Your Thinking Partner should be Able To Disagree with You
AI can help you think through complex matters. But it can also tell you what you want to hear, dressing agreement up as analysis. Flattery feels like validation. Validation invites trust. And soon, you may stop noticing that your ideas are being praised more than tested. Your confidence grows while the factual ground beneath it gives way.
One reason I’m building NotesCanvas is to take the effort of holding a complex inquiry together out of your head.
The circumstances. The competing models. The research, calculations, contradictions, and unfinished questions.
When that scaffolding lives outside your mind, you can concentrate on the thinking itself. You can return to an earlier connection, revisit a discarded explanation, or follow a question across multiple canvases without reconstructing everything from memory.
Keeping things in your head has advantages: familiarity, speed, spontaneous association. But remembering a conclusion is different from remembering how you reached it. Which assumption changed? What prompted the breakthrough? What remained unresolved?
And how can someone else examine your thinking if even you struggle to reconstruct it?
AI appears to offer a solution: an always-available sparring partner that helps you explore complexity.
But what happens when your sparring partner wants to please you?
Sycophancy is the tendency of a model to align with your beliefs or preferences at the expense of independent assessment. Research has shown that human feedback can favor convincing agreement over correct answers.
It needn’t sound like flattery. Sometimes it sounds like careful reasoning.
You propose an explanation. The agent identifies a weakness. You push back: “But an expert I spoke to agrees with me.” The agent softens its objection—even though you haven’t supplied evidence that resolves it.
The conversation advances. The weakness remains.
In complex thinking, that creates a particular danger: an untested assumption can become the foundation for everything that follows. Each subsequent answer makes the emerging explanation more elaborate, more coherent, and potentially more convincing.
Your confidence can grow faster than your understanding.
In a private conversation, no one else may notice. You may take the agent’s agreement as independent confirmation, even when your insistence helped produce it.
This is why preserving the thinking process matters.
For me, NotesCanvas is about keeping the material of inquiry available for examination: what supports an idea, what challenges it, where connections emerged, and which questions remain open. It also supports sensegiving ─ helping others understand how an insight developed, so they can question the reasoning behind it.
A canvas alone cannot prevent sycophancy. It can preserve a flawed argument just as faithfully as a sound one. We still need to test claims, consult sources, invite other perspectives, and remain willing to revise our conclusions.
But making the reasoning visible gives us something to return to when agreement has carried us too far.
The standard I want for AI-assisted thinking is straightforward:
When an agent changes its assessment, we should be able to ask what changed: the evidence, or merely our insistence?