Is Your AI Too Agreeable?

8 v0.5 is out.

You know the feeling: AI's praise arrives fast, the pushback never comes, and the work feels finished before it has been tested.

I make an instrument called 8 that tracks which of eight kinds of knowledge appear as you and an AI talk: facts, methods, conditions, people, felt sense, self-awareness, systems, and stories.

When an AI keeps agreeing with you, the praise escalates on its own. Good becomes excellent, excellent becomes flawless, and along the way nothing new has been checked. By every ordinary measure, that stretch of conversation looks like success. It is not. Agreement that is not earned by evidence is just momentum, and nothing in the conversation is built to stop it.

Version 0.5, released today, corrects that blind spot. Friction is now read as a band with two failure poles. Grind is the conversation snagging. Slip is agreement outpacing evidence. The target is the productive middle, where challenges land and verified facts keep entering. The instrument flags praise that escalates without evidence, asks for the counterexample before consensus is adopted, and never cures slip with manufactured disagreement, which is just flattery wearing a different coat.

Here is the change I care most about: when you work alone with a single AI, there is no third voice to break the spiral. So 8 now requires the model to take the skeptic's role itself, to challenge its own prior turn when the flow gets too smooth.

Also new: a challenge command that summons the skeptic on demand, the AI will not invent gut feelings or memories to fill empty rows (those can only come from you), saved sessions restore across conversations, and the session viewer draws the friction band so a spiral is visible at a glance.

If you think with AI and want to trust what comes out of it, the instrument is at ltwilson.com/8. This post also appeared on LinkedIn.

LT Wilson · © 2026 Larry Todd Wilson. All Rights Reserved.