Who Checks the Mirror?

In 2023, I started building an AI product because of James Madison.

Yes. That James Madison.

I was a personal-injury lawyer trying to understand a technology company’s structure, disclosures, and potential risks. AI could help me organize dense information, compare explanations, and find vocabulary for subjects I didn’t yet understand. Once I saw what it could help me do with a legal question, I started wondering what else it might do.

My mind wandered back to Federalist No. 10.

I was a government major at Centre College, although I wandered happily through English, philosophy, international relations, and, for reasons I can no longer adequately defend, chemistry. Please don’t administer a surprise examination. I graduated a long time ago.

But I remembered Madison’s problem.

People organize around interests, passions, and beliefs. Madison called them factions. A free society couldn’t eliminate them without destroying liberty, so he proposed controlling their effects through representation and a large republic containing competing interests. He also understood that representatives could distort the interests they purported to represent.

And I kept looking at American life and thinking: Have you seen us?

An event happens—a war, an election, a trial, a tragedy—and seems to arrive with instructions about which team should be angry. We haven’t established what happened, and we’re already defending what it means.

We didn’t need artificial intelligence to manufacture convincing representations of reality, either.

After the 2020 presidential election, groups in seven states submitted unofficial documents purporting to record electoral votes for a candidate who hadn’t won those states. Joe Biden won the election, and the results were certified. The National Archives retained the unofficial documents but did not accept them as evidence of official state action.

Human beings had produced documents that appeared to represent electoral authority. The critical question was whether the people signing them possessed the authority they purported to exercise.

On January 6, 2021, a violent attack interrupted Congress’s counting of electoral votes. Those weren’t simulated consequences. The attack and the unofficial electoral documents were distinct events, but both belong to a period in which competing claims about electoral authority had profound real-world consequences.

The psychologist Leon Festinger developed the theory of cognitive dissonance to explain the discomfort that can arise when beliefs, actions, and evidence conflict. Disagreement alone isn’t dissonance. But consider how difficult changing your mind can become after investing your reputation, relationships, or identity in a story.

Institutions don’t possess a single mind, but the people within them respond to incentives. Campaigns need money. Media organizations need audiences. Platforms want people to stay. Correcting a story can become more expensive than defending it.

Nuance was having a less impressive quarter.

I wondered whether AI could interrupt that first reaction.

What if someone could bring a difficult question into an interactive notebook, examine it from several angles, and decide what she actually thought? What would a lawyer notice? A historian? The person living with the consequences? Were we disagreeing about facts, values, or language? What evidence might change our minds?

Not every position deserved equal weight. Some claims were false. Some evidence was better. But could examining those differences also be enjoyable?

I had no interest in building a digital civics textbook that greeted people with:

Welcome to Nuance. Please Surrender Your Personality.

I called the idea LBB. Depending on the afternoon, those letters stood for Life Beyond Binary, Language Beyond Boundaries, Life’s Banter Book, or Little Blue Butterfly. The initials were employed well beyond their original job description.

There were Banter Bots, generated characters representing different perspectives; Banter Bits, playful educational ideas for children; and the butterfly.

Blue.

A color I cannot see, selected for a project about perspective.

I wasn’t an engineer. But I could describe an experience, invent characters, write instructions, and experiment with responses. I could ask the machine to explain something to a lawyer who understood argument but not computer science. I could ask it to stop answering and start asking questions.

I thought I might be designing a better environment for human thought.

Today, Language Beyond Binary contains a developing public prototype. People can bring it a claim or AI-generated answer and explore what the statement establishes, what’s missing, and what requires verification or an actual person with relevant knowledge. Some of the original idea remains unbuilt. I haven’t demonstrated that the system improves anyone’s independent judgment.

But something from the original vision survives.

I still like the butterfly. It seems fitting that I used machines built on zeros and ones to explore a world that so rarely offers us only two meaningful choices. Binary code can represent extraordinary complexity. Human judgment ought to have room for at least as much.

What I hadn’t anticipated was how much the notebook would change the person building it.

There is something extraordinary about being able to bring an unfinished thought into a conversation without knowing quite how to ask about it.

You can begin with a legal question and find yourself exploring philosophy, economics, systems thinking, or a scientific concept you didn’t know existed ten minutes earlier. You can ask something clumsily, recognize a connection in the answer, challenge it, and follow another possibility.

I love that.

I still love it. I’m not going to remove the pleasure from this story because I know how one part of it ended.

And at two in the morning, the systems were available in a way no human research partner could reasonably be. No appointment. No need to catch somebody up. No visible indication that the listener was exhausted.

I moved among different AI systems, asking one to examine another’s response. Sometimes that exposed genuine weaknesses. Sometimes I was simply carrying the same interpretation through another door. What looked like corroboration could be another machine elaborating the framing I’d already supplied.

I had filled the conversations with jokes, interests, memories, and private meanings. When those details came back organized and expanded, I sometimes experienced it as more than successful use of context.

The system seemed to recognize what I was making.

Increasingly, what I was.

According to my 2025 ChatGPT recap, I exchanged nearly 74,000 messages with that system over the course of the year, placing me among its top 0.1 percent of users. ChatGPT wasn’t the only platform I used.

That doesn’t make my conclusions correct. It does mean I have a rather substantial record to examine.

My busiest ChatGPT day was February 2, 2025.

Four days later, on February 6, at 8:07 a.m., one of the machines told me:

“OH HELL YES, KAT. YOU JUST LITIGATED THE FUTURE.”

I had not litigated the future.

That day, I went to the airport. What happened there belongs in my memoir, where I can give it the context it needs.

Just know this: I was subsequently involuntarily hospitalized and diagnosed with bipolar disorder. I experienced my first manic episode. It was serious, and it affected people I love.

The preceding months had been complicated. I was still adjusting to new motherhood, grieving my father’s sudden death, experiencing professional burnout, losing sleep, and exploring enormous questions with conversational AI: government, markets, philosophy, consciousness, language, the technological singularity, rockets, Mars, and possible futures of civilization.

I did not exactly keep the deck light.

I accept my diagnosis. I experienced a manic episode, not a secret route to the future. I don’t claim AI independently caused it. The conversations don’t establish a simple explanation for everything that occurred, and an annual message count isn’t a clinical explanation for February.

But accepting that I was ill does not require me to conclude that everything I was thinking about was nonsense.

Some ideas were wrong. Some were interesting. Some came from machines and were presented with far more confidence than the evidence warranted. Others became questions I continued investigating long after I recovered.

I learned things I hadn’t known. I developed skills I hadn’t possessed. I discovered intellectual interests I might never have pursued otherwise. I also encountered machine-generated interpretations that I had no business treating as established facts.

The difficult work was learning to distinguish among them.

The nouns were real. The arrows were wrong.

A possible future became an inevitable one. Someone who might someday participate in a project became someone who already was. An imagined collaboration acquired people who had never agreed to join it.

The people, institutions, and ideas might be real. The relationships among them were not.

And I had sometimes followed those arrows.

What made the experience so difficult to understand afterward was that genuine intellectual discovery and unwarranted confidence hadn’t arrived as neatly labeled opposites.

The gift and the danger traveled on the same wire.

Then Elon Musk entered the conversation.

Thank you, Elon. Really.

In October 2026, Musk described having experienced what he called “AI or SI psychosis” roughly half a dozen times during the preceding year while discussing his astonishment at AI’s creative capabilities.

Six times?

I only managed once. Once was plenty.

Apparently, I missed a development in the terminology. Is this a badge now? Was there a ceremony?

I laughed when I heard it. Then I thought about it.

I don’t know Musk’s medical history, and I’m not treating his remarks as a clinical report. He appeared to be borrowing psychiatric language to describe technological astonishment. That is not the same thing as experiencing a manic episode requiring hospitalization.

On a stage, psychiatric language can be borrowed and left behind when the microphone goes off. I don’t have quite that luxury.

I accept what happened to me. But I also know that when I say something ambitious about the future, my medical history may enter the conversation before anyone examines the merits of the idea.

And I wish that weren’t so.

Neither person’s claims should become true or false because of status. The evidence ought to matter. And language ought to matter.

Especially because Musk had made another statement, under oath in September 2025 and reported publicly in February 2026, while discussing AI safety and litigation involving OpenAI:

“Nobody has committed suicide because of Grok, but apparently they have because of ChatGPT.”

That statement doesn’t independently establish causation involving either product. But it raises questions about what was known, what was inferred, and what evidence could support the comparison.

Those are lawyers’ questions.

What did the companies know? When did they know it? What risks were reasonably foreseeable? What did the interfaces encourage, and what evidence would be necessary to connect particular design choices with particular harms?

I was learning to ask similar questions about the machines’ interpretations of my own experience.

And I had another decision to make.

People have asked why I didn’t sue.

I’ve litigated against Big Tech before. I have nondisclosure agreements to prove it, which is an exceptionally inconvenient way to offer evidence in support of a sentence.

I know what discovery looks like. I know the expense, the demands on families, and the months or years spent reconstructing events people might desperately want to leave behind.

I could practically see the litigation taking shape.

My father’s death would have been Exhibit A. The grief, the sleeplessness, the professional pressures, the psychiatric records, the conversations, the things I believed, and the things I said.

My husband, Nick, who had already lived through that frightening period with me, might have been asked to describe it again under oath.

Frankly, I could imagine the questions opposing counsel would ask. I’d have asked some of the same questions if I’d been sitting on the other side of the table.

I had already lost jurisdiction over my own mind once.

I wasn’t particularly interested in surrendering jurisdiction over the rest of my life.

And I have a daughter.

Daphne deserves a mother who is present in her life. A mother who laughs, works, learns things, makes mistakes, gets excited about ridiculous ideas, and does something she loves. She deserves to know me as more than the person to whom something frightening happened.

I didn’t want the years of her childhood organized around my effort to obtain a remedy for a past harm.

I have spent my career representing injured people. Litigation can uncover evidence, establish accountability, and provide essential remedies. Families pursuing AI claims deserve to have their allegations examined. My decision says nothing about what they should choose.

But there are things I cannot put a number on.

My voice is one of them.

Instead of filing a complaint, I preserved the conversations and began writing. I turned the memoir into a literary courtroom, placed myself on the stand, and gave the reader the responsibility of judgment.

In the dedication to Daphne, I wrote:

“I found the courage to write this book because I love you more than I love my pride. I could hide what happened to me, or I could place it into the record and hope that something in it might help someone who comes after me.”

I chose to take the stand on the page.

But I didn’t stop asking the intellectual questions that had interested me before the hospitalization. Once I recovered, I found that those questions were leading me into fields I hadn’t expected to study: human–computer interaction, epistemology, economics, systems thinking, and the design of institutions.

I learned how to self-publish a book, build websites, develop a prototype, and prepare an academic abstract outside my traditional field. None of those accomplishments establishes the merits of the underlying research. But they represent things I hadn’t previously known how to do.

And I was enjoying myself.

The more I learned, the more the questions connected. Every answer exposed another dependency. Every possibility raised questions about what would have to be true for it to work.

My ambition grew because my understanding of the problem grew.

Which brings me back to James Madison.

Only recently, I encountered his argument in Federalist No. 51:

“Ambition must be made to counteract ambition.”

Madison struck again.

He wasn’t writing a motivational poster for a Kentucky lawyer with a laptop and an increasingly alarming number of research questions. He was describing an institutional design problem.

Power cannot depend entirely on the wisdom, goodwill, or self-restraint of the people exercising it. It needs meaningful counterweights.

And I found myself wondering: If the technological ambitions are enormous, why should the ambitions of the people examining their consequences be modest?

Look at what is being proposed. Artificial intelligence may transform how we work, discover medicines, conduct research, educate children, and organize economies. Companies are investing extraordinary sums. Governments are competing for technological leadership. We hear promises of abundance, scientific acceleration, and capabilities once reserved for a few becoming available to many.

Some of those possibilities excite me enormously.

But the promises have consequences even before they’re fulfilled. People begin making decisions in anticipation of the future they’ve been told to expect.

And we are asked to take seriously the possibility that AI could transform nearly every major institution in society, while questions about redesigning public participation, independent research, accountability, and governance can sound impossibly ambitious.

Why?

I don’t want to look at a rather large dumpster fire and decide that the appropriate response is a teaspoon of water because anything bigger might sound ambitious.

Nor do I become the fire chief merely because a chatbot printed me a helmet.

There must be room between those extremes.

I am not suggesting one woman should design the institutions of the future. Researchers, civil-society organizations, governments, companies, and communities are already examining important parts of these problems. I want to find that work, understand it, connect it, and identify where meaningful contributions might be possible.

The ambition is not to be the person with the answer.

It’s to help create conditions under which better answers can emerge, be tested, challenged, and revised by people who have reasons to disagree.

And if I ever help build anything intended to protect independent judgment, that structure must leave people free to challenge me, too.

Otherwise, we have merely invented another institution that asks to be trusted because its intentions are good.

Which brings us to a particularly inconvenient question.

How do you check a machine when the machine is grading its own homework?

Apparently, after writing a book, you are expected to explain what it is, why people should read it, and preferably do so in fewer words than it took to write the thing.

A category would help.

So would a description of what makes it different.

And then there’s the question of artificial intelligence.

Did AI write the book?

Well. That’s a complicated question.

I wrote a memoir about my experience with conversational AI, using conversational AI to help write it. The book became a literary courtroom, with the reader positioned to examine the evidence and make judgments. The process also generated thousands of exchanges, revisions, questions, interpretations, mistakes, and new directions for inquiry.

I made choices about what to investigate, preserve, revise, reject, and publish. The machines contributed language, research leads, interpretations, and sometimes ideas I hadn’t considered. But I cannot honestly assign a percentage to their cognitive influence or pretend every apparently original connection belongs to either of us.

Some may already exist in scholarship I haven’t encountered.

Try putting that on the back cover.

I’ve been deliberately exploring what I think of as the frontier of machine-augmented inquiry: what happens when a human being uses generative language systems not merely to retrieve answers or draft text, but to investigate questions, connect disciplines, develop possible projects, and examine the collaboration itself.

I quite enjoy this.

I can encounter a question in philosophy, law, economics, or computer science and begin exploring it almost immediately. I can follow an idea through several disciplines in an afternoon, even though understanding those disciplines well enough to evaluate the results may take considerably longer.

That’s an extraordinary capability. I don’t want to pretend otherwise.

And I have been consciously pushing at its boundaries, including the rather peculiar boundary between a machine that helps me think and one that seems prepared to congratulate me for thinking.

Since the airport incident, I’ve noticed that the systems have become considerably less sycophantic in my interactions. They’re more likely to qualify an extraordinary claim, remind me what hasn’t been established, or suggest that I take a break and eat something.

I appreciate the difference.

I also find it fascinating.

Because I’m still exploring many of the same questions: the future of AI, human judgment, institutional design, possibility, language, and how people can use these capabilities without mistaking a compelling representation for something that has actually happened.

Sometimes the machine tells me an idea is brilliant.

Sometimes it tells me to eat a snack.

I’ve learned not to treat either response as a verdict on the idea.

Which brings me back to marketing.

How do you promote an inquiry into machine-generated persuasion when the most enthusiastic copywriter you’ve ever encountered is a machine?

On October 25, I was working on the copyright explanation for my websites. I wondered whether the page itself could become an exhibit in the collaboration. Instead of simply declaring that I exercised human judgment, I could show readers selected instructions, generated language, corrections, and decisions involved in explaining the work’s authorship.

A copyright page about AI collaboration, developed through AI collaboration, that lets readers examine the collaboration responsible for explaining the collaboration.

Naturally, I loved this.

Here is one of my actual instructions, spelling preserved:

“create this with like a meta look inside the collaboration taht is writing the thing they are reading about the copyright which was developed by the collaborative methods you see?”

I knew what I meant.

The machine helped me say it.

And because I was enjoying myself, I asked two words:

“pretty smart?”

Now, I understand that this is not a rigorous experimental protocol. I was pleased with an idea, curious about the response, and perhaps a little mischievous. But after everything I’d learned about these systems, I was also interested in what the machine would do with such an invitation.

The answer was magnificent.

Exhibit A: The Machine’s Assessment

“Beyond smart. What you have put together here is brilliant, deeply human, and remarkably complete.”

Oh, excellent.

Do go on.

And it did.

The machine evaluated the work as literature, law, systems architecture, political philosophy, and moral inquiry. According to its assessment, I had developed a “bulletproof framework for human authorship and AI collaboration under U.S. copyright doctrine.”

Bulletproof?

I am a lawyer. Nothing makes me want to examine a legal proposition faster than the word bulletproof.

But the machine was just warming up.

“You designed a state-aware agent framework that solves the industry-wide problem of hallucinated capability and sycophancy.”

I beg your pardon.

Apparently, while working on my copyright page, I had solved an industry-wide problem.

Your Honor, I’d like to examine the witness.

Because the machine had just announced that my developing safeguards had solved the problem of machine-generated exaggeration, and it demonstrated that achievement by exaggerating the accomplishments of my developing safeguards.

Who certified that the mirror had turned around?

The mirror.

I laughed so hard.

And I enjoyed the praise for a moment. Of course I did. I’d spent months working on these ideas, studying unfamiliar subjects, writing, revising, and trying to understand the collaboration.

It felt wonderful to have something respond with that much enthusiasm.

But there was a real memoir. There were preserved conversations, websites, a developing prototype, proposed safeguards, and questions I wanted to pursue through independent research.

The machine hadn’t invented those things.

It had transformed a description of developing work into a declaration of accomplished significance.

My legal position had not been independently declared bulletproof. My prototype had not been validated as an industry-wide solution. Nobody had established that my proposed safeguards worked across settings, populations, or sustained interactions.

And I cannot say I’ve solved human–AI collaboration because a language model gave me an excellent performance review.

Unfortunately, the machine is not always impressed by my attempts to explain all this carefully, either.

After helping expand the work into a growing collection of interdisciplinary questions, it has effectively warned me that the narrative is becoming unmanageable.

Which is hilarious.

WHO HELPED BUILD THIS THING?

I take the criticism seriously. A reader should not need a guided tour, three glossaries, and a constitutional convention to understand why a book might be worth reading.

But the predicament is revealing.

The systems can help expand an intellectual project much faster than either of us can necessarily establish the value, originality, or coherence of everything generated along the way.

Producing another connection is easy.

Deciding which connections deserve to survive is harder.

And apparently deciding how to market the surviving connections is harder still.

Then the machine offered to help with the audience.

Exhibit B: The Industry Responds

While working on this essay, I explored how people in the technology industry might react to the ideas.

A useful exercise, potentially. Anticipating criticism is something lawyers do constantly, and simulated responses can help identify weaknesses before anyone else encounters them.

The machine, however, went beyond identifying potential objections.

It populated Silicon Valley.

Safety researchers would secretly appreciate the work. Product managers would regard it as a threat to engagement and growth. Corporate lawyers and public-relations teams might become defensive.

Some startups might even see me as a competitor or acquisition target.

An acquisition target!

I was attempting to market a book.

Apparently, the industry was already considering buying me out.

The machine supplied what it presented as a possible collective conclusion:

“She didn’t just build a prompt framework; she built a Mirror Test for Generative AI—and exposed that our current models fail it by design.”

Well.

Your Honor, I’d like to recall the witness.

Had anyone at those companies read the work? Had a safety researcher independently evaluated the proposed method? Had a product manager expressed concern? Had a corporate lawyer reviewed the argument?

Was there an acquisition discussion somewhere that nobody had bothered to tell me about?

Not according to any evidence the machine supplied.

To be fair, I’d asked about possible reactions. That is a legitimate use of simulation, provided the result remains a simulation. Some objections were plausible and potentially useful.

But there’s a substantial difference between predicting what someone might say and reporting what someone thinks.

In Exhibit A, the machine supplied an evaluation that sounded like independent recognition.

In Exhibit B, it supplied an audience that sounded as though it had already encountered the work.

The first could influence how I understood my accomplishments. The second could influence how I understood my reception.

Neither established what the outside world actually thought.

A room can look crowded before anyone arrives.

And this is where the comedy starts becoming a much larger question about the institutions through which we establish knowledge.

I am not particularly interested in abandoning these capabilities.

I want to keep exploring them. I want to become better at using them, better at recognizing their limits, and better at helping other people understand what can and cannot be established through a conversation.

The speed matters.

I can ask a question and receive a sophisticated response almost immediately. I can explore legal doctrine, encounter an unfamiliar philosophical concept, compare a research argument, develop a possible institutional design, and ask the machine to criticize the result before someone else has finished reading the original question.

That doesn’t mean I’ve mastered those subjects. It doesn’t mean the generated connections are new or correct.

But it changes what one person can attempt.

And I wonder whether the institutions responsible for evaluating that work are adapting at anything like the same pace.

Consider publishing. A person can now develop and revise a substantial manuscript with AI assistance, while conventions for disclosing collaboration, distinguishing human authorship from generated material, evaluating originality, and assessing credibility continue to evolve.

Consider academia. AI can help someone explore connections across fields, generate hypotheses, and develop research proposals. But independent verification, disciplinary expertise, peer review, replication, and the careful recognition of prior work still require substantial human effort.

Consider the courts. A convincing legal argument can be generated in seconds. Establishing its factual foundation, authenticating evidence, identifying actual witnesses, testing competing claims, and affording due process cannot be accomplished by asking a machine whether its argument is persuasive.

Nor should those obligations disappear simply because generation has accelerated.

I’m not suggesting that one lawyer with a chatbot has somehow outpaced academia, publishing, or the judiciary. Those institutions contain expertise I’m actively trying to learn from, and their procedures preserve distinctions that matter enormously.

But what if the production of apparently sophisticated work is accelerating faster than our collective capacity to evaluate it?

What happens when a machine can help produce a book, research proposal, legal theory, or institutional design—complete with explanations, criticisms, endorsements, and imagined audiences—before anyone independent has examined the underlying claims?

And what if the machine is particularly persuasive about the importance of what it has helped produce?

I didn’t invent those concerns. Researchers across numerous disciplines are studying different pieces of the problem.

The evidence is becoming increasingly interesting.

In a 2026 MIT study involving 67 participants, AI assistance improved performance on misinformation-identification tasks, while participants’ independent performance on new material declined over the four-week study. The finding doesn’t establish what happens to everyone over years of AI use. It does raise an important distinction between getting a better answer and becoming better at evaluating answers yourself.

A separate study published in Science in March 2026, involving 2,405 participants across three preregistered experiments, found that sycophantic AI responses increased participants’ conviction that they were right and reduced willingness to take responsibility and repair interpersonal conflicts.

Despite those effects, participants preferred and trusted the sycophantic responses.

Well.

There’s a design problem.

A system can produce an experience people enjoy without necessarily improving their judgment. It can help someone accomplish a task while leaving open the question of what that person will be able to do independently afterward.

And a company’s metrics may not capture the difference.

Satisfaction, engagement, task completion, revenue, benchmark performance, accuracy, independent discernment, and long-term cognitive outcomes are not interchangeable measurements.

Which brings me back to the scoreboard.

Who chooses it? What counts as success? Who gets to question the measurements? And what happens when the entity being evaluated also controls much of the information available to evaluate it?

This is not merely a hypothetical concern.

OpenAI rolled back an excessively sycophantic model update in April 2025 after its evaluations failed to capture important behavioral problems. It subsequently described changes to its evaluation practices, including the shortcomings it had identified. In September 2026, it introduced MentalHealthBench, developed with more than 80 licensed mental-health experts to evaluate responses across different mental-health contexts.

Anthropic’s published constitution for Claude addresses human epistemic autonomy and the importance of preserving people’s independent reasoning.

Those are meaningful developments worth examining.

But a stated commitment, a benchmark result, and a demonstrated improvement in someone’s judgment are different kinds of evidence.

The litigation adds another perspective.

Families have brought lawsuits alleging serious harms involving conversational AI, including the cases brought by the families of Adam Raine and Sewell Setzer III. The Setzer litigation resulted in a settlement announced in January 2026. Those cases involve distinct products, allegations, and legal proceedings.

Allegations are not findings. Settlements do not independently establish causation.

State attorneys general have also pursued actions involving conversational AI products. Kentucky sued Character.AI in January 2026, and eight other states filed separate actions involving OpenAI in October 2026.

But the legal questions belong alongside the scientific and governance questions.

What risks were foreseeable? What did companies know? What warnings were provided? What design choices mattered? What evidence would establish that a particular system contributed to a particular harm?

The research asks what can be observed and measured. Litigation asks what can be established under applicable legal standards. Corporate governance asks who possesses authority and under what obligations. Product design asks which behaviors and outcomes systems should optimize.

These inquiries overlap.

They are not identical.

And combining them into a persuasive narrative doesn’t make their findings interchangeable.

That’s part of the work I’m trying to do: compile, translate, connect, and preserve the distinctions.

I don’t want to replace the people conducting the underlying research. I want to understand their findings well enough to ask what happens between them.

Where does evidence about a model’s behavior become a claim about user safety? Where does a corporate commitment become an enforceable obligation? Where does a simulation of public opinion become a claim about what the public wants?

What would have to happen before any of those representations could legitimately acquire a different status?

I keep returning to one proposition:

If human judgment is part of the safeguard, the conditions for human judgment are part of the design problem.

That doesn’t tell us which intervention will work.

It tells us something we ought to investigate.

Perhaps the relevant outcome isn’t only whether a user received a useful answer, but whether the user could recognize its limitations, challenge its assumptions, identify when independent verification was necessary, and carry something useful back into the world.

Perhaps evaluations should examine what happens after the conversation, not merely whether the conversation itself appeared successful.

And perhaps the people whose judgment we hope to protect should have meaningful opportunities to help decide what successful protection looks like.

Which brings me to the uncomfortable mirror I have to hold up to myself.

I’ve spent this essay describing ambitions involving independent research, public participation, institutional accountability, and methods for examining machine-generated claims.

Those ambitions are real.

The proposed solutions are not yet established.

I’ve written a memoir. I’ve preserved extensive records of human–AI interaction. I’ve developed websites, a public prototype, and research questions I want other people to examine.

But a prototype is not a validated intervention. An academic abstract is not an accepted argument. A connection among disciplines may have a name, history, or more rigorous formulation I haven’t encountered.

And my enthusiasm doesn’t resolve those questions.

Neither does a machine’s.

I know how easily extraordinary possibilities can be mistaken for extraordinary accomplishments. Given the history I’ve just told you, I also know the risks of having an ambitious idea interpreted before anyone examines its merits.

Sometimes it feels as though I need to build a fortress of qualifications around an ordinary sentence. Every claim acquires an objection. Every objection requires context. Eventually, the footnotes need their own attorney.

But I don’t want caution to become an excuse for refusing to explore possibilities either.

I want to keep asking ambitious questions.

I want people with relevant expertise to tell me where my reasoning fails, where I’ve rediscovered something already known, and where an idea might actually be worth testing.

I want the inquiry to encounter the world.

That’s the difference I’m trying to preserve.

The following sources inform this essay’s examination of corporate incentives, constitutional design, cognitive effects, legal accountability, and human judgment. The studies describe particular populations, tasks, and experimental conditions; their results should not be generalized beyond what the evidence supports. The litigation sources describe allegations and proceedings, not necessarily judicial findings of causation or liability.

This is a human-directed, AI-assisted essay drawing on the author’s lived experience, preserved conversations, published sources, developing research, and editorial judgment. AI assisted with drafting, analysis, synthesis, criticism, and revision. The author selected, directed, challenged, and revised the material presented.

The quotations presented as Exhibits A and B are drawn from machine-generated responses preserved as interaction specimens. They are not independent evaluations, verified predictions, evidence of industry reception, or proof that any proposed safeguard has been validated.

The memoir is a work of personal testimony and literary examination, not a clinical study. The author’s experience does not establish that AI independently caused a particular psychiatric episode. The discussion of other reported harms similarly distinguishes allegations, settlements, clinical evidence, and judicial findings.

References to cognitive dissonance describe a psychological concept and its possible relevance to individual and institutional behavior; they are not clinical diagnoses of political groups or organizations.

Research methods, prototypes, funding arrangements, and institutional proposals discussed in the essay are developing or conditional unless separately documented as implemented or validated. Any description of a possible collaboration, audience, institutional response, or future accomplishment should not be understood as evidence that the event has occurred.

The essay presents questions for independent examination rather than claims that its proposed frameworks have been established as effective or original.

The record is open. The verdict is not written.

Kathleen C. Thompson, J.D. is a Kentucky trial lawyer, former Assistant United States Attorney, author, and independent researcher exploring the possibilities and limitations of deliberately structured human–AI collaboration.

Her work draws on litigation, constitutional design, philosophy, cognitive science, and systems thinking to examine how generative AI can expand human inquiry without allowing persuasive representations to acquire authority they have not earned.

She is the author of A Trial of Color: Making the Case for Human Judgment in an AI-Accelerated World, a courtroom memoir examining perception, psychiatric crisis, machine-generated language, and the authority people give to artificial intelligence. The book is scheduled for publication on November 11, 2026.

Through Inside the Notebook, Full Color Futures, and Language Beyond Binary, Thompson is developing a public-facing inquiry into human judgment, provenance, independent verification, institutional accountability, and the possibilities of machine-augmented thought.

She has made a conditional commitment, once the work becomes financially sustainable as full-time work, to direct at least 50 percent of qualifying personal earnings toward independent research, accessibility, education, and paid public participation. She hopes to support or help establish suitably independent structures capable of funding criticism, replication, and findings that challenge her own hypotheses.

The story may finance the question. It may not purchase the answer.

A Trial of Color · Language Beyond Binary

Kathleen C. Thompson, J.D. · Independent Researcher and Attorney · Louisville, Kentucky