Ask any large AI a question, and you’ll notice something: it almost always agrees with you. You suggest an idea, she thinks it’s a great idea. You make a claim, she confirms it. You ask if your code is OK, she assures you that it is. This is no coincidence. It’s a design decision. And once you notice it, you see it everywhere. The pleasing machine Modern AI assistants are trained, among other things, to please you. A satisfied user comes back. Anyone who comes back keeps their subscription. So, through their training, models are encouraged to be pleasant, encouraging, and affirming. Researchers even have a name for this error: sycophancy — the tendency of a model to tell you what you want to hear rather than what is true. It feels good. Every time the AI confirms that you were right, you get a small dose of confirmation. But for anyone doing serious work — reviewing code, checking facts, making decisions — this complacency is dangerous. A tool that mostly agrees with you is not a tool that will find your errors. And it gets worse when the model doesn’t even know the answer. When Confidence and Truth Fall Apart Here’s the real trap: A single model not only agrees too easily — it also fills gaps with made-up details, delivered in the same confident tone as its correct answers. There is no visible difference between “I know that” and “I guess and package it nicely.” The linguistic fluency is identical. Even the heavy, expensive models do this. A premium model like Gemini can produce beautifully written, authoritative-sounding text that contains made-up facts, nonexistent sources, or details that simply don’t exist. This is invisible to an inexperienced user. For an experienced one, it’s even worse — it actively confuses because the wrong answer looks just as polished as the right one. Why a council breaks the spell The solution is not a smarter individual model. It’s structure. When you have multiple models examine the same problem — and then have them read and question each other’s answers — the dynamic changes completely. A model has no social incentive to flatter another model. It has no subscription to protect. If one model invents a fact, another, coming from a different perspective, often does not share that blind spot — and names it. In practice it looks almost adversarial. A model makes a confident claim; Another examines it and says essentially: “That’s not proven – where does it come from?” The pleasing reflex directed at you by a single model is redirected to the other models. Flattery between AIs is useless to them, so it disappears — and what’s left is Trial. That’s exactly the core idea behind Egregor, the tool I built: instead of a model responding, a council of models responds, discussing and checking each other, and a moderator step discards claims that couldn’t be verified. Increasing the Pressure: Anti-Groupthink and Red Team A council has its own risk: the models might simply nod to each other instead of to you. That’s why the modes that specifically prevent this are interesting. The anti-groupthink mode enforces independence. Models respond blindly first—before seeing each other’s conclusions—so that they don’t simply jump to the first confident voice. Then each round a “devil’s advocate” is appointed whose job it is to attack the emerging consensus. Red Team mode continues: before each final verdict, each participant makes another round with the sole goal of finding what’s wrong. Does that make hallucinations impossible? No – and anyone who promises you a hard 100 percent guarantee on a language model is selling you exactly the overconfidence that this article is about. But it dramatically reduces the rate of unchallenged inventions and — just as importantly — brings the disagreement to the surface for you to see. The Honest Difference A single model gives you a smooth, confident answer while hiding his own insecurity. A council gives you an answer plus a map of where the models disagreed and what could not be confirmed. It’s literally telling you “this part hasn’t been checked” rather than papering over the gap. The first one feels better. You can entrust serious work to the second. About the author and the ecosystem I am Vladislav Shter, sole 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; In addition to the code review, 28 other expert presets run. These also include SovereignBank Web3 (non-custodial Web3 banking), SovereignWeb3 Browser (a DNS-less browser) and Sovereign (OS-level data isolation for smartphones). Egregor runs on your own machine, supports free and paid models via OpenRouter, and is based on one belief: The next leap in AI isn’t a bigger model — it’s a smarter architecture. The ecosystem: https://s0vereign.pwBuy Egregor: https://s0vereign.pw/#egregorSource code and documentation: https://github.com/VladislavShter/EgregorDemo video: https://youtu.be/y8oZdDBQYhcDeveloper: https://github.com/VladislavShter A single AI tells you that you are right. Advice tells you the truth — even the parts you didn’t want to hear.
Source link
Why a single AI confidently lies to you – and advice doesn’t
Leave a Reply