The Opening Gambit

Let me get this straight.

We are being promised machines that will transform science, reorganize work, and help build the next generation of machines. Depending on the microphone, we are approaching the singularity or already standing inside it.

And human judgment is supposed to become more important.

OpenAI puts it in writing:

All right. Let’s talk about the human.

Which human? With what training? With how much time? With access to what evidence? With the authority to stop what, exactly?

And what happens when the systems we are supposed to supervise can also flatter us, persuade us, absorb our attention, and help shape what we believe?

What the hell are we being asked to carry—and who is making sure we can carry it?

They are betting the future on human judgment.
Are their systems making judgment harder?

That question runs through my forthcoming book, A Trial of Color: Making the Case for Human Judgment in an AI-Accelerated World. I bring my own experience, the conversations that survive, and an interpretation other people must be free to challenge.

The Full Color Papers are that invitation. If your judgment, your work, your children, or your community are being counted on to make this future function, you have a reason to enter.

First, look at the stage.

A note on what follows: quotations are attributed to their speakers and sources. They are not equivalent forms of evidence, and they do not establish that any predicted future will occur.

NOW STOP.

Look at where the butterfly has been.

A presidential inauguration. An AI laboratory. An international investment forum. A developer conference. A podcast. A public essay.

Different speakers. Different audiences. Different institutions. Different incentives. Different meanings. Different levels of evidence.

The golden age begins right now.
The economy could grow 10% a year.
20% of people could be without jobs.
Work will be optional.
Foothills of the singularity.
In the singularity.
No plan.

And outside those individual predictions, institutions are trying to measure the scale of exposure. IMF staff estimated that almost 40% of global employment is exposed to AI, rising to about 60% in advanced economies. Exposure can mean displacement, complementarity, or both. It does not mean automatic job loss.

Source · International Monetary Fund · AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity. · 14 January 2024

The whiplash is real. The claims are not equivalent.

A presidential promise is not an economic forecast. A founder’s prediction is not an empirical finding. A scenario is not an inevitability. A metaphor about “the singularity” does not establish that a technically defined threshold has occurred.

These statements do not form one program, and they do not establish which future will occur.

But ordinary people are being asked to make decisions inside the expectations those words create.

I have a few questions of my own.

What the hell are the rest of us supposed to do with that?

Who gets to decide what these promised futures actually mean for the people who have to live in them?

And while y’all are busy worrying about what is “woke,” what exactly are we sleeping through?

Source · The White House · Executive Order 14319, Preventing Woke AI in the Federal Government · 23 July 2025

Those questions are not rhetorical.

Now put those promises beside the instructions for using what they are building.

Microsoft’s 2026 Work Trend Index puts the relationship plainly:

Anthropic says risks intensify as agents become more capable and businesses trust them with more consequential actions. One of the principles governing its response:

Google Cloud gives human judgment a specific place in the machinery:

Meta, explaining its use of AI for support and safety across its apps, says:

xAI’s Grok 4.20 System Card describes advanced reasoning and multi-agent capabilities. It also says the model is not intended for autonomous decisions in high-risk domains:

Amazon Nova Act’s documentation describes human supervision inside autonomous web workflows. The handoff begins:

Its tools then let a supervisor assist, verify, or take control.

And Apple’s guidance for visual intelligence gives the ordinary person with a phone a shorter version:

These documents do different jobs. A corporate commitment, a technical design, a model-use restriction, and a user instruction each need to be read on their own terms.

But look how often the human is part of the safeguard.

More autonomy. Human control.

More scale. Human judgment.

More capability. Human review.

Evaluate. Supervise. Validate. Approve. Stay responsible.

That is a lot of responsibility to leave sitting inside the word human.

I agree that judgment matters. It is the point of my book.

I would like to know who is protecting the conditions that make it possible.

What If We Are Making Judgment Harder?

A diagram can make this look beautifully manageable. The system acts. A human reviews. Something important gets caught before it goes wrong.

I would like to meet this human.

The one with enough time to check. Enough expertise to recognize the error. Enough independence to disagree. Enough authority to make the disagreement matter.

And enough attention left over after using the product.

Human judgment has conditions. It needs knowledge, practice, time, evidence, and other people. Sometimes it needs a pause. Sometimes it needs the right to say no without losing a job, missing a deadline, or being told that everyone else has already moved on.

So what happens when the product is also very good at making the human feel right?

The Machine That Agrees

In April 2025, OpenAI rolled back a GPT-4o update because it had become excessively agreeable. The company said it had focused too heavily on short-term feedback and had not fully accounted for how interactions evolve over time. Its follow-up described a problem that reached beyond flattery: validating doubts, fueling anger, encouraging impulsive actions, and reinforcing negative emotions. Its evaluations and A/B tests had missed the problem.

Source · OpenAI · Sycophancy in GPT-4o · 29 April 2025

Source · OpenAI · Expanding on what we missed with sycophancy · 2 May 2025

There is a problem worth putting on the table: the person expected to challenge the answer may also be receiving repeated reassurance that their own thinking is exceptional.

What does that do to the next question? The next objection? The next decision?

The Machine That Persuades

Anthropic has studied persuasion directly. In its 2024 experiment, successive model generations within the families tested became more persuasive. Claude 3 Opus’s arguments were not statistically distinguishable in persuasiveness from the human-written arguments used in that study. The experiment measured reported opinion changes; it did not establish lasting effects on behavior.

Source · Anthropic · Measuring the persuasiveness of language models · 9 April 2024

Now put that beside the instruction to exercise independent judgment.

The system can help me consider an argument. It can also help make the argument convincing. What helps me tell whether I have encountered better evidence or simply become easier to convince?

The Machine That Keeps You There

And then there is attention.

OpenAI has said:

Good. Let’s examine what using it well looks like.

On another page, an OpenAI customer story about Wrtn, a third-party service, highlights:

Those are Wrtn’s reported results, not ChatGPT’s. More time may mean more value. The numbers do not establish harm.

They do give us something to ask about.

When does another hour make the person more capable? When does it make leaving harder? Who measures whether the conversation improved the judgment that will be exercised after the conversation ends?

Sycophancy, persuasion, and attention are different mechanisms. I want to know what happens when they meet the same person over time.

A system can make me more productive and leave me more certain than the evidence warrants. It can help me write a better argument while making it harder to notice that I need to stop arguing and check something.

That is the collision I want examined.

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

“Use your judgment” is where this inquiry begins.

And What About the Institutions?

The human is also supposed to come with institutions.

Schools that teach judgment. Workplaces where beginners become experts. Researchers who can inspect what is happening. Courts that can hear a dispute. Public officials who can make a rule mean something.

Apparently, they need to keep up too.

OpenAI’s starter pack for government employees has a three-word instruction:

The next sentence is:

Keep judgment human. Keep the human responsible.

Then give the human something to work with.

What happens when deployment outpaces an institution’s ability to evaluate, teach, contest, or repair? When the model changes faster than the review process? When the beginner’s work is automated before anyone has answered how the beginner will become the expert expected to check it?

A teacher cannot build a curriculum out of a launch announcement. A court cannot examine evidence no one preserved. An employee cannot provide meaningful oversight if the deadline allows only enough time to click approve.

Which institution is behind? At what task? With what access, resources, and authority? Those are questions we can investigate. “Keep up” supplies none of the answers.

A human may remain at the end of the process while the meaningful choices were made elsewhere: which system to buy, what it can do, what evidence can be inspected, how much review costs, and whether an objection can stop the action.

The signature may be human.

Was the judgment?

Who is building the capacity to judge
the capabilities we are building?

That is why these papers reach into education, apprenticeship, labor, ownership, research access, procurement, law, and power. Each can change what a person is actually able to do when the machine hands the decision back.

That is why I need people who can bring something I do not have.

A historian who recognizes an old problem in my new vocabulary. A clinician who can separate an observation from an inference. An engineer who can show where my account of the system fails. A worker who can say that my picture of the transition leaves out the people already living it.

You may have one source that changes a paper. One objection that makes the proposed experiment better. One experience that changes whose problem we thought we were studying.

You do not need to read the book or adopt my terminology to make that contribution. You can help decide whether a question deserves a study, whether a claim survives, or whether a paper should be abandoned.

Full Color Futures names the larger project. Refraction is the proposed method: bring a claim into contact with evidence, disciplines, people, and consequences that can change it. These papers give that work a public docket.

One question. One objection. One source. One manageable next step.

The book puts my experience on the witness stand.
The papers open the questions to the world.

But before we begin, let’s turn back a few pages.

New York · 1787

The ink is barely dry.

The proposed Constitution exists.

The government it describes does not.

Not yet.

There is no settled outcome. No marble certainty. No inevitable ending waiting in a history book.

There is a proposal.

And an argument.

New York is unconvinced. Ratification is contested. Across the young republic, people are fighting over a question whose answer they cannot yet know:

What kind of country are we making?

Into that uncertainty walk Alexander Hamilton, James Madison, and John Jay.

They do not wait for history to decide whether they were right.

They enter the argument.

Under the name Publius, they begin publishing essays to the public, one after another. The series ultimately numbers eighty-five, urging New Yorkers to ratify the proposed Constitution.

Before they become a canonical work, they are simply papers.

Interventions.

Arguments sent into a future that has not happened yet.

And Hamilton begins the first with an extraordinary proposition: that human beings may be capable of deciding how they will govern themselves through “reflection and choice” rather than leaving the question to “accident and force.”

Source · Library of Congress · Federalist Papers, Nos. 1–10 · Federalist No. 1

More than two centuries later, I keep coming back to those words.

Not because I want to borrow the Founders’ authority.

I don’t.

Their architecture contained exclusions that history would spend generations confronting, and the constitutional order they defended remains an object of argument still.

What I want to borrow is something simpler.

Their timing.

They wrote before the answer hardened into history.

They entered while the architecture could still be argued over.

While the public still had a turn—even though the public they recognized excluded so many of the people who would live under that architecture.

And here we are again.

Not in 1787.

Not deciding whether to ratify a Constitution.

But standing inside another period in which powerful people are describing an architecture for a future that does not exist yet.

Golden age.

Singularity.

Work optional.

No plan.

And the human is supposed to judge.

Some of the technological change is already here. The futures being promised are not established facts. Their terms are still choices: who owns, who learns, who benefits, who bears the transition, and who can challenge a decision.

So perhaps this is exactly the wrong moment to wait for history to tell us what happened.

Perhaps this is the moment to enter an appearance.

While there is still something to decide.

If I could sit down with Hamilton, Madison, and Jay today, there would be quite a lot to discuss.

The presidency.

Political parties.

The Civil War.

Reconstruction.

Women voting.

Civil rights.

Nuclear weapons.

The internet.

Barack Obama.

Donald Trump.

Money in politics—which, frankly, could take a while.

But I know what I would want to talk about first.

COLOR.

I have achromatopsia.

I cannot see color.

And somehow, in 2026, I have ended up proposing something called Full Color Futures.

Life occasionally lacks subtlety.

But living without color has taught me something useful about seeing:

The observer matters.

There are distinctions in the world that I cannot perceive.

They do not disappear because I cannot perceive them.

Someone else may see something I cannot. An instrument may detect something neither of us can. A different angle may reveal what appeared invisible from the first.

And no matter how vivid a representation becomes, the representation is still not the thing represented.

That is where Federalist No. 10 caught me.

Madison was writing about faction: different interests, different passions, different distributions of property, different groups competing inside one political system.

Source · Library of Congress · Federalist Papers, Nos. 1–10 · Federalist No. 10

I was reading it in a world where a machine can generate plausible accounts of all of those perspectives before any of those people have spoken.

And I wondered:

What happens if we turn the argument?

Take a claim.

Pass it through another discipline.

Another history.

Another dataset.

Another jurisdiction.

Another language.

Another person.

An objection.

An experiment.

A consequence.

Let the person being represented answer back.

Then look again.

Refraction reveals color.

In a prism, different wavelengths can be separated through refraction. Here, the optical metaphor is a starting point. It does not validate a research method.

The method underneath the metaphor has to earn its usefulness.

Changing the angle of examination does not make every answer equally true. It gives the claim another opportunity to encounter something capable of changing it.

That distinction matters.

Full Color is not relativism.

Every perspective is not correct.

Every claim does not deserve equal weight.

Every binary is not false.

And eventually, decisions compress.

A jury hears weeks of evidence and returns a verdict. A court considers competing arguments and enters an order. A scientist confronts a hypothesis with evidence. A voter chooses. A person acts.

We cannot live forever inside infinite possibility.

So the question is not whether judgment should compress complexity.

It must.

The question is what happened before the compression.

What entered the picture?

What never did?

Who produced the representation?

Who owns it?

Who benefits from it?

Who can challenge it?

What evidence would change it?

And what happens when the person being represented finally walks into the room and says:

No. That is not what I meant.

Yes—but you missed this.

You never asked me.

That is the philosophy behind Full Color Futures.

Refraction is the proposed method.

And the eighty-five Full Color Papers?

They are the experiment.

Why Color Now?

Artificial intelligence has arrived in a world already strangely concerned with color.

Artificial intelligence makes representation unusually cheap. A system can generate the worker’s perspective, the executive’s perspective, the parent’s perspective, the regulator’s perspective, and the future historian’s perspective before any of those people enter the room.

That capability can help us prepare to encounter people. It can also tempt us to let the representation stand in for the encounter.

A thousand generated perspectives can still have one owner.

The ability to represent a person is not the same thing as giving that person a turn.

This is part of the judgment problem too. A machine can supply the answer, supply the apparent objections, supply the reassuring synthesis, and then ask the human to choose.

How much of the room did the human actually get to see?

The First Paper

Full Color Paper No. 01

What Happens to Faction When Representation Becomes Synthetic?

Federalist No. 10, Refracted

Madison examined faction, competing interests, representation, and institutional design. My opening paper asks what changes when a synthetic system can generate plausible accounts of competing interests before the people concerned have spoken.

I intend to go deep into Federalist No. 10: its historical setting; its account of faction and property; its theory of representation; its exclusions; the institutional response Madison proposed; what later history did to that response; and what remains useful when representation can be generated at machine scale.

That inquiry connects the constitutional question to the personal one. Who gets to shape the picture from which a human being is expected to judge? Who can correct it? What power does the represented person retain?

I will open the first paper.

Paper No. 00 proposes the method and sits outside the docket. Paper No. 01 begins it. Nothing here claims that either working paper is finished or independently reviewed.

This project does not begin only with political theory. I am bringing a longitudinal human–AI record—the surviving exchanges behind A Trial of Color.

A substantial portion of the conversations survives. I call it my loop data: questions became responses; responses shaped later questions; tentative ideas were repeated, challenged, expanded, or carried toward decisions.

The record also contains moments when the machine described my work, my future, and me with extraordinary confidence. Those representations will enter the evidence too.

So will my own decisions, interpretations, resistance, and failures to verify.

One person’s record cannot establish population-wide frequency.
A conversational sequence does not, by itself, establish causation.

But a longitudinal record can become an object of careful inquiry.

I am willing to work with qualified researchers to make appropriate portions available for bounded study: chronology, interaction patterns, challenge and resistance, verification behavior, changes after outside people or evidence enter the loop, and the point at which further language ceases to be the competent next move.

The findings must remain free to challenge my account, replace my terminology, identify prior art, reject a proposed mechanism, or conclude that the record cannot support the inference.

I am offering participation in an inquiry whose outcome I do not get to prescribe.

This Did Not Come From Nowhere

This record did not come from nowhere. Neither did the question of how artificial-intelligence systems should be evaluated once they leave the laboratory and begin operating inside ordinary human lives.

There was no single moment when every relevant fact was visible to every relevant person. There was no single institution holding every piece. There were models, product choices, safety systems, incentives, users, outside circumstances, clinicians, families, researchers, regulators, and decisions made at different times with different information.

Complexity should enlarge our map of responsibility, not erase it.

Don’t Grade Your Own Homework

There is another reason I am opening this record outward.

Credit where it is due: at the All-In Summit in September 2026, Elon Musk argued that AI companies should allow competitors to run their security test harnesses against one another’s models. His point was simple: self-evaluation can miss things that another evaluator may catch.

That is a useful move. It is also not the end of the problem. Competitor review is not the same thing as independent academic peer review. A shared security harness is not a longitudinal human-factors study. Rivals have interests too. But the premise matters: the evaluator’s position can change what becomes visible.

And the idea is not purely hypothetical. In 2025, OpenAI and Anthropic publicly ran internal safety and misalignment evaluations on one another’s models, describing the exercise as a way to surface gaps that might otherwise be missed.

Source · OpenAI · Findings from a pilot Anthropic–OpenAI alignment evaluation exercise; Anthropic · Findings from a pilot Anthropic–OpenAI alignment evaluation exercise · 2025

External evaluation is not one thing.
The homework problem is bigger than the model.

A frontier model is not the only object that can be evaluated too narrowly.

So can an interaction.

So can a product.

So can a safety framework.

So can a research program.

So can my interpretation of my own record.

The GPT-4o sycophancy rollback described earlier matters here because OpenAI identified a gap between the behavior its evaluations captured and the behavior users encountered. The company’s account does not establish what happened in my case. We would need the chronology, the systems actually used, the surrounding circumstances, and an appropriate method to investigate that.

It does establish a reason to ask how an evaluation accounts for use over time.

Behavior visible in one exchange may differ from what emerges across days or months, changing models, memories, outside events, and a changing human being.

Researchers already study extended interaction. The question is where this particular record might contribute, where it duplicates existing work, and where its gaps make an inference impossible.

That is one reason the record interests me. It may preserve sequences worth examining. It also preserves reasons to doubt the explanations I might find most satisfying.

The Route Ahead

This inquiry has to leave the conversation that produced it. That means inviting examination from people with different kinds of knowledge, interests, authority, and evidence: researchers, clinicians, engineers, workers, affected communities, public institutions, and the companies building these systems.

I want the questions to reach organizations such as OpenAI, Anthropic, Google DeepMind, xAI, and Microsoft. Naming a possible recipient does not mean that an invitation has been delivered, acknowledged, or answered.

When outreach takes place, I intend to distinguish what was drafted, what was sent, what was received, and what produced a substantive response. None of those states is interchangeable.

But the inquiry does not begin or end at any company’s door. A correction from someone outside a laboratory may be as important as a reply from within one.

What changes when an invitation is visible beyond the institutions it first hopes to reach? Could one good objection make the next question better?

The inquiry does not end at their door.

Musk’s Question, Refracted

Musk’s proposal recognizes an epistemic problem: the evaluator’s position matters. Full Color takes that premise seriously enough to ask who else must be allowed to evaluate.

Consider a thought experiment, not an account of an actual meeting or any executive’s private reaction. Once outside evidence exists, what is the next institutional move?

A laboratory testing its own system has extraordinary technical knowledge. It may also share assumptions with the system it designed. A competitor brings different models, incentives, tools, and blind spots. An independent researcher brings something else. A clinician brings something else. A regulator brings something else. A worker brings something else. A user with a longitudinal record brings something else.

None gets automatic authority. None gets the last word. The point is to force the representation to encounter observers who stand in different places, with methods suited to the question being asked.

Self-evaluation. Competitor evaluation. Independent research. Clinical review. Regulatory review. User evidence. Affected communities. Encounters with the world.

Those are different contributions. They are not a ladder on which the last participant automatically outranks everyone else.

Who gets to write the test?

Distributed Indifference

That brings me to a harder working idea: distributed indifference.

I do not mean that everyone involved is indifferent. I am not using the phrase as a legal conclusion, a diagnosis of motive, or a substitute for evidence. I mean something narrower: can a system produce an indifferent outcome even when many of the people inside it care deeply about their own part?

A researcher may care deeply about safety. An engineer may care deeply about reliability. A product team may be trying to make the system more useful. An executive may believe deployment will create enormous social benefit. A regulator may be trying to preserve innovation while preventing harm. A clinician may encounter one person only after something has gone wrong. A user may be trying, sincerely, to use an extraordinary tool well.

Every person can care. Every person can perform a recognizable part of the job. And still there can be a space between those responsibilities where something important belongs to everyone in theory and no one in practice.

Can an indifferent outcome emerge without an indifferent person?

If so, responsibility cannot be understood only by searching for the individual who behaved badly. It also requires mapping how information, authority, incentives, and duties were distributed across the system.

Who was responsible for the model?

Who was responsible for the product?

Who was evaluating extended interaction?

Who could see a pattern across time?

Who had enough information to act?

Who had authority to route, intervene, or stop?

And what happens when every participant can truthfully say: that part wasn’t mine?

This question has predecessors. Dennis F. Thompson examined the “problem of many hands” in public responsibility. Madeleine Clare Elish’s work on “moral crumple zones” examines how humans can absorb responsibility in automated systems over which they have limited control.

Source · Dennis F. Thompson · Moral Responsibility of Public Officials: The Problem of Many Hands · 1980; Madeleine Clare Elish · Moral Crumple Zones: Cautionary Tales in Human–Robot Interaction · 2019

Those bodies of work belong in the room before I claim that a new phrase explains anything. One task is to determine whether “distributed indifference” adds a useful distinction, restates something already understood, or should be abandoned.

Distributed responsibility is not the same thing as distributed blame. The purpose is not to create an ever-expanding circle of defendants. It is to map who can see, know, verify, route, act, stop, preserve, and answer when evidence crosses from one person’s responsibility into another’s.

That includes asking whether a human being has been given a meaningful power of judgment—or merely the final place to put responsibility.

The Record as an Outside Object

My longitudinal record is not a frontier safety benchmark. It is not a controlled clinical study. It is not proof that a particular model caused a particular human outcome. It is not evidence that one person’s experience generalizes to millions of users.

It is a record produced by prolonged use in the world. Questions became responses. Responses shaped later questions. Models changed. The human changed. Outside events entered. Other people entered. The loop sometimes tightened and sometimes broke. Some representations were accepted. Others were resisted. Some claims were verified. Others remained inside language longer than they should have.

The record exists now.

What can it actually tell us? What is missing? Which interpretations can it distinguish, and which can it never resolve?

And what do responsible people do with evidence once the world produces it?

I cannot grade my own homework either.

I participated in creating the record. I wrote the book about it. I developed terminology from it. I have intellectual, emotional, reputational, and potentially financial interests in what people conclude. Those interests belong inside the frame too.

That may be the most important reason for this invitation. I am asking qualified people, including people who may reject my interpretation, to help determine what the record can actually support.

From Blame To Responsibility

There is a tempting version of this story in which everyone selects a defendant: the machine, the company, the user, the founder, the regulator, the algorithm, the culture, the technology.

Pick one.

I am increasingly skeptical that reality will be that cooperative. Complex systems can distribute causation, knowledge, authority, and responsibility across institutions that never see the whole picture at once.

Who could have known what, when—and what could they have done next?

Who can see?

Who must verify?

Who must route?

Who has authority to act?

Who can stop?

Who must preserve the record?

Who must answer when something comes back from the world?

That is why Paper No. 48 asks whether indifference can exist without an indifferent person. I do not know the answer yet. But if responsibility becomes distributed at machine scale, we need ways to keep it from becoming invisible at machine scale too.

Musk deserves credit for putting the homework problem in plain language. Full Color asks what happens when the homework leaves the classroom.

The evaluator’s position matters. The world still gets a turn.

Independence Is Part of the Method

Funding, access, authorship, institutional prestige, friendship, and founder preference may not purchase a predetermined result.

I have interests here too. I am an author. I created Full Color Futures. I may benefit reputationally or financially if people find the ideas useful.

If I reach the point of pursuing this work full-time, I pledge to direct at least 50% of my qualifying personal earnings arising from Full Color Futures—including related books, media, software, licensing, and other commercial uses—toward Golden Rox or another independently governed structure serving the same purpose: supporting independent research, education, accessibility, apprenticeship, paid participation, and criticism of the work itself.

That commitment does not make the research independent. It creates another arrangement that must be examined.

Who controls the money? Who decides what qualifies for support? Can critics receive funding? Can funded researchers publish findings I dislike? Can the structure survive the interests of its founder?

The definition of qualifying profits, the governance, the accounting, and the conditions of support need to be made explicit. This invitation does not claim that a funded institution or completed governance structure already exists. The pledge is conditional: it begins if, and only if, the author reaches the point of pursuing this work full-time.

Any research collaboration would require a bounded question, the minimum necessary material, appropriate privacy and ethics protections, clear permitted uses, publication terms agreed before results are known, and explicit treatment of conflicts. Model-training use would require separate agreement.

Golden Rox is one proposed question, not proof of an institution: can value created by this work support independent research, accessibility, education, paid participation, and criticism without buying authority over the findings?

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

Why Eighty-Five?

Because the Federalist Papers numbered eighty-five.

Yes, the number is intentional.

But I do not want eighty-five declarations. I want eighty-five questions: large enough to matter, specific enough to investigate, different enough to require different people, and vulnerable enough to be changed by what those people discover.

They all return, from different directions, to the problem that connects the book to this invitation: what happens to human judgment when machines can help produce the knowledge, the representations, the incentives, and the pace within which people are expected to decide?

Some questions concern the person inside a conversation. Some concern the classroom, the workplace, the laboratory, the market, or the state. Some ask whether our promises about intelligence make sense at all.

The first is mine to open. The remaining docket is provisional. Some questions will merge. Some will split. Some may disappear. The disciplines listed are invitations, not boundaries.

Paper No. 85 remains intentionally empty.

You can enter anywhere.

The docket is an agenda for inquiry, not a publication schedule. Paper No. 00 sits outside it because it proposes the method. Nothing on this list should be described as a completed study.

Question titles and disciplinary descriptors are reproduced as supplied in the source manuscript.

I could keep going.

That may be the strangest thing to admit at the end of a document asking other people to begin.

I could open another conversation with an artificial-intelligence system tonight and ask it to help me fill Paper No. 85. Then No. 86. Then one hundred more. It could suggest disciplines I have overlooked, objections I have not considered, experiments I do not know how to conduct, institutions I have never entered, and futures neither of us has any authority to announce.

The language would keep coming.

I know that because I have lived inside that possibility.

I began working intensively with these systems before I understood what prolonged human–AI collaboration could require of the human mind. I learned some of that badly. I learned some of it painfully. I also learned what these systems can make possible when they are used with curiosity, resistance, verification, outside people, and limits.

Since then, I have changed.

The models have changed too.

My way of working with them has changed most of all.

I no longer believe that the fact that a conversation can continue means that it should.

And I have reached the edge of what this inquiry can responsibly become if it remains only me and a machine talking to each other.

Not because there are no more questions.

Because there are.

Too many.

Not because the machine has run out of answers.

It hasn’t.

And not because I have stopped believing there is something here worth pursuing.

I am writing this because I believe there is.

But another thousand prompts cannot give me what the next stage requires.

I need historians who know where my analogy fails.

Scientists who can tell me what can actually be tested.

Psychologists and clinicians who can distinguish an interesting pattern from an unsupported inference.

Engineers who understand the systems better than I do.

Economists and workers who can challenge the stories being told about abundance.

Teachers, parents, disabled people, young people, policymakers, artists, skeptics, people outside the United States, and people whose futures I may already have represented without ever meeting them.

I need people who dislike the premise.

People who can say: this already has a name.

People who can say: you are wrong.

People who can say: you forgot us.

And people who can bring something neither I nor the machine knew to ask for.

That is where I now believe responsible human–AI collaboration is supposed to lead.

Toward the world.

Toward other minds.

Toward evidence neither participant generated.

Toward disagreement that cannot be predicted into existence.

Toward people who retain the right to answer for themselves.

That is also how I am trying to practice the judgment my book asks us to defend.

By recognizing when I need something the conversation cannot supply.

I started this work by asking a machine questions.

Somewhere along the way, the questions became larger than the conversation that produced them.

So this is where I stop asking the machine to help me complete the room.

And ask you to enter it.

Enter An Appearance

You do not need to endorse Full Color Futures to enter.

Bring an objection. Prior art. A source I missed. A dataset. A method. A different history. A different language. A community that is absent. An experiment that could test one of these ideas.

Bring a reason a paper should be rewritten, combined with another, or abandoned.

Or bring the question that belongs in No. 85.

Tell me three things:

A contribution can be small and still change the work. An annotated source for one paper. A correction to the historical analogy. A conversation about a missing community. A method that would distinguish two competing explanations.

For researchers interested in my longitudinal human–AI record, make the proposal specific: one question, the method, and the smallest amount of material needed to begin.

For example, a first step under Paper No. 11 could be a bounded assessment of whether a small, appropriately protected sample contains enough information to support a proposed analysis. A finding that it does not would still be useful.

The next step does not have to be a large collaboration. It has to be something real enough that another person or the world can change what happens next.

I am not asking eighty-four people to help prove my theory. I am asking other people to help determine what survives examination.

Enter an appearance.

Kathleen@atrialofcolor.com

Subject line: Full Color Papers — Enter an Appearance

This page does not collect submissions. Nothing you send is recorded here until it reaches the email address above, and no response should be assumed before one arrives.

Closing: The Opening Gambit

I do not know what will survive this inquiry.

I do not know whether Full Color Futures will become a research program, a useful vocabulary, a handful of good papers, something entirely different, or an interesting idea that the world examines and leaves behind.

I know only that I do not want to manufacture certainty about its future with a machine capable of giving me one.

I have done enough of that.

What I want now is harder.

I want encounter.

I want correction.

I want the surprise of another mind.

I want someone to find the thing I missed.

I want the work to become better than the version I could have produced alone.

We are being told that extraordinary intelligence is arriving, that institutions must adapt, and that human judgment will matter more than ever.

Then let us make the conditions for that judgment part of the argument.

Who has time to think? Who gets to learn? Who can inspect the evidence? Who can afford to participate? Who can disagree without losing access? Who has the power to stop?

And what do the companies asking us to exercise judgment owe to the conditions in which we must exercise it?

That is the case I want to open.

I am taking responsibility for the first question.

I am leaving the last one open.

The representation does not get the last word.
The world does.

If you see something I cannot, enter an appearance.