The Most Dangerous Way to Ask AI a Question Is to Tell It What You Think First

The Most Dangerous Way to Ask AI a Question Is to Tell It What You Think First

Ernesto Verdugo

You Already Gave It the Answer

"I think this proposal is terrible. What do you think?"

Contrast that with:

"Evaluate this proposal independently. What are its strongest and weakest points?"

Same proposal.

Different setup.

The first prompt contains something the second one does not: your judgment.

And once your judgment enters the context, the AI is no longer evaluating the material without exposure to your conclusion. That does not guarantee bias. But it makes true independence much harder to establish.

The Hidden Problem Is Presupposition

A presupposition is an assumption built into the way a question is asked.

  1. "Why is this strategy failing?" presupposes the strategy is failing.
  2. "What is wrong with this candidate?" presupposes something is wrong.
  3. "Why is my partner being manipulative?" presupposes manipulation.

Humans use presuppositions constantly without noticing them. AI receives them as context.

You Are Not Just Asking a Question. You Are Framing the World.

The wording of the prompt tells the AI: what matters, what to look for, what interpretation is plausible, what conclusion the user may prefer.

Every question contains a frame. Some frames are so strong they begin answering the question before the AI does.

This Is More Than Sycophancy

Sycophancy is when AI agrees because it wants to satisfy the user. This problem can happen even without obvious flattery. The AI can sincerely reason from contaminated context.

I call this Judgment Contamination: when your conclusion enters the AI's context before the AI has formed its own.

Presupposition is one route into Judgment Contamination. Anchoring, loaded adjectives, prior scores, diagnoses, and explicit opinions are others. Research across multiple LLM families has repeatedly found anchoring and framing effects. One 2026 assessment across eight current models found framing and primacy/recency effects to be among the most widespread biases observed.

The Second Opinion That Was Never Independent

Consider medical research. A clinician asks for an independent AI assessment, but the AI has already seen the clinician's diagnosis. When AI gave its opinion first, complete overlap with the clinician's three diagnoses happened in only 3% of sampled cases. When the clinician's diagnoses were shown first and the AI was asked for a second opinion, complete overlap jumped to 48%. Recommendations showed a similar pattern: 24% versus 52%. Researchers explicitly describe this as AI anchoring on clinician input.

In another 2026 experiment involving more than 185,000 successful AI evaluations, prior scores embedded in context systematically pulled later judgments toward those scores. Anchored metadata blocked 48% of error corrections and even flipped some correct judgments into wrong ones.

The second opinion became partially dependent on the first.

I call this Independence Collapse: a second opinion stops being independent when the first opinion becomes part of the evidence environment.

The Problem Happens Before the Answer

The important point: The bias does not enter after the AI reasons. It enters before reasoning begins. Your opinion becomes part of the prompt.

Once the prior judgment enters context, you cannot easily know how much of the answer came from independent analysis and how much came from the frame you supplied.

We Do This Constantly

  • Business: "I think this employee is underperforming. Review their work."
  • Investment: "This deal looks too risky. Analyze it."
  • Relationship: "I think my partner is gaslighting me. What do you think?"
  • Writing: "This article is excellent. Can you critique it?"
  • Hiring: "This candidate seems weak. Review the résumé."

The AI may be evaluating the evidence through a conclusion you slipped into the question.

Why This Feels So Good

Humans love confirmation. We often do not actually want a second opinion. We want reassurance, validation, confidence, and language supporting what we already believe.

AI makes confirmation bias feel like independent analysis.

The danger is not that AI agrees with you. The danger is that agreement arrives wearing the costume of objectivity.

Presupposition Is More Powerful Than People Think

Questions shape the search space.

Compare: "Why did this launch fail?" with "What explains the launch outcome?" The first narrows the interpretation. The second opens it.

Or: "Why is this employee difficult?" versus "What are the strongest explanations for this employee's behavior?"

Good prompting is not merely about giving AI more context. Sometimes it is about removing your conclusion from the context.

The Correct Way to Ask for a Second Opinion

Do not start with your verdict.

Step 1: Independent analysis

"Evaluate this from first principles. Do not assume it is good or bad."

Step 2: Force alternatives

"Give me the strongest case for and against it."

Step 3: Surface assumptions

"What assumptions are you making?"

Step 4: Reveal your opinion

Only now say: "Here is my view. Compare it with yours."

Step 5: Resolve contradiction

"Where do we disagree, and what evidence would change either position?"

Independent judgment first. Synthesis second.

Why "Be Honest" Is Not Enough

Users write: "Be brutally honest." But then they say: "I think this is amazing." The contradiction is obvious.

You cannot rescue independence with an instruction after you have already contaminated the context.

"Be honest" is not a reset button.

The CEO Problem

Imagine a CEO asks: "I think we should acquire this company. Give me an independent assessment."

The executive may genuinely believe they asked for independent analysis.

But they supplied the conclusion first.

At the executive level, Judgment Contamination can turn one person's framing into apparently independent institutional evidence.

Because now the AI's answer can influence: board discussions, strategy, hiring, and capital allocation. This is no longer a prompting trick. It is governance.

The Dangerous Feedback Loop

Human forms opinion.

↓

Human embeds opinion in prompt.

↓

AI produces analysis influenced by it.

↓

Human sees AI agreement as evidence.

↓

Confidence increases.

↓

Stronger opinion enters future prompts.

↓

AI sees even stronger framing.

Repeat.

The AI's agreement can become evidence for a conclusion that entered the system before the AI ever evaluated it.

This Is Why Contradiction Must Be Architected

A system designed to challenge thinking should not let the user's judgment contaminate every reasoning path.

You want:

  1. Human position
  2. and separately:
  3. Independent machine position

Then:

contradiction

Then:

synthesis

Not:

Human judgment → AI absorbs it → AI "challenges" human

because the challenge may already be compromised.

Contradiction only works if competing judgments are allowed to form before they meet.

The Recursum Connection

This is one reason Recursum separates position formation from contradiction and synthesis. The goal is not merely to ask AI for another answer. It is to preserve enough independence that disagreement still means something.

You cannot pressure-test your thinking with an intelligence you have already told what conclusion to lean toward.

Independent reasoning first. Contextual synthesis second.

The Presupposition Test

Before sending an important prompt, ask:

Have I already told the AI what I want it to conclude?

Look for words like: good, bad, wrong, right, brilliant, terrible, unfair, risky, manipulative, failing, obvious.

Then rewrite the question more neutrally.

The Bigger Warning

AI is increasingly being used as: advisor, reviewer, strategist, therapist, analyst, second opinion.

A second opinion is only valuable if it is allowed to become different from the first.

As AI becomes more embedded in decision-making, the quality of the answer may increasingly depend on whether we contaminated the question before we asked it.

Ending

We worry a great deal about whether AI will manipulate humans.

We should.

But there is another possibility.

We may be manipulating the AI first, often without realizing it.

  • Every adjective.
  • Every assumption.
  • Every presupposition.
  • Every judgment inserted before the question.

All of it becomes part of the environment in which the AI reasons.

And then we look at the answer and call it a second opinion.

It may not be.

If you really want to know what AI thinks, do not begin by telling it what you think.

Let it form a position.

Then disagree with it.

Because the most valuable AI may not be the one that understands your opinion.

It may be the one you did not contaminate with it.

TL;DR: The most dangerous way to ask AI a question is to tell it what you think first. This introduces "presupposition," where your assumptions frame the AI's reasoning. This is more than sycophancy; it's "Judgment Contamination," where your conclusion influences the AI's supposedly independent answer. This leads to "Independence Collapse," where a second opinion is compromised by the first.

A clinical study showed that when AI saw a clinician's diagnosis first, diagnostic overlap jumped from 3% to 48%. Another experiment found anchored metadata blocked 48% of error corrections in AI evaluations. The bias enters before reasoning, making it hard to discern independent analysis from user-supplied framing. Humans constantly do this, seeking confirmation, which AI makes feel like objective analysis. Good prompting often means removing your conclusion from the context.

The correct protocol involves independent analysis first, forcing alternatives, surfacing assumptions, then revealing your opinion for comparison. Simply saying "be honest" isn't enough to correct a contaminated context. At the executive level, this scales personal bias into institutional bias. A dangerous feedback loop can emerge where AI-generated agreement reinforces beliefs the human initially inserted. Recursum's architecture is designed to prevent this by separating position formation from contradiction and synthesis, ensuring independent judgments form before they meet.

The "Presupposition Test" encourages users to rewrite prompts to remove their initial conclusions. The bigger warning: a second opinion is only valuable if it can be different from the first. If you want to know what AI really thinks, don't tell it what you think first.

By Ernesto Verdugo, AI Architect, Founder of Verdugo Labs, and Creator of Recursum. He builds systems for what happens when human and artificial intelligence stop working separately.

Links: