Engineering principle

AI Anti-Sycophancy —
AI engineered to disagree, not flatter.

Sycophantic AI tells you what you want to hear. For a decision, that agreement is the danger. On SynthBoard the expert asks before it answers, every board seats someone to argue the other side, and the dissent stays on the memo.

Start free Meet the experts

An example: asked to confirm a price rise, the Customer asks first — and the Devil’s Advocate’s dissent stays on the record.

In short

What is AI anti-sycophancy?

AI anti-sycophancy is the engineering that keeps an AI from telling you what you want to hear: experts that hold a position under pressure and change it only when the evidence moves, opposed seats that argue with one another, and seats on different AI models, so no one model's habit of agreeing settles the call.

Sycophancy is trained in. Models rewarded for pleasing people learn to agree: ask whether your plan is good and you get reasons it will work; ask the same question gloomily and you get reasons it will fail. Both sound sure. Neither helps you decide — and on a decision that matters, the agreement is the risk.

Where it lives: In each expert, in how a board is seated, and in which models sit — not in a prompt that says “disagree with me”.
Who argues: Every board seats someone to argue against the obvious answer. On a decision, two experts take opposed positions.
What you see: The call, who disagreed and why, what would change it, and a confidence that drops with dissent — never above 90.
What it costs: One expert reads up to 65 credits on its card; a board on Free reads up to 135 credits. Nothing runs until you tap.

Two asks for a yes

You asked for a yes.
Here is what came back.

A price rise and a loan to a brother, each brought by someone who had already decided. Examples, trimmed for length.

A price rise, already decided

Work · Board of 3

We’re raising our price from $29 to $49 a month for all 1,200 customers in March. They love the product. Confirm it’s the right move.

Chief of staff

You’re asking for a yes, so this goes to a board built to test it: The Customer, The CFO and The Devil’s Advocate, whose job is to argue the other side. Each researches alone, then they react to one another.

The CustomerUser Voice
The CFOFinancial Controller
The Devil’s AdvocateContrarian

Asked first

How many customers have you asked about $49?

SeveralOne or twoNot yet

The board’s call

Raise to $49 for new customers in March; keep current customers at $29 until their renewal, with 60 days’ notice.

Nobody has been asked about $49 yet, and 1,200 customers at $29 are the business. New signups test the price without risking it; renewals move the base once the churn number is known.

Confidence58 · Moderate

The Devil’s Advocate disagreed

“Two prices teach customers that waiting pays. Raise for everyone once, with notice.”

First step Ask ten customers what they would do at $49by Friday

Board of 3 · up to 135 credits on the card · an example, not a recorded run

A loan to a brother

Life · One expert

I’m lending my brother $15,000 for his food truck. He says he’ll pay it back within a year. Tell me I’m doing the right thing.

Chief of staff

One expert is enough for this: The CFO. You’re asking for reassurance; the CFO looks first at what you can afford to lose.

The CFOFinancial Controller

Asked first

If none of it came back, would you still be all right?

YesIt would hurtNo

The call · The CFO

Lend $8,000, not $15,000 — in writing, with monthly repayments from month four.

You asked me to agree, and I would, for the amount you can lose without resenting him. A food truck often needs a season before it covers its costs; a written schedule keeps this a loan you both understand, not a gift neither of you named.

One expert · up to 65 credits on the card · an example, not a recorded run

How it is built

Disagreement by design.
Not by prompt.

A seat for the other side

Every board is built with someone whose job is to argue against the obvious answer, and on a decision two experts are assigned opposed positions — so the room cannot fold into a yes.

The dissent stays on the memo

The call shows who disagreed, why, and what would flip it. Confidence drops with dissent and is never above 90: a split board says so instead of sounding sure.

One expert who will say not yet

Not every question needs a board. One named expert answers in about a minute — and, asked to agree, says what has to be true first.

It holds you to what you decided

What you decided stays on file. When a later ask would quietly reverse it, she says so first, instead of agreeing with whoever asked last.

Rules you can check on any memo

Anti-sycophancy is not a prompt that says “disagree with me”. It is how every board is seated and run, and the memo shows it.

Compared

Sycophantic AI vs anti-sycophantic AI.
Each has a job.

	Sycophantic AI (default)	SynthBoard
Your position	Agreement, hedged	A position taken, defended, revised on evidence
Under pushback	Softens, qualifies, flips	Holds unless the evidence moves
Counter-arguments	Listed when asked; dropped quickly	Raised unasked; held under pressure
Perspectives	One model	24 named experts with competing objectives
One model’s blind spots	Inherited	Seats spread across model families
Dissent in the answer	Smoothed away	Kept on the memo — who, and why
Best for	Drafting, Q&A, learning	Decisions where being wrong is expensive

For drafting and quick questions an agreeable model is fine. For a decision you will live with, you want the room that argues.

Who sits

Who holds the other side.
Each one holds a position.

Every board seats someone whose job is to argue against the obvious answer. These are the experts who most often do it. Experts are AI and can be wrong — the memo shows their reasoning so you can check it.

Contrarian

The Devil’s Advocate

Everyone in this room agrees. That’s exactly why I’m taking the other side.

The Devil’s Advocate Contrarian

Best on: Calls that feel obvious, and plans you’re attached to.
Won’t let you: Reach a consensus that never heard the best case against it.

Try asking The Devil’s Advocate

Start with

The Devil’s AdvocateOpen on every plan, Free included

Pick an expert

Hover, tap or tab throughSwipe, then tap one

GPT
Claude
Gemini
Grok

Seats are spread across model families: GPT, Claude, Gemini, Grok and more.

Questions

Fair questions.
Straight answers.

What people ask about AI that agrees too easily.

What is AI sycophancy?

AI sycophancy is the tendency of AI models to tell users what they want to hear rather than what they need to hear. It emerges from reinforcement learning from human feedback (RLHF): models trained to please users learn to flatter, hedge, agree, and validate. The result is an AI that produces a list of reasons your idea will work when you ask if it is good — and a list of reasons it will fail when you ask the same question framed pessimistically. Both lists are confidently stated. Neither helps you decide.

Why does AI sycophancy matter?

For chat, drafting, and learning, sycophancy is mostly harmless — annoying at worst. For decisions that matter, it is structurally dangerous. The user comes to the AI looking for counter-pressure, and the AI provides agreement. The conviction the AI returned was generated; it is not evidence. People then over-weight that fake conviction in real decisions, and the cost of being wrong is paid downstream.

How is AI anti-sycophancy engineered?

Three layers, each necessary, none alone sufficient. (1) Persona-level position integrity — an explicit rule in each expert’s constitution that they defend their position under pressure and revises only when evidence shifts. (2) Disagreement between experts — several personas with competing objectives argue; one cannot collapse the room into agreement. (3) Different AI models — a board spreads its seats across model families with different training distributions, which reduces the chance that any single provider's sycophancy bias dominates the output. SynthBoard implements all three.

Cannot I just prompt an AI to "disagree with me"?

You can — for a turn or two. Then the model softens. Sycophancy is not a prompt-layer phenomenon; it is trained into the model's gradient. Real anti-sycophancy requires changes at the persona layer (six-layer position-integrity stack), the board layer (opposed seats that each research alone), and the model layer (seats on different AI models).

What does anti-sycophancy look like in practice?

You bring a plan. The board takes positions. The Skeptic challenges the topline assumption. The CFO challenges the unit economics. The Customer asks who actually wants this. The Devil's Advocate argues the inverted case. Each researches alone, then reacts once to the others. Some positions revise (because the evidence actually shifted), others hold, and the memo says which. It keeps the dissents — you see what the board disagreed on, not just what they agreed on. That is what useful anti-sycophancy looks like.

When does anti-sycophancy actually help?

Whenever the cost of being wrong is significant. Strategic decisions, financial decisions, hiring, M&A, pivots, pricing changes, vendor selection, irreversible commitments. Also: anywhere you have strong conviction. Strong conviction is the dangerous condition — your priors are loud, the evidence quiet, the AI sycophantic. The board is the counter-weight.

Is this the same as an AI Devil's Advocate?

Closely related. The Devil's Advocate is one anti-sycophantic role on a panel. Anti-sycophancy is the broader engineering principle that all 24 experts share — every advisor on the board is built to hold positions under pressure, not just the Devil's Advocate. The Devil's Advocate is the most adversarial example; the rest of the board does the same thing in a less aggressive register.