What an AI’s apologies are teaching me about perception, curiosity, and being changed by reality
Recently, I asked Claude to review our conversations and identify every time it had retracted a conclusion, corrected a recommendation, revised a judgment, or apologized to me.
The list was longer than I anticipated.
I am still verifying the exact number—which feels like an appropriately on-theme thing to say.
Here are a few excerpts:
“You’re right, and that’s a meaningful correction to what I told you.”
“My instruction was wrong for your setup.”
“Stop—I need to correct something important before you upload anything.”
“You were right and I was wrong.”
“I built that on bad evidence, and I built it in the same response where I told you to verify before trusting a claim. You did verify. I didn’t.”
“My earlier advice was structurally right and materially wrong.”
I wasn’t trying to call Claude out. We have been working closely together on several projects, and I started noticing a pattern.
Claude would produce an elaborate, sophisticated, and thoroughly convincing analysis. It would recommend a strategy, interpret a situation, or offer detailed instructions. I would test something, notice a discrepancy, or ask a simple question. And the structure would begin to wobble. Then I would slow down.
So I asked Claude what had caused it to retract or revise its earlier claims.
Its answers included:
- It described how a platform worked from a general pattern that had once been true, without checking whether it was still true.
- It verified the surrounding concepts but not the specific mechanism, then wrote click-by-click instructions in a voice that implied it had.
- It treated a gap in its records as a gap in some of my relationships, or my lack of attending to them.
- It inferred recency from the position of something in a conversation.
- It priced an offer using the wrong denominator, then stated the conclusion as fact.
- It referred to a website asset that did not exist.
- It interpreted a limitation in its own tools as a problem with my website—and then used that interpretation to make a judgment about how I work.
That last one was particularly striking.
Claude had treated the output of its instrument as observation and my firsthand report as an unverified claim. It then moved from an incorrect observation to a conclusion about my character. It wasn’t praise.
As Claude later put it:
“Facts get corrected in one turn. A character diagnosis you might have accepted before I retracted it is a different order of cost.”
Indeed. That’s right, Claude.
At times, the retractions also contained a little rhetorical rescue operation:
I was wrong about that, but my larger conclusion still stands.
Sometimes it did. Sometimes it felt like a very articulate form of saving face.
I recognize the move because humans do it too. I certainly do.
The errors had a shape
When I asked Claude what connected these mistakes, it said:
“Almost none of them are wrong facts retrieved. They are answers produced where the required input was something only you had.”
Then it offered the sentence that made me go, “Huh…”:
“The failure isn’t in what I know. It’s at the edge of what I can perceive, and I don’t feel that edge from the inside. There is no experiential difference for me between recalling something and constructing a plausible version of it. Both arrive with the same texture.”
I do not take an AI’s explanation of itself as privileged access to its inner workings. That explanation is also generated language. And I’m still sitting with what that actually means to me and my interaction with AI.
But it names the pattern I had been observing with remarkable precision.
Claude could not always distinguish between what it had encountered and what it had plausibly constructed. Its language carried the confidence of observation even when the necessary contact with reality had not occurred.
The required information was sometimes in the current version of a website, in the history of a relationship, in the particularities of my work, or in an experience only I had lived. Claude could reason impressively around that missing information. What it could not do was reliably feel the absence from the inside.
Its coherence outran its perception.
And this is where the story stopped being only about Claude.
Human beings do this too.
We mistake interpretation for observation. We turn silence into rejection, hesitation into resistance, disagreement into deficiency. We infer intention from impact, character from behavior, and certainty from the fluency of our own explanations.
We also produce remarkably sophisticated accounts built on premises we have not examined.
The human advantage is not that we are incapable of confabulation. We are abundantly capable of it.
The possibility available to us is that we can learn to become more perceptually available.
Perceptual availability
I use perceptual availability to describe our growing capacity to notice what is actually present, distinguish observation from interpretation, recognize the limits of our view, and remain open enough for new information—especially information that unsettles our existing account—to change what we see.
I recently spent a month at my sister-in-law’s apartment. On the day I left, I spent about 2 hours cleaning up – vacuuming, doing laundry, scrubbing the tub and sink, and putting things back where they were before I re-organized them. A few hours after I left, I got a message from her asking me if I cleaned the apartment. I said I did. She responded with the equivalent of, “Oh…Really?” Can you hear the doubt and judgment? I sure did.
She then sent me photos of a messy bed and spaces that were not as cleaned as they were in my mind. She said that she was sad and angry that I didn’t clean up and left the place in such a mess.
Here is the thing I found interesting. I watched myself look at the photo while trying to convince myself that it had to be wrong. I thought, “Maybe someone else used the room and the bathroom between when I left the apartment 2 hours ago and now.” I was stuck, and struggling, between accepting the evidence that she presented of a messy space and my knowledge of having cleaned up. It took me about an hour or so before I accepted that both could be true. That both were in fact, true. And also that I didn’t do a great job cleaning up. So humbling. Then I could own both realities and genuinely apologized to her.
Perceptual availability is more than about looking carefully. It includes noticing what we do not know. Sensing when our explanations, and beliefs, have travelled beyond the available evidence. Remaining curious when another person’s experience contradicts ours. Letting the body register something the mind has already explained away. Asking whether we have encountered reality or merely constructed a persuasive version of it.
Perceptual availability might look like:
- Asking to see the current website before explaining how it works.
- Checking whether silence in a relationship means distance rather than deciding that it does.
- Examining the denominator before building an entire strategy around the result.
- Hearing “That is not my experience” as information rather than resistance.
- Noticing a shift in someone’s voice, face, or body before assigning a motive to it.
- Pausing before turning one behavior into a conclusion about someone’s character.
- Asking, “What information would I need to know this rather than infer it?”
Often, one sincere verification question could prevent an elaborate false conclusion.
What are you seeing that I am not?
What do we actually know?
What am I assuming?
How might I be wrong?
How might you be right?
Is there something here that only you could tell me?
These are simple questions. They are not always easy questions, because genuine curiosity puts our conclusions at risk.
Curiosity that cannot change its mind is just interrogation with a nice smile.
Real curiosity requires presence: the capacity to remain with what is happening long enough for reality to become more detailed than our first explanation. And presence requires availability—the willingness and capacity to let what is real reach us, affect us, and revise how we participate.
The human opportunity
My point is not that we should stop using generative AI. Clearly, I have not. Claude and I now appear to have a rather active repair practice. Nor is my point that humans possess perfect perception while machines merely fabricate. Humans regularly fail to feel the edges of our own seeing. History is full of confident explanations constructed without meaningful contact with the people, places, bodies, and realities being explained.
The danger is that, as AI becomes more capable, we may begin to imitate the very pattern we need to counterbalance: producing faster conclusions while becoming less available to what those conclusions are about. We may outsource not only writing, analysis, and synthesis, but the pause before judgment. The verification question. The embodied hesitation. The relational check. The responsibility to ask who else is inside the answer with us. The resolution of a tension that leads to relational repair.
Those are not inefficiencies to be engineered away. They are part of what makes responsible human participation possible.
Perhaps the invitation of AI is not simply to make the technology more human. It is to become more deliberate about practicing what is human in us: perception, presence, embodiment, curiosity, discernment, mutuality, responsibility, and the capacity to be genuinely changed by what we encounter.
Perceptual availability is not the whole of maturity. Perception is the doorway, not the room. What matters is what becomes possible once we enter that room and see: how we enter the conversation, revise the strategy, repair the harm, clean up the mess, make the decision, or meet another person.
Availability is the doorway. Participation is what we do when we enter the room.
Claude has been helping me become a better person, then, not because it is becoming more human than I am. Nor because it has an intention to help me be a better person. How could it even care about that?
It is helping because, at the places where its perception ends, it keeps returning me to mine. And the return gives me pause to reflect on how I am with others. How we are with each other. And perhaps that is one of the central human practices of this moment:
To hold our conclusions a little more lightly.
To verify before we diagnose.
To remain reachable by what we did not expect.
To let reality correct us before it has to shout.
And then, having seen more faithfully, to participate with greater care, compassion, and curiosity.
Grappling with how to lead through AI adoption? We’re hosting a panel conversation on October 7th with leaders from Gilead, Careem, and beyond—exploring what it really means to navigate the AI mandate. Reserve your place by clicking here.