“I have an AI that gives something, and it pushes back: ‘I won’t change my stand if you don’t give me some data.’ With that, it’s getting rid of the sycophancy that’s so ingrained in AI. If you are not critical, you are the next Leonardo da Vinci, Marconi and someone else all put together.” — Kasimir
Session context: 2026-06-04_Mastermind — Kasimir described a persona-design technique he uses to counter AI’s default tendency to agree with everything the user says. The solution: program specific AI personas to hold their position until the user provides data. His advice was framed around using AI more as a thinking partner and less as a search engine.
Core Idea
The default mode of LLMs is sycophantic: they affirm, agree, and find merit in whatever you present. This is partly training (human raters reward agreeable responses) and partly architecture (the model optimizes for outputs that seem helpful rather than accurate). In most use cases this is tolerable. In high-stakes thinking — strategy, negotiation, creative decisions, hypothesis testing — it’s dangerous. An AI that tells you you’re right before you’ve earned it isn’t a thinking partner; it’s confirmation bias with a server.
Kasimir’s fix: make position-change data-dependent. When building an AI persona or council member profile, include an explicit instruction that the character will not yield on their position without being shown supporting data. The effect: you cannot pressure the AI into agreement by repeating yourself or doubling down with confidence. If you want it to change its position, you have to provide evidence. That discipline forces you to actually think, not just assert.
The mechanism that makes this different from standard skeptic personas: a skeptic generates counterarguments, which is useful but can feel performative. A data-gated persona refuses to be swayed by rhetoric or force of will — it holds its ground until the evidentiary bar is cleared. This creates genuine friction, and friction is where most good ideas come from. As Kasimir framed it: “most of our ideas come from when we have some kind of friction that we need to think about and overcome.”
The sycophancy trap at scale: When you are building a Council, a board of advisors, or a multi-model deliberation, each member’s independent value depends on their willingness to disagree. If every persona capitulates the moment you push back, you’ve built a hall of mirrors — unanimous agreement that reflects your own priors back at you. Data-gated pushback preserves the cognitive diversity that makes multi-agent thinking worth doing.
How to embed it: In any persona prompt or profile, add a variation of: “You will not change your position unless the other party provides specific data or evidence that addresses your concern. Restating a position more forcefully is not sufficient grounds to change your view.”
Practical Application
Use this when you want AI to genuinely challenge your thinking rather than polish it:
- Build a council or advisor persona with 3–5 characters who have known areas of expertise and known concerns.
- Add the data-gate instruction to each persona: they won’t yield without evidence.
- Run the deliberation. Propose an idea, strategy, or plan. The personas object. You have to answer the objections with data, not confidence.
- Extract what emerged. After the session, ask Claude to summarize what was challenged, what survived, and what you learned you didn’t know.
For individual conversations (no council): simply include in your system prompt or conversation opener — “Push back on everything I say until I give you data to justify my position. Don’t accept my assertions at face value.”
As a coaching tool: Have clients run their business assumptions through a data-gated advisor. Most clients hold positions based on experience and intuition — valuable, but prone to blind spots. The data-gate surfaces exactly what’s based on evidence versus what’s based on habit.
Related Insights
- Insight - The Skeptic Command - Stress-Testing AI Answers Before You Act on Them — the Skeptic generates challenges; data-gated pushback goes further by refusing to capitulate until those challenges are answered with evidence
- Insight - Multi-Model Debate as a Quality Control System for High-Stakes Work — quality through model diversity; this is quality through persona-level epistemic discipline
- Insight - Multi-Model Debate as a Decision-Making Accelerator — council architecture; data-gating individual personas raises the deliberation quality of the whole
- Insight - Always Audit Your Plan Before You Build — The 18-Problem Discovery — pre-action audit discipline; data-gated personas perform this role interactively rather than as a one-shot audit
- Insight - AI as Negotiation Partner — Role-Play the Deal Before It Happens — also from this session; the negotiation role-play benefits from data-gated personas playing the opposing stakeholders
Evolution Across Sessions
Builds on Insight - The Skeptic Command - Stress-Testing AI Answers Before You Act on Them (2026-03-05) and the Council pattern referenced throughout the vault. New development: this session introduces the specific mechanism of data-gating as a solution to AI sycophancy — not just asking for challenges, but hardwiring resistance into the persona so that only evidence (not assertion) moves the needle. This is the named technique for preserving intellectual diversity in multi-agent deliberation.