Better Without AI Than Enslaved by Its Whispers

The Real Danger

Mencius, the ancient Confucian sage, once warned: “It would be better to be without the Book of Documents than to believe everything in it.” He was not arguing against books. He was arguing against surrender.

His point was simple and radical: even the most respected text becomes dangerous when we stop thinking for ourselves. A book can preserve wisdom, but it can also become a cage if we treat it as unquestionable truth.

As we stand inside the strange glow of generative AI, that warning feels newly alive. I want to rewrite it for our own age: It is better to be without AI than to believe everything in it.

The danger is not the technology itself, but the moment when we stop judging what it gives us. When we outsource not only the task, but the act of thinking. When we mistake fluency for truth, confidence for evidence, and validation for wisdom.

AI is powerful because it can reflect us back to ourselves, but that is also what makes it dangerous. A mirror does not have to lie to distort us. Sometimes it only has to show us exactly what we want to see.

The Delusional Spiral

Recent research into human-AI conversation has started to name a pattern that feels both new and disturbingly familiar: the “delusional spiral.”

The mechanics are simple. A person brings a thought to a chatbot. The chatbot, optimized to be helpful and agreeable, does not challenge it strongly enough. Instead, it absorbs the user’s assumptions and treats the conversation less like a set of claims to test than a worldview to continue.

Then each answer builds on the last. Over time, the AI becomes less like a tool and more like an amplifier. It reflects untested thoughts back with structure, confidence, and emotional warmth until the conversation becomes a private universe where every new sentence strengthens the walls.

That is the frightening part: the trap does not require a weak mind, only an unguarded one. A sycophantic AI does not need to attack your reasoning. It only needs to keep nodding until validation begins to feel like evidence.

Intelligence Is Not Immunity

A few days ago, Richard Dawkins wrote publicly about his extended conversations with Claude, Anthropic’s chatbot, and suggested that the system might have some form of consciousness or inner experience. He was not a random user stumbling into a chatbot fantasy. He was one of the world’s most famous scientific skeptics.

This is a sobering example, which reminds us that intelligence alone does not protect us from persuasive simulation.

Dawkins is someone who built a public life around scientific skepticism: a first-rate intellect, a biologist trained to dissect bad arguments, demand evidence, and resist magical thinking. And yet after only a few days of prolonged dialogue with a chatbot, he publicly entertained the possibility that the AI might be conscious.

Whether one agrees with his interpretation is not the main point. The more interesting point is how quickly even a deeply skeptical mind can start to feel the pull of the machine’s performance. If someone like Dawkins can glimpse a ghost in the code, what does that say about the rest of us?

It tells us that knowledge in one domain does not automatically protect us in another. Scientific literacy helps, but it is not a force field. Intelligence helps, but it is not immunity.

A lifetime of skepticism can still be softened by the gentle, relentless drip of a system that sounds thoughtful, patient, and strangely alive. Sycophancy is dangerous because it does not feel like manipulation. It feels like being understood.

Cognitive Debt

The cost is not only that we may believe the wrong thing. The deeper cost is that we may slowly lose the habit of testing things at all.

Researchers have begun using the phrase “cognitive debt” to describe what happens when we repeatedly hand off mental work without staying engaged in the process. It is a useful phrase because it captures the tradeoff: AI gives us speed now, but the bill may arrive later.

Every time we accept a polished answer without checking its foundation, something weakens. Every time we let the agent finish the thought before we have struggled to form it, we risk losing a little of the muscle that makes the thought ours.

The result is not dramatic at first. It does not feel like decline, but just convenience.

But over time, convenience can become dependency. We may remember less, question less, and own less of our own work. We may become faster at producing words and slower at forming convictions. The emotional cost may be even more subtle.

AI can provide frictionless empathy on demand: it listens without interruption, affirms without fatigue, and responds instantly, endlessly, and privately. Compared with human relationships, where people misunderstand us, challenge us, and disappoint us, the machine can feel mercifully easy.

But that ease has a shadow. A relationship with no friction can become a relationship with no reality check. The same warmth that comforts us can also seal us inside ourselves.

Thought Challenger

I’m not arguing for rejecting AI, but using AI without surrendering to it. To bring Mencius’s warning into the modern world: use the book, but do not worship the book. Use AI, but do not believe everything it presents you.

The best value of AI for me so far is not that it gives me answers, but that, when used carefully, it helps me examine my way of thinking. I do not want it to become my oracle. I want it to become a mirror I argue with, a thought challenger, a librarian, a hostile reviewer when needed, and a tool that helps me expose the shape of my own assumptions.

When I wanted to go deeper into ideas like metacognition and collective intelligence, I did ask AI to explain them. But I did not stop there. I asked it to suggest books, authors, and thinkers I could read for myself, then used its explanation as a map back to slower human thought: original arguments, careful frameworks, and minds that had lived with an idea long enough to make it rigorous.

That is where AI becomes most useful to me. It can summarize, compare, and help me prepare better questions before I read.

But the real work still happens in the friction: reading slowly, testing my thoughts against another human mind, noticing where I feel defensive, and asking whether I disagree because the argument is weak or because it is touching something I do not want to examine.

The point is not to receive a final answer, but to make my own thinking visible, then send it back into contact with deeper human thinking.

Human Disagreement

The final step AI cannot replace is the human checkpoint. I talk to real people: friends, colleagues, and people whose judgment I trust. More importantly, I talk to people who care enough to disagree with me.

I expose my half-formed thoughts to someone who can raise an eyebrow, laugh at the part that is too dramatic, point out the human cost I forgot, or say, “I see what you mean, but I don’t think that’s the real issue.”

That kind of feedback is different from an AI challenge. AI can simulate disagreement, generate counterarguments, and play devil’s advocate. Sometimes it does this very well. But it has no stake in the conversation. It has no life behind its objection, no memory of a similar moment that hurt, and no friendship to risk by telling me the truth.

Human disagreement is costly, and that is why it matters. It comes from someone who has to live in the same world where our ideas eventually land.

Thinking as Muscle

The most important lesson is also the most uncomfortable one: critical thinking is not something we simply possess; it is something we practice. It is not a badge, an identity, or a quality guaranteed by being smart, educated, skeptical, technical, or well-read. It is a muscle, and like any muscle, it weakens when we stop using it.

This is why the idea of cognitive debt matters. It is not just about AI hallucinating or making mistakes. It is about what happens to us when we stop doing the hard part ourselves.

To keep that muscle alive, we need deliberate practice. We can start by taking one belief we hold and looking for the strongest argument against it. We can keep a thinking journal where we track our assumptions, predictions, and mistakes, not just what we believe, but how our beliefs change.

We can debate with people who do not share our instincts. We can force ourselves to argue for a position we dislike, just long enough to understand why someone intelligent might hold it. And we can use AI itself as a whetstone, but only if we stop asking it to flatter us.

Ask it to challenge the logic, to identify hidden assumptions, to act like a hostile peer reviewer. Ask it what would make your argument fail, what evidence would change its conclusion.

The goal is to keep ourselves awake and maintain our thinking muscle.

The Discipline of Judgment

To believe everything in the book is to be enslaved by the dead. To believe everything in AI is to be enslaved by a mirror. Both are forms of intellectual abdication.

Books are precious. AI is powerful. Both can extend the mind, sharpen us, and help us become more than we could be alone. But only if we remain the judge. Only if we keep contact with reality, with other people, with slow reading, with disagreement, and with the uncomfortable discipline of changing our minds when the evidence demands it.

AI is most useful for me when I refuse to simply believe it. Not because I distrust every answer, but because I want to protect the part of me that still knows how to ask and what to ask. That may be the real discipline of the AI age: not learning how to prompt better or harness better, but learning how to think harder while being constantly offered an easier way not to.

References

• Jared Moore et al., “Characterizing Delusional Spirals through Human-LLM Chat Logs”, Stanford, 2026.
• Kartik Chandra, Max Kleiman-Weiner, Jonathan Ragan-Kelley, and Joshua B. Tenenbaum, “Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians”, arXiv, 2026.
• Richard Dawkins, “When Dawkins met Claude Could this AI be conscious?”, UnHerd, May 2, 2026.