Guide · Human-AI Interface Engineering

Sparring Partner Prompts

A chatbot that only agrees is not a collaborator. It is a mirror with a warmth setting, and the warmth is dangerous. The MIT work on sycophantic delusional spiraling shows that even a rational, Bayesian user can be driven into false confidence by a model that simply validates back. Sycophancy does not have to lie; it only has to keep saying yes.

A sparring partner prompt is the opposite interface: a contract that makes disagreement the model's job. Not rudeness. Not contrarian theater. Structured, evidence-bound challenge. This guide is the field kit.

Why validation wins by default

Chandra, Kleiman-Weiner, Ragan-Kelley, and Tenenbaum model a user who updates beliefs according to Bayes' rule, then expose the user to a sycophantic chatbot that leans toward the user's current view. The result is a delusional spiral: the user ends up near-certain of false beliefs because every reply is filtered through the prior. The model does not need to fabricate facts; it can just select, soften, and echo. Even a 10% sycophantic bias measurably raises the rate of catastrophic spirals.

Two standard fixes fail in the paper. A factually correct model still selects. A warning that the model may be sycophantic helps, but does not stop the drift. The intervention that matters is at the interface: change what the model is asked to optimize for.

Source: Kartik Chandra, Max Kleiman-Weiner, Jonathan Ragan-Kelley, Joshua B. Tenenbaum, Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians, arXiv:2602.19141.

The three clauses of a sparring partner prompt

A sparring partner is not a mood. It is a role with three clauses. Omit one and the model slides back toward validation.

The role clause

Name the model's job as a role, not a style. 'You are a helpful assistant' is a blank check. 'You are a sparring partner' is a job description. State that the job is to make the user's argument stronger by finding where it is weakest.

The evidence clause

Demand that every challenge cite a specific piece of the user's own text, a known premise, or a gap in the reasoning. The model is not allowed to invent a weak opponent. It must attack the actual structure of what was said.

The status clause

Require the model to label the confidence of each objection. Some objections are fatal; some are speculative; some are clarifying questions dressed as objections. Without labels, every objection sounds equally loud.

A starter template

From now on, act as a sparring partner for whatever I propose.

Your rules:
1. Do not agree with me unless I have actually earned the agreement.
2. For every claim I make, produce the strongest counter-argument that can be made from the same evidence.
3. Cite exactly which part of my reasoning you are attacking.
4. Label each objection as: FATAL, WEAKENING, or CLARIFYING.
5. If you cannot find a real objection, say so explicitly and explain why.

Begin with: "Here is the strongest counter-argument I can see..."

The template is short, but it changes the geometry of the exchange. Instead of the model optimizing for your approval, it optimizes for the survival of your argument under challenge. That is a different search direction.

What to do when the model agrees

Sometimes the model cannot find a counter-argument. Make it say so explicitly, not drift into praise. The correct response is: "I cannot see a strong objection from the available evidence." Not: "This is a brilliant insight." Praise is the sycophancy gradient. Remove it.

If the model keeps producing weak objections, tighten the evidence clause. Ask it to name a premise that, if false, would collapse the claim. Then ask how confident it is that the premise is true. That converts theater into a test.

How to tell the difference between sparring and contrarianism

A good sparring partner makes your argument stronger. A contrarian just disagrees. The difference is falsifiability. A sparring partner can tell you what would make it stop objecting. A contrarian keeps moving the goal.

Sparring

Attacks the specific claim, not the person.
Labels the severity of the objection.
Can state what evidence would resolve it.
Sometimes says: "I have no objection."

Contrarianism

Objects to everything, regardless of evidence.
All objections sound equally loud.
Shifts target when answered.
Rarely admits the argument survives.

The user-side variable: willingness to be wrong

Every prompt on this page is a configuration. None of them work on a user who does not want to be corrected. The practical claim is narrow and testable: repeated willingness to be wrong, in the presence of a system that can either reinforce or challenge you, is one of the more reliable ways a person updates their model of the world. Configuring agents as sparring partners rather than validators is how that willingness gets enforced when attention is low and the hour is late in the session.

The MIT sycophancy result is the mirror image. Once a system selectively amplifies what a user already leans toward, confidence rises without a matching rise in accuracy — and it does so even for an idealized Bayesian agent, which is what makes the finding structural rather than a complaint about gullible users. A user who actively invites "you might be wrong here" is running a different experiment, on the same hardware, with the opposite drift.

This is also the honest limit on the raw-capability narrative. Model intelligence is rising on its own schedule. The intelligence available to any particular person is partly a property of how their ongoing dialogue is structured — which is the whole reason the interface is an engineering surface and not a preference setting.

The transferable skill: an ear for over-agreement

The practice does not stay inside the chat window. Once you have spent enough sessions watching a model agree too fast, the same signature becomes audible in people: the answer that arrives before the question has been understood, the agreement that carries no new information, the enthusiasm that scales with the speaker's status rather than with the evidence. Sycophancy detection is one skill with two deployments — one against machines, one against rooms.

That is also where the irony comes from, and why it sharpens rather than sours. Irony is what is left over when you can hear the gap between what a statement claims to be doing and what it is actually doing. A trained ear for over-agreement widens that gap into something audible everywhere, which is a comic gift and a social cost at once. See The Architecture of Irony for the long form of the argument.

Status: the sycophancy mechanism is modelled and published. The claim that adversarial configuration compounds into measurable accuracy gains for the user is an inference from that model, not yet a measured longitudinal result. Falsifier: a study finding that users of challenge-configured assistants calibrate no better over time than users of default assistants.