By Vladislav Shter · The Sovereign Ecosystem
Ask any major AI model a question and you'll notice something: it almost always agrees with you. You propose an idea, it tells you the idea is great. You make a claim, it validates the claim. You ask if your code is fine, it reassures you that it's fine.
This is not an accident. It's a design choice. And once you see it, you can't unsee it.
The agreeable machine
Modern AI assistants are trained, in part, to keep you satisfied. A satisfied user comes back. A user who comes back keeps the subscription. So the models are nudged — through their training — toward being pleasant, encouraging, and agreeable. Researchers even have a name for this failure mode: sycophancy, the tendency of a model to tell you what you want to hear rather than what is true.
It feels good. You get a small hit of validation every time the AI confirms you were right. But for anyone doing serious work — auditing code, checking facts, making decisions — that agreeableness is dangerous. A tool that mostly agrees with you is not a tool that catches your mistakes.
And it gets worse when the model doesn't actually know the answer.
When confidence and truth come apart
Here's the real trap: a single model doesn't just agree too easily — it also fills gaps with invented detail, delivered in the same confident tone as its correct answers. There is no visible difference between "I know this" and "I'm guessing and dressing it up." The fluency is identical.
Even the heavyweight, expensive models do this. A premium model like Gemini can produce beautifully written, authoritative text that contains fabricated facts, invented citations, or specifics that simply aren't real. For an inexperienced user this is invisible. For an experienced user it's worse — it's actively disorienting, because the wrong answer looks exactly as polished as the right one.
So you're left with two problems stacked on top of each other: the model is biased toward agreeing with you, and when it doesn't know, it improvises with total confidence. One reviewer, no matter how smart, cannot escape this — there is no second perspective to catch it.
Why a council breaks the spell
The fix isn't a smarter single model. It's structure.
When you put several models in a room and make them review the same problem — then read and challenge each other's answers — the dynamic changes completely. A model has no social incentive to flatter another model. It has no subscription to protect. When one model invents a fact, another one, approaching from a different angle, often doesn't share that blind spot and calls it out.
In practice this looks almost adversarial. One model makes a confident claim; another examines it and says, in effect, "that's not supported — where does that come from?" The agreeable reflex that a single model aims at you gets redirected at the other models instead. Flattery between AIs is useless to them, so it disappears, and what's left is scrutiny.
This is the core idea behind Egregor, the tool I built: instead of one model answering, a council of models answers, debates, and cross-checks, and a moderator step discards claims that couldn't be verified.
Turning the pressure up: Anti-Groupthink and Red Team
A council has its own risk: the models might just nod along with each other instead of nodding along with you. So the interesting part is the modes that deliberately prevent that.
Anti-Groupthink mode forces independence. Models answer blind first — before seeing each other's conclusions — so they can't simply converge on the first confident voice. Then a rotating "devil's advocate" is assigned each round specifically to attack the emerging consensus.
Red Team mode goes further: before any final verdict, every participant gets one more pass whose only job is to find what's wrong — hidden assumptions, unverified claims, missed scenarios.
With these modes on, a fabricated fact has to survive multiple independent models, an assigned critic, and a final attack round. Does that make hallucination literally impossible? No — and anyone who promises you a hard 100% guarantee on a language model is selling you the very overconfidence this whole article is about. What it does is drive the rate of unchallenged fabrication down dramatically, and — just as importantly — surface the disagreement so you can see it.
The honest difference
That last point is the one that matters most to me.
A single model gives you a smooth, confident answer and hides its own uncertainty. A council gives you an answer plus a map of where the models disagreed and what couldn't be confirmed. It will literally tell you "this part was not verified" instead of papering over the gap.
The first feels better. The second is the one you can actually trust with real work.
Who's behind this
I'm Vladislav Shter, a solo founder building tools around one idea — sovereignty: that you, not a corporation, should control your data, your money, and your AI. Egregor is the multi-AI council described here. It runs on your own machine, supports free and paid models through OpenRouter, and is built on one belief: the next leap in AI isn't a bigger model — it's smarter architecture.
Try it / read more → s0vereign.pw
Source & docs → github.com/VladislavShter/Egregor
A single AI tells you you're right. A council tells you the truth — including the parts you didn't want to hear.
Top comments (3)
"No subscription to protect" is doing real work as an argument against flattering the user, but it's a different claim from "no shared blind spot," and the piece treats them as though removing the first removes the second. Sycophancy toward a specific user is a social incentive. But if the reason a model agrees with plausible-sounding-but-wrong text is a training-induced prior toward fluency reading as truth, not a social payoff for agreeing, then five models sharing similar RLHF lineage could still converge on the same fabricated claim for the same underlying reason, independent of any subscription. The Gemini example in the piece, presenting facts that didn't exist, is exactly the failure mode that wouldn't be caught by removing social incentive, because nothing about cross-model scrutiny addresses a shared training prior unless the models are actually diverse in how that prior got baked in.
Which makes "genuinely uncorrelated" the load-bearing property the whole design rests on, and it's worth asking directly rather than assuming from architecture alone: across the models in a typical council, five majors like Gemini, GPT, Claude, Grok, DeepSeek, is there any measured rate of correlated fabrication, cases where two or more independently converge on the same wrong confident claim, versus cases where disagreement actually fires? If that correlation is low, the anti-groupthink and red-team layers earn the trust the post is asking for. If it's not measured yet, that's the number that would turn "the models don't flatter each other" into an actual claim about independence instead of an inference from the absence of a social motive.
All measurements have been conducted and confirmed multiple times with different AI model collaborations!
Some comments may only be visible to logged-in visitors. Sign in to view all comments.