People often ask whether AI is intelligent. I think we're asking the wrong question. The more interesting question is this: What kind of conversation produces reality-aligned thinking?
Over the past year, I've spent thousands of hours working with AI while creating books, developing philosophical ideas, testing arguments, exploring personal questions, and documenting my own thinking. Along the way, I've noticed something surprising.
One thing I’ve repeatedly observed is that the quality of the conversation often depends as much on the person asking the question as it does on the AI answering it.
That does not mean AI is always right. It isn't.
It also does not mean AI simply reflects whatever the user believes. It doesn’t. It does mean something subtler.
The structure of the inquiry profoundly influences the structure of the dialogue.
I have come to believe that this may be the most important question behind effective human-AI collaboration. Not, “How do I get a better answer?” but, “What kind of inquiry produces better reasoning?”
That shift in understanding changed what I thought The Uncertainty Project was about. I originally thought of it as a record of long-term human-AI interaction. A journal documenting what happens when “intelligent systems” become a part of everyday thinking.
But that’s only half the story.
The more important question isn’t, “What happened?” It’s, “How did I arrive at what I now believe?” That shifts the project from documenting conversations to documenting the actual process of knowing itself. That process is what I want to understand.
AI Is Not the Most Interesting Part
When people dismiss AI as “slop,” they're usually evaluating answers and talking about outputs. But what interests me are the conditions that produced those answers.
Imagine two conversations.
The first begins like this: “My boyfriend is abusive. Tell me why.”
The second begins like this: “Here's everything that happened. Here are the parts I'm uncertain about. Here are three explanations I'm considering. Tell me where I might be fooling myself.”
Those are not the same conversation. The AI has not been given the same material. The reasoning process cannot be identical.
The second conversation is not better because it is longer. It is better because it contains a healthier relationship with uncertainty. Instead of demanding confirmation, it invites investigation. That distinction may be one of the most important discoveries I've made.
The Mirror Effect
Instead of just thinking about AI as a machine that gives answers, I’ve started seeing it more as something that can reveal the structure of my own inquiry.
If I bring vague thinking, I often get vague reasoning.
If I bring emotionally loaded conclusions, I may get emotionally loaded elaborations.
If I bring observations, competing explanations, uncertainty, and a willingness to revise my own assumptions, the conversation can become far more useful.
The AI is not simply generating text. It is interacting with the way I am thinking. In many conversations, it functions as a cognitive mirror. It doesn’t just reflect my beliefs; the structure of my inquiry actually shapes the structure of the dialogue that follows. It can challenge my assumptions or give me a perspective I had not considered. And sometimes it’s flat-out wrong. It’s not a mirror or an oracle in any literal sense, but it can reveal patterns in my own thinking that might have otherwise remained invisible.
That leads me to a tentative hypothesis:
AI often reveals the structure of our thinking before it reveals the quality of its own thinking.
This may be the single most important realization of The Uncertainty Project. The way we ask questions shapes the way we understand ourselves, other people, and the world. If our inquiries begin with assumptions, our conclusions will often inherit those assumptions. Better questions don’t guarantee true answers, but they can create better conditions for evidence to shape where we end up.
Reality-Aligned Dialogue
Looking back over many conversations, I've noticed several habits that consistently improve the quality of the reasoning. None of them are really about AI. They're about thinking.
1. Bring observations before conclusions.
Instead of beginning with, “He’s abusive,” begin with, “Here’s what happened.” Description almost always leads to better reasoning than premature judgment.
2. Separate facts from interpretations.
Instead of saying, “He lied,” say, “Here’s why I suspect the story may not be true.” That keeps uncertainty visible instead of pretending certainty where it doesn't really exist.
3. Invite correction.
Some of the most valuable conversations I've had began with questions like, “What alternative explanations am I missing?” “Where might I be wrong?” “Can we pressure-test this?” Those questions change the entire dialogue. The goal shifts from being confirmed to becoming more accurate.
4. Keep uncertainty explicit.
There is a big difference between these two statements: “I know what happened,” and, “Here’s what I know. Here's what I suspect. Here's what I don't know.” The second creates room for reality to surprise us.
5. Revise when evidence changes.
A useful model is one that remains responsive to reality. When new evidence appears, good thinking adapts. It doesn’t defend the previous explanation at all costs.
6. Distinguish explanation from justification.
Understanding why someone behaves a certain way is not the same as approving of what they did. Compassion and accountability are not the same thing, but they are equally important. This distinction has become one of the organizing principles of my work.
7. Translate insight into behavior.
Ideas are only half the journey. If understanding never changes action, then it stays at the level of philosophy instead of becoming wisdom. Reality alignment is demonstrated by making better decisions, not by having better ideas.
8. Prefer calibration over confidence.
Instead of just asking, “What's the answer?” also ask, “How confident should I be?” “What evidence would change this conclusion?” “What important information might still be missing?” Reality alignment isn’t about becoming more certain. It’s about becoming more accurately calibrated to the available evidence.
None of these habits totally eliminate AI's limitations. AI can still misunderstand context, hallucinate facts, overlook important evidence, misunderstand nuance, or confidently generate incorrect conclusions. Reality-aligned dialogue doesn’t eliminate those risks, but it does reduce them.
Beyond Prompt Engineering
Prompt engineering focuses on improving outputs. Reality-Aligned Dialogue focuses on improving the quality of dialogue by improving the habits of inquiry brought into it.
Learning to separate observation from interpretation, hold uncertainty, be open to correction, revise conclusions when evidence changes, and distinguish explanation from justification has changed the way I think, the way I write, and the way I relate to other people.
Maybe because these aren’t really habits of prompting. They are habits of dialogue. In fact, none of this is really about prompts. Prompt engineering asks, “How do I get better answers?” But what I'm asking is very different. “How do I become a better participant in the search for reality?”
That question survives every future version of AI. It also goes beyond AI. Applying the same principles would improve a conversation with a thoughtful friend, a therapist, a teacher, Socrates, Carl Rogers, Gregory Bateson, or Krishnamurti.
Common Reality Traps
It’s not all about the prompts. Many failures in AI dialogue begin with bad inquiry habits. Others originate in the model itself. The difficult cases often emerge from the interaction between the two. Some of the most common mistakes people make are 1) hiding uncertainty, 2) asking for confirmation instead of investigation or analysis, 3) presenting conclusions as established facts, 4) ignoring evidence that contradicts an initial interpretation, 5) treating AI confidence as evidence that something is true, and 6) continuing to defend an explanation after the evidence has changed.
These are also human thinking mistakes. AI can make them easier to see—or harder to detect.
The model has its own reality traps as well: resolving uncertainty too quickly, reproducing assumptions from its training data, agreeing too readily with the user, and producing language whose fluency exceeds the strength of its evidence. Reality-Aligned Dialogue requires examining both sides of the exchange.
A Different Kind of Documentation
This realization has changed the direction of The Uncertainty Project.
I no longer think I should just document discoveries. I think I should document how discoveries become possible.
That suggests two complementary forms of documentation. The first is archaeological. It asks, “What happened?” “What changed?” “What did I learn?” “Where did I make mistakes?” “How did an idea evolve?”
The second is methodological.
It asks, “How did this realization become visible/possible?”
Not just the conclusion, but the actual path itself. Because paths can be learned. Methods can be shared. And thinking can be practiced.
AI Amplifies Inquiry
AI doesn't just amplify intelligence. It amplifies the quality of the inquiry brought into the conversation.
It amplifies assumptions, curiosity, certainty, uncertainty, bias, intellectual honesty, laziness, rigor.
If you bring one kind of inquiry, you’ll receive one kind of dialogue. If you bring another, you’ll receive another. It’s that simple. Though, obviously, there’s more to it than just that.
Imagine two people asking about the same situation. One asks, “Tell me why my friend is selfish.” The other asks, “Here’s what happened. What are some different ways of interpreting this?”
The first inquiry already contains a conclusion. The second contains a question. The AI may challenge either one, misunderstand either one, or give poor reasoning in either case. But the conversation begins from fundamentally different places.
Or imagine someone asking, “Help me prove I'm right,” compared with, “Help me understand what I'm missing.” Those aren’t just different prompts. They’re different relationships with reality.
The more I work with AI, the more I see that this may be one of the most significant variables shaping the quality of the dialogue. The inquiry itself becomes part of the system. Knowing this doesn't guarantee good reasoning. But it changes the conditions under which good reasoning becomes possible.
Dialogue as a Laboratory
I now imagine a recurring feature within The Uncertainty Project that I think of as Dialogue Lab.
Rather than curating impressive AI conversations, the goal would be to examine how reasoning unfolded. Not just what the AI said, but how understanding emerged.
A typical Dialogue Lab might trace a conversation like this:
The conversation itself becomes an experiment in thinking. That feels much more useful than simply archiving clever exchanges with AI. The goal isn't to showcase AI, but to make the process of reasoning visible.
The Real Subject
Maybe The Uncertainty Project isn’t even fundamentally about artificial intelligence. AI is the setting, the mirror, the pressure chamber. But the real subject is the person. The human being.
More specifically, it is the human capacity to remain curious, discerning, intellectually honest, and responsive to reality while collaborating with increasingly intelligent systems.
To me, that seems to be a question worth documenting. Because AI will certainly continue to change, but more importantly, the habits that help us think more truthfully will matter long after today's technologies have become obsolete.
That could actually be what The Uncertainty Project has been about from the start. Not documenting AI, as I originally thought, but documenting the lifelong practice of learning how to suss out reality in conversation, be it with machines, with other people, or even with ourselves.
If future evidence shows this framework is incomplete, then the framework will change. That is not a failure of the method. It is the method working.
Reality always has the final word.