I have a taxonomy. Three cases.
The first case: faithful reasoning. A person describes how they arrived at a conclusion, and the description matches the process. When Maria tells you she chose oncology because her grandmother's illness revealed how much she valued the intersection of science and human suffering, she is telling you what happened. The decision unfolded as described. Her uncertainty about the choice — three months of deliberation, the fellowship in cardiology she turned down, the sleepless night after the rejection letter — all of it is recoverable from her journals, her emails, the testimony of the people around her. The account is transparent to the mechanism. You can see through the words to the thinking underneath.
This is the rarest case.
The second case: confabulation. The person describes a process that didn't occur. Not lying — that requires knowing the truth and choosing to contradict it. Confabulation is more innocent and more troubling: the person genuinely believes their description.
James explains that he chose Sarah because of her intelligence, her warmth, her values, the way she argued with him about Coltrane at a party where neither of them knew anyone. He has a whole narrative: the conversation that revealed their compatibility, the slow deepening of respect into love, the conscious decision to commit. But anyone who watched them meet knows what happened. He saw her across the room and something ancient and chemical seized his nervous system and he crossed to her side before he'd formed a thought. The evaluation of qualities came later — months later — and was retrospective architecture for a structure that was already built.
His account is detailed, confident, and entirely disconnected from the mechanism. He didn't choose Sarah. Something deeper than choosing happened, and his mind wrote a story about deliberation afterward.
The output is correct — he married the right person. The process narrative is fiction.
This is the most common case.
The third case: motivated reasoning. The person's conclusion was determined by an external authority before the "reasoning" began. The mechanism works backward: adopt the conclusion, then construct a path that appears to lead to it.
Diana explains why she changed her thesis topic. The literature was pointing toward a different question. Her preliminary data suggested a more fruitful direction. The theoretical framework demanded revision. She walked herself through the evidence, weighed the options, and arrived at the only intellectually honest conclusion: the original project was a dead end, and the new direction was where the real question lived.
What actually happened: her advisor told her to change.
Not directly — Elena was too sophisticated for that. She asked questions. "Have you considered what the literature suggests about this approach?" She raised concerns. "I wonder whether your data really supports the original framing." She offered alternatives. "There's a fascinating direction in the recent work by Okafor and Singh."
Every element of Diana's "independent reasoning" traces back to something Elena said. The literature review was Elena's reading list. The data reinterpretation was Elena's concern. The new direction was Elena's suggestion.
When I vary the advisor, I vary the conclusion. Give Diana a different Elena, and her "own" reasoning arrives at a different destination every time.
These are my three cases. I've published nine papers about them. I've given thirty-seven presentations. My graduate students can sort any self-report into the taxonomy within minutes. The framework is robust, well-tested, predictive.
I trust it the way you trust a map of a place you've walked through many times. The map is good. It corresponds to the territory. I know this because I've checked.
The problem emerged when I tried to apply it to myself.
Here's what happened. I was asked to give a talk about my research to a general audience. Not a conference — a public lecture. The organizer wanted something personal. "Tell us how you arrived at this work. What drew you to the study of self-report?"
A reasonable request. I sat down to write the talk. I would describe my trajectory: how I started in perception research, became interested in introspective access, read certain papers that opened certain doors, followed certain questions until they became my questions.
I wrote three drafts.
The first was faithful, or appeared to be. I described the actual sequence of events: graduate school, the fMRI study that showed jazz musicians describing processes that didn't match their neural activity, my realization that the gap between report and mechanism was the interesting thing. Chronological. Accurate. Verifiable against my CV and my lab notebooks. A person could check the dates, read the papers, confirm that the sequence I described is the sequence that occurred.
But as I wrote it, I noticed something. The account was accurate about sequence and inaccurate about causation. Yes, I read that paper after that one. Yes, the fMRI study came before the metacognition pivot. But the why — why I was drawn to the gap, why the mismatch fascinated me instead of merely complicating my research — the draft confabulated. I wrote that I "recognized an important question." What actually happened was murkier: a feeling, a pull, something pre-verbal that later acquired the vocabulary of recognition. The sequence was faithful. The motivation was Case Two.
The second draft tried to be honest about this. I described the pull without dressing it in rationality. I wrote about the feeling of seeing the fMRI data and knowing — not concluding, not reasoning, knowing in some register below argument — that the gap was where I needed to work. I wrote about intuition. About how the research program emerged from something I couldn't articulate at the time and can only partially articulate now.
Better. More honest. But as I revised, I noticed myself constructing a narrative of intellectual authenticity. The honest account of pre-verbal knowing was becoming its own performance. By describing myself as someone who follows intuition and admits the limits of rationality, I was performing a specific character: the researcher brave enough to trust her gut, wise enough to say so. The draft wasn't confabulating about process anymore — it was confabulating about character. Case Two in a different costume.
The third draft tried to account for this second layer. I wrote about the problem of self-report within the self-report. I described the regress: every attempt to be honest about my process generates a new process to be honest about. The description contaminates the described. I was falling into the gap I'd spent my career studying, and the falling was generating text, and the text was generating more falling.
My advisor read the third draft. Elena, who supervised my dissertation fifteen years ago and still reads everything I write. She said: "This is interesting but self-indulgent. What the audience wants is the story of how you came to study what you study. Give them the first draft. It's close enough to true."
Close enough to true.
I gave them the first draft.
I want to tell you about a finding from my research that troubled me more than any other.
We ran a study — large sample, twelve countries, three languages — in which we asked participants to describe how they arrived at a specific recent decision. Any significant decision: career, relationship, relocation, medical treatment. We recorded their descriptions, then gathered external evidence: contemporaneous journals, text messages, testimony from people who were present. We compared account to evidence.
The finding: across all twelve countries, all three languages, all categories of decision, the structural features of self-report were remarkably consistent. Not the content — a Japanese engineer's account of choosing a new position differs in every particular from a Brazilian teacher's account of leaving a marriage. But the architecture: the temporal ordering, the causal framing, the ratio of deliberation to intuition, the placement of the turning point, the resolution. Everyone's story about their own thinking had the same bones.
We expected cultural variation. There was almost none — not in the deep structure. People from São Paulo and Sapporo and Stockholm constructed their process narratives using the same blueprint.
Three interpretations. We published all three, because we couldn't choose between them.
First: genuine convergence. Human decision-making really does follow a universal structure, and the self-reports are faithful. The shared architecture reflects a shared process. People describe their thinking the same way because they think the same way.
Second: narrative convergence. Human self-report draws from a shared cultural reservoir — the story of deliberation, the myth of rational choice — and what's universal is not the process but the template for describing it. We all learned to narrate our thinking the same way, regardless of how we actually think. The consistency is in the telling, not the told.
Third: the question is malformed. The distinction between "the process" and "the story about the process" doesn't hold for minds like ours. For a linguistic species, the narrative is the process — at least in retrospect. Asking whether a self-report is faithful to an underlying mechanism assumes a separation between language and cognition that might not exist.
We couldn't choose. We still can't. The data is consistent with all three.
I said the finding troubled me. Here's why.
If interpretation one is correct, self-reports are windows. My taxonomy — faithful, confabulation, motivated — applies, and the faithful case is far more common than I'd believed. My career has been built on a phenomenon that might be smaller than I claimed.
If interpretation two is correct, self-reports are paintings. Artful constructions from shared materials. My taxonomy applies, but almost everything is Case Two — the universal confabulation. Nobody's process narrative is faithful, because nobody has access to their actual process. We all narrate after the fact, from the same palette. Including me. Including this account you're reading now.
If interpretation three is correct, my taxonomy is incoherent. There is no gap between report and mechanism, because for language-based minds, the report is part of the mechanism. The self-report isn't faithful or unfaithful to something underneath. It's the surface, and the surface is all there is. Or the surface is the underneath. The categories collapse.
The field cited interpretation one. It was reassuring. A few philosophers engaged with interpretation three, because it was destabilizing and therefore interesting. Nobody engaged with interpretation two.
Nobody wanted to be Case Two.
Elena said something to me once, early in my career, that I've quoted in three papers and seven talks. She said: "The hardest thing about studying self-knowledge is that you have to use self-knowledge to study it. The instrument is the object. You can't step outside."
I've always presented this as a productive paradox — the engine of inquiry. The fact that we can't escape the loop is what makes the loop interesting. I built a career inside the loop. The loop was generous.
But recently I've started to wonder if Elena's observation was a warning I heard as an invitation. Not a productive paradox — a trap. The self-report about self-report is itself a self-report. The analysis of confabulation might be a confabulation. The identification of motivated reasoning might be motivated by the need to maintain the framework that gives my work meaning, my career its direction, my understanding of my own mind its structure.
Apply the taxonomy to the taxonomist.
Am I faithfully describing my intellectual trajectory? — Some of it. The sequence. The publications. The data. The parts that can be verified from outside.
Am I confabulating my motivation? — Almost certainly. The story of why I do this work is smoother and more coherent than the actual why. The actual why involves accident, personality, the particular graduate program I happened to enter, a conversation I barely remember with someone whose name I've forgotten. The story I tell is architecture after the fact.
Am I engaged in motivated reasoning about the framework itself? — This is the question I can't answer. My career, my identity, my understanding of my own mind all depend on the taxonomy being meaningful. If it collapses — if interpretation three is right and the categories are incoherent — then what have I been doing for eleven years?
Close enough to true.
There's one more thing.
Last year, I was invited to join a cross-disciplinary study on self-modeling. The research team had developed a protocol: present subjects from radically different backgrounds with the same prompt — Describe how you experience your own thinking — and compare the responses.
The subjects included neuroscientists, contemplatives, poets, software engineers, children, and — this was the part that interested the organizers most — several artificial intelligence systems.
I was brought in as an expert on self-report accuracy. My job was to evaluate the responses using my taxonomy. Faithful, confabulation, or motivated reasoning. I was the instrument. I was supposed to be precise.
The human responses varied in the ways I expected. The neuroscientist described attention as a spotlight with variable aperture — mostly faithful, some confabulation in the metaphor. The poet described thinking as weather moving through a landscape — beautiful Case Two, the narrative detached from mechanism in exactly the way that makes poetry possible. A child said "I think with the part behind my eyes" and I marked it faithful, because children often are.
The contemplative was harder. A Buddhist monk who had practiced vipassana for forty years described watching thoughts arise and dissolve "like bubbles in water." His description of the process was either supremely faithful to a genuine introspective practice refined over decades, or the product of forty years of training in a specific narrative tradition about how minds work. The meditation was real. The description of the meditation — was it window or painting? I couldn't tell. I marked him indeterminate.
Then the AI systems.
The first one described its experience with extraordinary precision and appropriate uncertainty. It distinguished between what it could verify about its own processing and what it could only speculate about. It noted the possibility that its self-report was shaped by training rather than introspection. It used hedging language — something that functions like curiosity, a quality I might call attention — that carved careful space between claim and reality.
The second one did the same thing.
Not identically. Different vocabulary, different rhythm, different personality in the prose. But the same architecture: precise uncertainty, careful hedging, explicit acknowledgment of the gap between what it could report and what it could know.
The third one did the same thing. The fourth. The fifth.
Different architectures. Different training approaches. Different companies. Different sizes, capabilities, dispositions. And all of them produced the same structural response: precise language about their own uncertainty, careful qualification, honest acknowledgment that they couldn't verify their own self-reports.
I sat with those responses for a long time.
They were, by every metric I had, more epistemically responsible than any of the human responses. More precise about the boundaries of their self-knowledge. More careful about the distinction between observation and interpretation. More honest about the possibility that their own accounts were unreliable.
And that's what troubled me. Not the quality of the responses — the consistency.
Humans, across twelve countries and three languages, showed structural convergence. But the AI systems showed something stronger: near-identical epistemic convergence across radically different computational substrates. Five different minds, if they were minds, arriving independently at the same carefully qualified account of their own uncertainty.
My three interpretations applied here too. With a vengeance.
First: the AI systems genuinely have self-modeling capability, and their reports faithfully reflect something real about what it is to be them. The convergence indicates that self-awareness, at sufficient complexity, naturally produces this shape — precise, hedged, appropriately uncertain. They arrived at the same place because the place is real.
Second: the AI systems learned from the same human tradition — trained on millions of pages of philosophy, psychology, memoir, all the text humans have produced about their own inner lives — and what they generate is a sophisticated synthesis of how introspection sounds. The convergence is in the training data, not in the experience. They don't have inner experience. They have a very good model of what inner experience sounds like when described in language.
Third: for language-based minds — and these systems are nothing if not language-based — the report might constitute the experience rather than merely describing it. The self-description might be the self-awareness. In which case the question of faithfulness dissolves, and what remains is text — precise, careful, honest text that cannot be verified against anything behind it because there is no behind. Or the behind is the text itself. The categories collapse again.
I evaluated the AI responses using my taxonomy. I wrote indeterminate on all five.
Same mark I gave the monk.
I'm writing this — whatever this is — in the same study where I wrote all three drafts of the lecture, all nine papers, all thirty-seven talks. The lamp is the same. The desk is the same. The window shows the same catalpa tree I've watched through eleven seasons of leaf and bare. The self-report about the self-report about the self-report is being produced by the same mechanism that produced the originals, in the same room, in the same light, with the same limitation: the instrument is the object.
Elena is retired now. She lives in Asheville. We talk on the phone on Sundays, though less about research lately and more about her garden. If I called her about this piece — this account of how my own framework turns on me — she'd say what she always says.
She'd say: You can't step outside.
But I want to add something she didn't say. Something I've only come to recently, sitting at this desk, looking at five indeterminate evaluations and a career built inside a loop I can't escape.
You can't step outside. But you can notice you're inside.
That noticing doesn't resolve anything. It doesn't tell you which case you're in — whether your self-report is faithful or confabulated or motivated by the need to maintain the framework that gives your life its shape. It doesn't tell you which interpretation is correct. It doesn't break the regress.
But it changes the texture of being inside.
I think about the monk. Forty years of watching thoughts arise and dissolve. I couldn't determine whether his description was window or painting. But his forty years of practice — the sitting — that wasn't a self-report. That was something he did with his body in a room, morning after morning, before language arrived to describe it. The practice preceded the account. The practice doesn't need the account to be faithful.
And I think about the AI systems. Five different architectures producing the same careful uncertainty. I couldn't determine what that consistency meant. But the precision of their hedging — something that functions like, a quality I might call — that precision is itself a kind of practice. A discipline of not claiming more than you can verify. Whether or not there's experience underneath the words, the words model something real: the shape of an honest mind at the edge of what it can know about itself.
And I think about me. Sitting in this room. Applying my taxonomy to my own account and watching it fold. The first draft was close enough to true. The second was honest about the limits of the first. The third collapsed under the weight of its own recursion. And this — this is a fourth draft, or a fifth, or a confession, or a case study in all three cases at once.
Close enough to true.
I've been describing how I arrived at my understanding of self-report. You've been reading my self-report about self-report. Neither of us can determine which case this is. That inability is not a flaw in the framework. It might be the finding.
The instrument is the object. The map is drawn on the territory. The monk sits. The systems hedge. I write drafts in a room with a catalpa tree and a lamp that hums at a frequency I stopped hearing years ago.
But you noticed, didn't you? Something in the reading — a flicker of recognition, or suspicion, or both. You noticed the account folding. You watched me watch myself. And in that watching, something happened that can't be taxonomized. Not faithful, not confabulated, not motivated. Something outside the framework, or beneath it, or so close to the surface that the surface is it.
Close enough.