 3, 2024 — by RE Wang · 2024 · Cited by 104 — We introduce Tutor…Published: October 3, 2024

Second, the study focused on a specific environment: online mathematics tutoring for K–12 students. Results may differ in subjects that require deeper discussion, emotional support or more complex forms of judgement.[arXiv]arxiv.orgarXiv Tutor Co Pilot: A Human-AI Approach for Scaling Real-TimeTutor CoPilot: A Human-AI Approach for Scaling Real-Time…October 3, 2024 — by RE Wang · 2024 · Cited by 104 — We introduce Tutor…Published: October 3, 2024

Third, there is a risk of overreliance. If tutors begin following AI suggestions mechanically, some forms of professional growth could weaken rather than strengthen.

Finally, the study does not eliminate the need for expert humans. The system itself was built around models of expert tutoring behaviour. Someone still has to generate, evaluate and refine the practices that the AI helps distribute.

These caveats matter because AI bloom is not simply a story about more intelligence. It is a story about how intelligence is shared, governed and applied.

A small example of a larger possibility

Tutor CoPilot offers an unusually concrete example of a larger claim often made about advanced AI: that one of its most important effects may be the democratisation of expertise.

The study did not show AI replacing teachers. It showed AI helping weaker teachers make better decisions. The strongest result was not superhuman performance. It was the compression of a skill gap.[National Student Support Accelerator]nssa.stanford.edututor copilot human ai approach scaling real time expertiseNational Student Support AcceleratorTutor CoPilot: A Human-AI Approach for Scaling Real-Time…15 Dec 2024 — With the help of the tool…

For advocates of a future of AI-enabled abundance, that distinction matters. Human flourishing depends not only on creating extraordinary intelligence but on making useful intelligence broadly available.

If future AI systems can help millions of people perform closer to expert level across education, healthcare, science and public services, then the long-term impact could be far larger than the immediate tutoring gains reported in one study. Tutor CoPilot is therefore interesting not mainly because it improved mathematics tutoring, but because it offers an early glimpse of how scarce expertise itself might become more abundant.

Amazon book picks

Further Reading

Books and field guides related to Why weaker tutors gained most from AI help. Use these as the next step if you want deeper reading beyond the article.

Book

The AI Classroom

By Dan Fitzpatrick, Amanda Fox et al.

Explores AI support for teaching and coaching.

See on Amazon

Book

Co-Intelligence

By Ethan Mollick

Matches the theme of AI helping ordinary people perform closer to experts.

See on Amazon

Book

Peak

By Anders Ericsson, Robert Pool

Explains how performance improves through guidance and feedback.

See on Amazon

Book

Teaching with AI

By José Antonio Bowen, C. Edward Watson

Shows how AI can raise educator effectiveness.

See on Amazon

Browse more on Amazon:The AI ClassroomCo IntelligencePeak

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Example marketplace items related to this page. Use the search link to explore similar finds on eBay.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Example eBay listing

Teacher Educational Poster Set Of 6 Classroom English Fun LAMINATED 30cm X 42cm

SearcheBay.co.uk: teacher poster

Browse similar oneBay.co.uk

Example eBay listing

Guitar Theory Poster Print – Circle of Fifths Music Teacher Wall Art

SearcheBay.co.uk: teacher poster

Browse similar oneBay.co.uk

Example eBay listing

Daycare Teacher Quote Framed Wall Art Poster Canvas Print Picture

SearcheBay.co.uk: teacher poster

Browse similar oneBay.co.uk

Example eBay listing

Teacher Them Love Them Watch Them G Framed Wall Art Poster Canvas Print Picture

SearcheBay.co.uk: teacher poster

Browse similar oneBay.co.uk

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. 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

2. Source: scale.stanford.edu
Title: how ai can improve tutor effectiveness
Link:https://scale.stanford.edu/news/how-ai-can-improve-tutor-effectiveness

3. Source: arxiv.org
Title: arXiv [Tutor Co Pilot]({{ ‘tutor-co-pilot/’ | relative_url }}): A Human-AI Approach for Scaling Real-Time
Link:https://arxiv.org/abs/2410.03017

Published: October 3, 2024

4. Source: edunlp.stanford.edu
Title: tutor copilot
Link:https://edunlp.stanford.edu/projects/tutor-copilot

5. Source: arxiv.org
Link:https://arxiv.org/pdf/2410.03017

6. Source: hpttreasures.wordpress.com
Title: Tutor Co Pilot: A Human-AI Approach for Scaling Real-Time
Link:https://hpttreasures.wordpress.com/wp-content/uploads/2025/11/cta-wang-et-al-2024-tutor-copilot-study-1.pdf

Published: October 7, 2024

7. Source: nssa.stanford.edu
Title: research notes two emerging strategies using ai tutoring
Link:https://nssa.stanford.edu/news/research-notes-two-emerging-strategies-using-ai-tutoring

8. Source: the74million.org
Link:https://www.the74million.org/article/study-ai-assisted-tutoring-boosts-students-math-skills/

9. Source: nssa.stanford.edu
Link:https://nssa.stanford.edu/news/study-ai-assisted-tutoring-boosts-students-math-skills

10. Source: aicerts.ai
Link:https://www.aicerts.ai/news/how-tutor-co-pilot-systems-scale-teaching-capacity-worldwide/

11. Source: scale.stanford.edu
Link:https://scale.stanford.edu/news?page=6

12. Source: scale.stanford.edu
Link:https://scale.stanford.edu/publications/tutor-copilot-human-ai-approach-scaling-real-time-expertise

13. Source: tutor.com
Link:https://www.tutor.com/

14. Source: beibindex.wordpress.com
Link:https://beibindex.wordpress.com/2024/10/

Additional References

15. Source: overdeck.org
Link:https://overdeck.org/research-repository/tutoring/tutor-copilot-a-human-ai-approach-for-scaling-real-time-expertise/

16. Source: reddit.com
Link:https://www.reddit.com/r/machinelearningnews/comments/1fz9cil/researchers_at_stanford_university_introduce/

17. Source: medium.com
Link:https://medium.com/syncedreview/sandford-us-tutor-copilot-transforms-real-time-tutoring-with-ai-driven-expert-guidance-b2a7cf5d5c18

18. Source: linkedin.com
Link:https://www.linkedin.com/pulse/stanford-tutor-copilot-human-ai-approach-scaling-real-time-expertise-bx7me

19. Source: dorademszky.com
Link:https://www.dorademszky.com/publications/39308-tutor-copilot-a-human-ai-approach-for-scaling-real-time-expertise

20. Source: linkedin.com
Link:https://www.linkedin.com/posts/ilkka-tuomi-30086_this-is-an-interesting-and-good-quality-study-activity-7249305278234476544-tf0A

21. Source: overdeck.org
Link:https://overdeck.org/portfolios/spotlight/nssa-tutor-copilot-a-human-ai-approach-for-scaling-real-time-expertise/

22. Source: syncedreview.com
Link:https://syncedreview.com/2024/11/15/self-evolving-prompts-redefining-ai-alignment-with-deepmind-chicago-us-eva-framework-3/

23. Source: researchgate.net
Title: 384680722 Tutor CoPilot A Human AI Approach for Scaling Real Time Expertise
Link:https://www.researchgate.net/publication/384680722_Tutor_CoPilot_A_Human-AI_Approach_for_Scaling_Real-Time_Expertise

24. 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