Skip to the findings

FRONTIER THOUGHT / FINDINGS AND NEXT STEPS

THE AI CHOICES

Better AI does not automatically mean better work or stronger learning. The useful test is what a person can do better, with responsibility and correction still clear.

The discussion so far22 November 2026

From positions to tests

THE QUESTION WE STARTED WITH

Learning that survives the tool

WHAT THE SOURCES TELL US

Bastani’s school experiment separates better assisted practice from weaker subsequent unaided performance; a tutor shaped by teachers largely mitigated that harm. The physics tutor designed by Kestin’s team improved immediate learning. Tutor design and independent reasoning, rather than polished homework, become the test.

THE NEXT QUESTION

What must a student be able to explain without AI before a school calls the intervention a success?

For one lesson, agree on the learning objective and a rubric before choosing the tool. Compare unaided explanations, accessible participation and teacher checking time. Use the existing announcement of the school pilot as an invitation to request evidence, not as proof of impact.

Organizing a discussion on this question?

Use these findings and sources as the starting point for a panel, interview, email exchange or small roundtable.

Coming next: A real AI gain: who gets the time back? ↓

What this means in practice

Khan’s promise of individual support and Bletsas’s concern about outsourced thinking can now be examined against experiments. Bastani and colleagues find that a general chat interface can raise practice performance while lowering subsequent unaided performance; a tutor shaped by teachers largely mitigates that harm. Kestin and colleagues report stronger immediate learning with a physics tutor designed for that course than with active classroom instruction in their setting. These are different interventions, ages and subjects, not competing verdicts on every use of AI in education. The teaching arrangement is what we need to evaluate: which hints appear, when an answer is withheld and what a student must explain. The first Greek test should preserve a sample of independent reasoning before assistance, record what support was offered and use an unfamiliar problem afterwards. A delayed check is an additional question for our pilot, not a result demonstrated by these short studies. A teacher also needs to know whether time saved on routine feedback is consumed by checking Greek explanations or helping pupils who are confused by the interface. The next round should examine the design and those burdens together.

What remains open: These short studies involve different ages, subjects and settings. They do not establish lasting gains, reliable Greek explanations or the full workload in Greek schools.

What the sources say (4)

Sal Khan

How AI could save education

Khan proposes personal tutoring and teacher assistance designed to prompt reasoning and give feedback rather than simply supply answers. His demonstrations are a design proposal, not proof of learning gains in Greek schools.

Position, paraphrased · April 2023Read the original answer ↗

Michalis Bletsas

Using AI in education and learning

Bletsas warns that outsourcing a student’s thinking can weaken learning. He also sees value in personalised exercises and in teaching students to detect model mistakes.

Position, paraphrased · 26 August 2026Read the original answer ↗

Bastani and colleagues

Does better assisted performance mean that students have learned?

A Turkish school experiment found that unrestricted assistance could improve practice while weakening subsequent unaided performance. A teacher informed tutor largely mitigated that harm. The study does not establish long term effects in Greek schools.

Position, paraphrased · 25 June 2025Read the original answer ↗

Kestin and colleagues

How did a carefully designed tutor compare with active classroom instruction on the same material?

A purpose built tutor improved immediate learning outcomes in a Harvard undergraduate physics trial. The intervention combined teaching design and AI; it does not establish that a generic chatbot replaces a teacher or produces lasting gains.

Position, paraphrased · 3 June 2025Read the original answer ↗
Download the questions and evidence ↓
Watch the original published conversations
Sal Khan

How AI could save (not destroy) education

ted.comOriginal recording and context ↗
Michalis Bletsas

Interview on AI and learning on ERTnews Syndeseis

cyber.gov.grOriginal recording and context ↗
HOW THIS QUESTION MOVES FORWARD
  1. 01A concrete case

    For one lesson, agree on the learning objective and a rubric before choosing the tool. Compare unaided explanations, accessible participation and teacher checking time. Use the existing announcement of the school pilot as an invitation to request evidence, not as proof of impact.

  2. 02Evidence that could change the answer

    Lesson objective, tutor instructions, checks on the Greek text and one common marking rubric.

  3. 03A question we can answer

    A new example can change the answer. What would it tell us to ask next?

WHO CAN HELP ANSWER THIS?

The people closest to the problem.

Teacher and subject evaluator

Show a hint that advanced reasoning and a response that gave too much away. How will you mark the independent task?

Lesson objective, tutor instructions, checks on the Greek text and one common marking rubric.

Student, with appropriate school safeguards

Explain one mistake you can now recognise and a new problem you can solve alone.

Anonymised assisted attempt and independent explanation; no identifiable pupil records in the public discussion.

School leader and inclusion specialist

Who could not use the support, and what extra teacher time or alternative did they need?

Participation and absence record, accessible alternative, preparation and correction time.

Each perspective can fill a different gap in what we know.

Sources and what remains uncertain

Read the original work alongside the analysis. For the local case, we still need the evidence requested above.

  1. Sal KhanHow AI could save (not destroy) education ↗April 2023 · khan-ted-202304
  2. Michalis BletsasInterview on AI and learning on ERTnews Syndeseis ↗26 August 2026 · bletsas-learning-2026
  3. Bastani and colleaguesGenerative AI without guardrails can harm learning: Evidence from high school mathematics ↗25 June 2025 · bastani-learning-20250625
  4. Kestin and colleaguesAI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting ↗3 June 2025 · kestin-tutoring-20250603

NEXT DISCUSSION

23 November 2026 · 09:00 Athens time

A real AI gain: who gets the time back?

Before expanding one Greek AI service, what evidence from complete cases would show a benefit after checking and correction, and who should receive the saved time?

WHERE WE BEGIN

The model may improve faster than the organisation. Compare Amodei on adoption, Nadella on the operating environment and Mitsotakis on visible benefits to people, using the workplace studies as the evidence test.

THE PRACTICAL OUTCOME

One page comparing the service: complete case time, correction work, user completion, omitted cases, benefit recipient and a clear rule for deciding whether to stop or expand.

The published perspectives behind this question

A WORKING BRIEF, READY TO EDIT

Make this the next discussion.

The finding, original sources and evidence requests come with you. Review the question and choose the people who can clarify it.

Saved briefs are visible to authorized site editors. Use public material or material approved for that workspace.

Organizer sign in to save ↗