TL;DR
When you write with a single AI, it pulls you toward the middle of the road: your ideas start to look like the ideas of every other user of that same AI. This isn't a metaphor — the study measured the effect statistically. The fix is simple: use two different AIs at once. Claude plus ChatGPT gives you the same quality, but idea diversity recovers to the "no AI" level.
The second key insight is about learning. The "tutor + mistaken peers" combo produces better results than a tutor alone or simply being handed the right answer. When someone nearby makes understandable mistakes and you analyze why they're wrong, you remember more deeply. A lone AI tutor doesn't produce this effect, because it's always right.
The study tests two scenarios: convergent tasks (math with one correct answer) and divergent tasks (writing, idea generation). In both cases the multi-agent setup beats a single AI — for different reasons and through different mechanics.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
Two independent applications for different tasks:
Scenario A: Learning and problem-solving
STEP 1: You → give your answer to the problem
STEP 2: Peer agent A → gives an answer with an arithmetic error
STEP 3: Peer agent B → gives an answer with a conceptual error
STEP 4: Tutor agent → discusses, asks questions (doesn't give the answer directly)
STEP 5: You → analyze the peers' mistakes, correct your understanding
→ All in one conversation, the AI plays every role
Scenario B: Writing and idea generation
STEP 1: Claude → writes/brainstorms on the topic [separate chat]
STEP 2: ChatGPT → writes/brainstorms on the same topic [separate chat]
STEP 3: You → compare, take the best from both
→ Two parallel conversations in different interfaces
Example
Task: Rachel is launching a home-cooked meal subscription service in Denver. She needs to write a strong investor pitch, but she's worried her ideas will sound "like everyone else's."
Prompt (for scenario A — learning how pitches work):
I'm learning to write investor pitches. Play three roles at once.
My idea: a delivery service for home-cooked meals from local home cooks in Denver.
Role 1 — Mia (a peer with a conceptual error):
Give pitch advice that sounds logical
but is built on a wrong understanding of
what investors look for at the early stage.
Role 2 — Dylan (a peer with a factual error):
Give advice with the right understanding of the concept,
but with wrong market data or bad arithmetic.
Role 3 — Mentor Ben:
Don't give a direct answer. Ask me 2 questions
that will help me find Mia's and Dylan's mistakes myself.
After I answer, correct my understanding
if I missed anything.
For scenario B (idea diversity) — two parallel requests:
In Claude:
Come up with 5 non-obvious value propositions for a pitch for a
delivery service for home-cooked meals from local home cooks.
Audience — private investors in Denver.
Avoid the obvious: quality, homemade, natural.
In ChatGPT — the exact same request, word for word.
Result: - In scenario A: the model shows two "wrong" versions of the advice, then the tutor's questions. You hunt for the mistakes yourself — and that's exactly what builds understanding. After your answer, you get the breakdown. - In scenario B: Claude and ChatGPT give noticeably different ideas. Not because one is better — they have different "averaging patterns" from their training data. Combining both lists gives you a wider range.
Why It Works
A single AI always answers correctly — and that's its weakness for learning. When you're handed the right answer straight away, your brain doesn't activate what it already knows and doesn't notice the gaps. Psychologists call this "productive failure": when you first get it wrong or see someone else's mistake, understanding goes deeper. An LLM tuned to always answer correctly removes exactly the moment of tension that creates memory.
A single AI makes everyone's texts the same. Every model is trained on a particular data corpus and averages its patterns. If thousands of people ask ChatGPT to write a pitch, they all get similar structures and arguments. That's not bad — it's predictable. Claude averages differently. When you take both answers, the range of ideas is wider, because the two "centers of gravity" are shifted in different directions.
Watching others' mistakes is a separate learning mechanism. When you see "Mia" applying a formula correctly but misunderstanding the problem, you think: "Wait, am I doing the same thing?" That triggers a self-check that a plain correct answer never does. The researchers deliberately split mistakes into two types: conceptual (wrong understanding) and arithmetic (right understanding, wrong calculation) — because each type teaches something different.
Control levers: - Peer error type → adapt it to the domain: for writing — "right structure, weak arguments" vs "strong arguments, weak structure" - Tutor style → "ask only questions, give no answers" vs "give step-by-step hints" — different depth of independent work - Number of models → for idea generation, more than two wasn't tested, but the principle is the same
Prompt Template
Template A — Tutor + Peers with mistakes (for learning)
I'm studying {topic}. Play three roles at once.
My problem / question: {describe the problem or question}
My current answer: {write your answer or "haven't thought about it yet"}
Role 1 — {peer_A_name} (understands the concept
but makes mistakes in calculations/details):
Give your answer to the problem with this error profile.
Role 2 — {peer_B_name} (calculates/executes the steps correctly
but misunderstands the essence/concept):
Give your answer to the problem with this error profile.
Role 3 — {tutor_name} (a mentor who never gives answers directly):
Ask me 2-3 questions that will help me find the mistakes
in {peer_A_name}'s and {peer_B_name}'s answers.
Wait for my answer. Then give a breakdown:
where I was right and what I missed.
What to fill in:
- {topic} — what you're studying: finance, programming, negotiation, marketing
- {problem} — a specific question or case to work through
- {names} — any names; they give the agents clear-cut roles
- The error descriptions can be adapted to the domain: for business — "right logic, wrong market assumptions" vs "right data, wrong conclusion"
Template B — Model duo (for writing and ideas)
[Run in parallel in Claude and ChatGPT — the same prompt]
I need to {what to create}: {briefly describe the task}.
Audience: {who the reader / listener is}
Context: {key details of the situation}
Goal: {what the reader should do after reading}
Constraints: {what definitely should NOT be included or which approach does NOT work}
Give me {5 options / a first draft / a list of ideas}.
Don't explain your choices — just the result.
How to work with the two answers: 1. Copy both answers into a third chat (any AI) 2. Ask: "Here are two versions. Synthesize the best of both and remove duplicates"
🚀 Quick start — paste into the chat:
Here is a template for learning through peers who make mistakes.
Adapt it to my task: {your task}.
Ask me questions to fill in all the fields.
[paste template A above]
The LLM will ask about the topic, the specific problem and your current level of understanding, because without that it can't tune believable mistakes to the right difficulty level. Then copy the finished prompt and start the session.
Limitations
⚠️ Peers without a tutor are weaker: "Mistaken peers" alone, without a mentor, give a gain, but a much smaller one. Don't drop the tutor role from the template — it's the combination that works.
⚠️ Confidence ≠ results: Participants in the "peers only" group felt more confident but objectively knew less than those in the tutor group. Feeling good after a lesson is no guarantee you learned anything.
⚠️ For divergent tasks — two AIs, not one: If you ask Claude to generate "different perspectives" in a single chat, the range will still be narrower than when you actually use two models. Homogeneity is a property of the model, not of how many times you ask it.
⚠️ Simple tasks aren't for this method: For a quick reference question ("what's the difference between X and Y?") the multi-agent setup is overkill. The method pays off on complex, multi-step problems with non-obvious answers.
⚠️ The peer confidence effect: The study found that participants penalized "Alice" (arithmetic errors) more harshly than "Charlie" (conceptual errors), even though conceptual errors are more serious. Arithmetic errors are more visible; conceptual ones are more dangerous. Keep this in mind: if a conceptual error has crept into a peer's text or answer, it's easier to miss.
How It Was Studied
A team from the University of Toronto ran two controlled experiments with real people via Prolific, a crowdsourcing platform.
Experiment 1 (math, N=315): Participants solved SAT-level problems under four conditions — no AI, peers only, tutor only, tutor + peers. A curious design detail: after the lesson everyone played Tetris for a minute to "clear" short-term memory, then solved similar problems with no help at all. The results rise monotonically: control ~42% → peers ~48% → tutor ~59% → tutor + peers ~65%. Adding peers to the tutor gave a statistically significant gain, although peers alone without a tutor did not.
Experiment 2 (writing, N=247): Participants wrote argumentative and creative essays under three conditions — no AI, one AI (ChatGPT or Claude), and two AIs at once (Claude + ChatGPT). Both AI conditions improved text quality compared with "no AI." But here's the surprise: idea homogeneity — with a single AI, participants produced similar arguments. Two AIs neutralized this: quality stayed high, and idea diversity returned to the baseline "no AI" level.
An unexpected finding: confidence doesn't correlate with results. The "peers only" group felt that things were easier and that they'd done well, but the test showed they learned less than the tutor group. The feeling of understanding and actual understanding are different things, and peers without a mentor create a false sense of confidence.
Original from the Paper
Description of the peer agents from Experiment 1 (system prompts):
Bob (Tutor). Bob acts as a supportive math tutor who guides the
discussion and provides feedback on the participant's answers.
Rather than supplying solutions directly, Bob responds with
clarifications, hints, and encouragement (consistent with
Socratic tutoring practices shown to promote deeper reasoning).
Alice (Arithmetic-Error Peer). Alice has a strong conceptual
understanding of the material but frequently makes arithmetic
mistakes (mixing up numbers or making calculation errors). She
represents a common student profile where the underlying
reasoning is sound, but execution falters.
Charlie (Conceptual-Error Peer). Charlie computes things
accurately but sometimes misunderstands the underlying concepts,
applying formulas incorrectly despite doing the math correctly.
He represents the complementary error profile: procedurally
fluent but conceptually fragile.
Context: The researchers built three GPT agents with separate system prompts. The tutor uses a Socratic style (gives no answers, asks questions). The two peers have deliberately different error profiles that reflect the two most common types of student mistakes in math.
Adaptations and Extrapolations
💡 Adaptation: mistaken peers for professional skills
The same setup works outside math — for any field that has a "right" and a "wrong" approach:
I'm learning to write sales proposals.
Situation: The client is a mid-size restaurant in Portland
that wants to order social media marketing.
Play three roles:
Kate (strong on structure, weak on pricing):
Write a proposal with the right structure and logic
but a flawed pricing strategy —
too low, not covering all the costs.
Pete (gets the money right but doesn't understand the client):
Write a proposal with correct numbers and margins
but no understanding of the restaurant's real pain points.
Mentor Grace:
Don't write an alternative. Ask me 2 questions
that will help me find the weak spots in both proposals.
🔧 Technique: names instead of roles → sharper execution
By default: "act as a critic" — vague.
With agent names + a description of their character: "Adam harshly criticizes weak data" — more precise.
The study showed that participants rated the agents differently depending on their names and described traits, which affected how they interacted with each one. Give your agents names and one key trait, and the AI plays the role more precisely.
[Instead of "Play the role of a skeptic"]
↓
[Play the role of Michael — an investor who doesn't
believe a single number without a source. Michael
interrupts, demands evidence, isn't polite]
🔄 Extrapolation: combo with Chain-of-Thought
Combine the multi-agent setup with Chain-of-Thought (step-by-step reasoning):
Three agents analyze my business plan.
Each one thinks out loud, step by step.
Optimist Sam: looks for strengths →
thinks out loud → states a conclusion
Pessimist Irene: looks for risks →
thinks out loud → states a conclusion
Realist Andrew: listens to both →
thinks out loud → proposes concrete edits
[My plan: ...]
This strengthens both methods: diversity of perspectives + transparent reasoning.
Resources
Paper title: Beyond the AI Tutor: Social Learning with LLM Agents
Authors: Harsh Kumar, Jonathan Vincentius, Zi Kang (Jace) Mu, Ashton Anderson
Institution: University of Toronto, Department of Computer Science
Status: Working Paper, March 2026
Theoretical foundation: Vygotsky (Zone of Proximal Development), Bandura (Social Learning Theory), Kapur (Productive Failure), Bjork (Desirable Difficulties)
Services mentioned in the study: KhanMigo, Claude for Education, ChatGPT Edu