Dual Explanation (E+/−): how not to fall for a convincingly wrong AI answer

Can be done in one request or two separate ones

STEP 1: Get the AI's answer to your question

STEP 2: Ask the AI for a dual explanation →
  — Arguments FOR the answer being correct
  — Arguments AGAINST (why the answer might be wrong)

STEP 3: Read both blocks → make the call yourself

It all happens in one chat. No need to switch models or tools.

Task: Ethan is a founder who uses ChatGPT for legal questions about a contract with an investor. The model confidently explains that the anti-dilution clause protects his interests. The explanation is long and sounds professional. But Ethan isn't sure — should he trust it?

Prompt:

You just explained to me that the anti-dilution clause in the contract protects my interests 
as a founder.

Now do the following:

**Arguments FOR** — why your conclusion is right and this clause really protects me. 
Give 3-4 specific arguments.

**Arguments AGAINST** — why your conclusion might be wrong or incomplete. 
Where exactly could you have made a mistake? What nuances did you miss? What conditions make 
this clause useless or even dangerous for me? Give 3-4 specific arguments.

Don't defend your original answer. The arguments against must be just as 
strong as the arguments for.

Result: The model will produce two equally weighted blocks. The "against" block will surface things it never mentioned on its own: for instance, that "weighted average" vs "full ratchet" wording changes the meaning completely, or that the clause only kicks in under certain financing conditions. After reading both blocks, Ethan will spot concrete risks — and go to a lawyer with the right questions instead of false confidence.

The LLM's weakness: By default, the model generates coherent, persuasive text. When you ask it to explain an answer, it explains it in its own favor. That's not deception — it's just how generation works: the model continues the narrative that started with the answer. How convincing an explanation sounds has no correlation with whether the answer is right.

The LLM's strength: The model can simulate roles and perspectives. Explicitly ask it to "find arguments against," and it switches into critic mode and generates text from a different starting point. It doesn't "change its mind," but it starts from a different position.

How the method uses this: Dual explanation forces the model to start twice, from two different positions. The "for" arguments are its usual justification mode. The "against" arguments are a forced search for weak spots. When you read both blocks, you get a real picture instead of a one-sided defense. The illusion of certainty disappears.

Control knobs: - Explicitly demand equal strength — add "the arguments against must be no weaker than the arguments for." Otherwise the model offers limp objections. - Number of arguments — 3-4 per side is the sweet spot. Fewer is superficial; more is the same ideas in different words. - Add "what might you have missed?" — a separate trigger for blind spots that won't show up in standard "arguments against." - Level of skepticism — for maximum criticism, add a role: "Answer as a skeptical expert who disagrees with this conclusion from the start."

You just answered my question: {your question}.

Your answer was: {model answer}.

Now give a dual explanation:

**Arguments FOR** (why your answer is correct):
— Give {number} specific arguments supporting your answer.

**Arguments AGAINST** (why your answer might be wrong):
— Give {number} specific arguments. Where could you have made a mistake? 
  What did you miss? Under what conditions is your answer wrong?

The arguments against must be just as strong as the arguments for. 
Don't defend your original answer.

Placeholders: - {your question} — the original question you asked - {model answer} — what the model replied - {number} — usually 3-4; 2 is fine for simple tasks

🚀 Quick start — paste into the chat:

Here's a template for a dual explanation. Adapt it to my task: {your task}. 
Ask me questions to fill in the fields.

[paste the template above]

The LLM will ask what exactly you're checking and which answer you want to challenge — because without a specific answer and question, it can't produce pointed criticism.

⚠️ Familiar territory: If you know the topic well, dual explanation adds less. The study showed that when users know the answer themselves, a plain reasoning summary is enough. Dual explanation is most useful exactly when you're unsure.

⚠️ Doesn't work for facts with one clear answer: "How many days are in February?" — a symmetric debate is artificial here. The method is for tasks with real uncertainty or expert judgment.

⚠️ The model may "cave" to its first answer: The "against" arguments sometimes come out limp — the model subconsciously defends what it already said. A firm instruction helps: "the arguments against must be no weaker than the arguments for," plus a separate request to play the skeptic.

⚠️ It's not a substitute for verification: The method reduces false trust but doesn't eliminate errors. If the answer really matters, professional review is a must. The model generates persuasive arguments for both sides — not the truth.

Researchers at Arizona State University ran a clever experiment: how do you test whether AI explanations help if you remove people's ability to check the answer themselves? They took problems from IIT JEE-Advanced — one of the toughest entrance exams in the world in physics, chemistry and math. The participants were recent high school graduates: smart enough to read the explanations meaningfully, but not experienced enough to solve the problems on their own. A perfect test bench.

Each participant saw 8 problems with DeepSeek R1's answers, half of which were deliberately wrong. That matters: if the AI is always right, there's no point learning to double-check it. Participants were split into 5 groups by explanation type: answer only, full chain of thought, short reasoning summary (like GPT/Claude), a "for" justification, and the dual "for and against" explanation. 25 people per group, plus a cash bonus for correct judgments — so people actually tried instead of clicking at random.

The most surprising result: being able to see the model's reasoning didn't help catch the mistake. Users with the reasoning trace were wrong almost as often as those who saw just the answer. Dual explanation, on the other hand, produced a statistically significant drop in false trust — and it was the only condition where people both accepted correct answers better and rejected wrong ones better.

The researchers tested several generation templates. Here's the prompt logic for the E+/− condition (dual explanation):

Generate a post-hoc explanation for the following AI-generated answer 
to a {subject} question.

Question: {question}
AI Answer: {answer}

Provide:
1. Arguments FOR why this answer is correct (merits)
2. Arguments AGAINST why this answer might be incorrect (demerits)

Present both sides with equal rigor. Do not favor the AI's answer.

Context: Generation was done with DeepSeek V3 on DeepSeek R1's answers to JEE-Bench problems. Participants were shown the final block with two columns of arguments.

💡 Adaptation for checking business decisions

Dual explanation is especially powerful for irreversible decisions: hiring/firing, signing a contract, launching an ad campaign, choosing a contractor. By default, AI helps you execute what you've already decided — not challenge it.

I'm about to {decision}. I'm already leaning toward it.

Give a dual explanation:

**5 strong arguments FOR** this decision — why it's the right call.

**5 strong arguments AGAINST** — why I might regret it. 
Which risks aren't obvious? What am I most likely underestimating? 
Under what conditions will this decision turn out to be a mistake?

The arguments against must be just as convincing 
as the arguments for. Imagine you're the opposing counsel.

🔧 Technique: add "blind spots" → find what the model left unsaid

Standard "arguments against" usually cover the obvious risks. To dig out the non-obvious:

...and add a separate section:

**What I probably didn't ask but should know** — 
which aspects of this question did you skip because I didn't ask,
but which could change the answer?

This is a trigger for known unknowns — things the model knows but left out of the answer because you didn't ask.

🔧 Technique: a critic persona → sharper arguments against

A faceless "arguments against" yields soft criticism. A specific persona makes it sharper:

Give the arguments AGAINST in the voice of [Mark Cuban / a seasoned lawyer 
who has seen 100 contracts like this / a venture investor 
who has already been burned this exact way] — someone who distrusts 
this answer from the start.

The model simulates the role — and the criticism gets more specific and tougher.

Paper: Evaluating the False Trust engendered by LLM Explanations

Authors: Vardhan Palod, Upasana Biswas, Subbarao Kambhampati — School of Computing & AI, Arizona State University

Dataset: JEE-Bench — IIT JEE-Advanced problems in math, physics, chemistry

Related ideas: Contrastive explanation generation [22]; xAI trust calibration (Bansal et al., Schemmer et al., Zhang et al.)