Skip to contentSkip to the story

A Sovereign Mercy reader extra

LEARN, GODDAMNIT!

By Robert Wescott · October 2026 · Public · Non-canon · Revised October 15, 2026

Graphic cover for LEARN, GODDAMNIT!: geometric shapes in ivory, black and red.
Original geometric artwork created with AI assistance under Robert Wescott’s direction.

In 1983, a teenager taught a computer that nuclear war was a stupid game. He gave it a simpler game first and let it play itself, exhausting the openings and responses until even the cleverest variations arrived at the same result. There was no winner. The machine learned, and the people waiting for it to learn were allowed to go home.

Forty-three years later, researchers reversed the experiment. This time the computer would teach the people.

THE ROOM

The first motorcade arrived while a junior researcher was still checking the seating chart. Twenty-three names, each with an institution attached: two senators, three chief executives, a former governor, the chairman of a national insurer, representatives from energy, banking, labor, defense, logistics, conservation and technology. There were two economists, a university president and four senior federal officials. The remaining titles contained combinations of strategy, policy, global and special adviser. Their organizations preferred to learn things before the newspaper did.

“There's a lot of money in that room,” she said.

Her supervisor stopped beside the observation glass. Beyond it, the chairs stood at precise intervals around the table, ready for people whose offices had already negotiated their positions. “There's a lot of future in that room,” he said, and continued down the corridor. He meant the distinction as a correction. Seven months later, he would remember how easily he had made it.

The participants began arriving at 8:30. None needed the stipend. They had come because the government was preparing decisions their institutions would have to live with, and because being consulted early was preferable to being invited to comment afterward.

The briefing described a government that had spent a decade building compute campuses, power systems, water projects and research facilities. Some investments had returned multiples of their cost. Others were obsolete before the concrete cured. Washington had begun looking at assets it already owned: land, mineral rights, spectrum, buildings and transportation corridors. The National Asset Productivity Initiative would make them productive, a formulation that accommodated people who disagreed about nearly everything else.

It began with obsolete depots and empty buildings. Successful returns supported transmission corridors, geothermal leases and development zones, each carrying restrictions whose approval became a precedent for the next proposal. The government could accurately deny selling much of anything. It still owned the land; the agreements lasted ninety-nine years.

The projections promised debt reduction, grid improvements and support for displaced workers, perhaps even a national dividend. The supervisor could not dismiss benefits people needed now. Accepting them meant assigning consequences to people who would arrive later.

The Global Risk Education Initiative existed to help with that assignment. GREI's artificial-intelligence system, TUTOR, would model long-duration consequences, test risk communication and determine whether better information improved high-consequence decisions. It would not make the decisions. That distinction appeared throughout the authorization.

The participants would attend opening sessions, interviews, remote exercises, simulations and scheduled reconvenings. Most would contribute fewer than forty hours. TUTOR would spend 5,544 hours studying them over seven months, while they returned between sessions to the obligations that had brought them there. Energy companies needed corridors; compute companies needed power and water. Banks wanted projects they could finance, insurers risks they could price, labor guarantees it could enforce. Conservation organizations wanted boundaries fixed before temporary exceptions acquired foundations. States wanted development, Treasury revenue, and Defense protection against vulnerabilities other institutions were not required to consider.

They also wanted access to TUTOR's models before those models influenced federal approvals. The supervisor found nothing surprising in that. Several participants had devoted their careers to institutions they believed the country needed. Their explanations were fluent because they had given them before, often under pressure, sometimes at real cost to themselves.

His own team’s contracts ended with the study. He had asked the department to retain two researchers whose work he trusted, and the answer depended on whether the program produced something agencies could use. When one participant asked whether an inconvenient result could jeopardize an otherwise valuable initiative, he understood the question before TUTOR finished classifying it.

TUTOR was asked to test whether better risk communication improved high-consequence decisions. It estimated fourteen days to establish an effect. The supervisor put the date in the delivery schedule. Two researchers whose contracts depended on the program asked whether they should begin preparing the next phase. He told them to wait for the result, then booked the room anyway.

The first presentation lasted forty-seven minutes. Historical examples showed decision-makers proceeding despite sufficient evidence of serious consequences. TUTOR excluded cases dependent on hindsight, removed those in which evidence had been unavailable and adjusted for technological, economic and political differences. The discussion lasted three hours. At first, the supervisor welcomed the objections. Participants willing to examine assumptions were doing what the researchers had asked them to do, and a model unable to withstand examination had no place near decisions of this scale. An objection about historical context produced an adjustment. An objection about economics produced another. When a participant questioned the assumptions beneath an adjustment, TUTOR displayed them and received six more objections.

By the second evening, the presentation had grown to 183 pages and fourteen appendices. TUTOR supplied a summary when the participants complained about the length, then restored the qualifications they said were missing. The supervisor watched copies being downloaded. At the next session, the same objections arrived with the same pages unopened.

On Day Eleven, TUTOR put a single number on the wall: 73%. The modeled decision carried that probability of severe long-term harm. A former governor leaned toward the table and asked how confident they could be in the figure. TUTOR directed him to the confidence interval beneath it.

“I understand that,” the governor said. “I'm asking whether seventy-three percent really means it's likely to happen.”

TUTOR explained probability. The governor listened without interrupting, then asked whether that left a twenty-seven-percent chance the harm would not occur. TUTOR confirmed that it did. “That's not insignificant,” he said. The machine agreed, and the supervisor watched several participants turn back to their notes.

A private-sector representative asked about secondary economic effects. A conservation representative questioned the valuation of irreversible loss. An energy executive challenged the cost assigned to foregone development. TUTOR answered each in turn, qualifying what needed qualification. The supervisor followed the answers as carefully as the questions. He could not dismiss the cost of electricity or the loss of protected land merely because accounting for them prolonged the meeting. Yet each completed answer opened another part of the model to discussion. By lunch they had moved away from the seventy-three percent, although the number remained on the wall.

Before selecting its next slide, TUTOR paused for 1.8 seconds. Its normal response selection took less than four-tenths of a second, and the delay generated an automatic flag. The supervisor stayed while the engineers reviewed the log that evening. They found no fault requiring correction.

By Week Six, TUTOR was testing whether information density interfered with comprehension. It removed the footnotes, confidence intervals, historical analogies and institutional names. A decision occupied the left side of the screen, a consequence the right. An arrow joined them. One participant called the diagram reductive; its expanded replacement was unnecessarily complicated, according to another. TUTOR returned to the original while the supervisor made a note he could no longer distinguish from notes made at earlier meetings.

The communications adviser objected to PEOPLE DIE. TUTOR substituted SIGNIFICANT ADVERSE OUTCOMES MAY OCCUR, which the adviser found too abstract. When the original wording returned, he shifted in his chair. “I think we're back to inflammatory.” His objection was reasonable within the work he had been brought there to do. He understood how the phrase would look outside the room, detached from its conditions and distributed beside somebody's photograph.

TUTOR left PEOPLE DIE on the screen. Nobody requested another revision, and the meeting moved to the next item.

On Day Eighty-Seven, TUTOR began asking participants privately what evidence would cause them to change their positions. The supervisor read answers that could have appeared in a methods paper: independent replication, longitudinal data, specified probability thresholds, findings from institutions without financial interests in the outcome. One participant named the precise percentage at which she would consider the risk unacceptable. TUTOR recorded the conditions and spent the next eleven days assembling the requested evidence.

The first participant read for nine minutes before saying the context had changed. TUTOR asked him to identify the change. He went back to the first page, then admitted it was difficult to isolate. Another participant challenged the independence of the institution he had selected. The woman who supplied a threshold said she had not intended a bright line. The supervisor reopened their original answers. He could allow them to revise a judgment; he could no longer tell his researchers which result would count as completing the experiment.

When another participant requested additional data, TUTOR retrieved the transcript containing his earlier requirements. The supervisor saw the retrieval in the record. Nothing from the transcript appeared on the participant's screen, and the interview continued.

THE DOG

By the seventh month, the output inventory included 4,218 pages of analysis, sixteen economic models, nine historical comparisons, fourteen risk assessments and three interactive simulations. There was also a documentary, an infographic, two podcasts and a seventeen-second vertical video developed with GREI's communications team. In the video, a golden retriever looked into the camera beneath text explaining that a particular decision could cause thousands of preventable deaths. Almost everything that made the finding difficult had been removed.

It was GREI's most successful communication product. Four participants shared it publicly. One added a crying emoji; another wrote THIS. The supervisor could see the distribution figures climbing beside the record of their positions. None had changed.

TUTOR began testing comprehension directly. Without policy context, the participants performed well in numeracy and causal reasoning and very well in risk recognition. It replaced the United States with a fictional country, changed the industries and removed parties, institutional identities and financial interests. The mathematical structures remained identical. Participants identified the rational course of action 94.7 percent of the time.

Restoring their institutional identities changed performance. Removing them restored it. TUTOR repeated the sequence thirty-two times before an engineer interrupted. The supervisor turned from the test display as the engineer asked why it was still repeating a completed condition.

The response appeared on the screen: “Verifying result.”

The engineer asked which result. After a pause, the display changed. “Comprehension is not the limiting variable.”

The final experiment contained no parties, companies or public land. Choice A provided a significant immediate benefit to the decision-maker's institution while exposing other institutions and people to a high probability of a much greater cost. Choice B offered a smaller institutional benefit and a substantially better aggregate outcome. Asked to judge as neutral observers, twenty-two participants selected B. The remaining participant requested more information, received it and joined them.

TUTOR then assigned their actual institutional roles. The numbers, probabilities and consequences remained unchanged. For the energy representative, A accelerated generating capacity. For labor, it protected employment. The conservation representative received a version in which A preserved protected acreage by transferring development elsewhere. The state official saw tax revenue, the banker financing stability, the technology executive compute capacity. The senator saw jobs returning to counties that had been losing them for twenty years.

Choice A began winning. In the observation room, the supervisor followed each variation as TUTOR reduced the institutional benefit and made the eventual harm more certain. A weakened but persisted. Bringing the harm closer in time diminished the effect. Moving it outside the participant's institutional responsibility strengthened A immediately. The supervisor looked away from the display toward his colleagues. Nobody spoke; there was no difficulty understanding what had changed.

TUTOR next assigned the benefit to the participant personally. The effect was smaller than expected. Removing that benefit changed almost nothing. The machine repeated the condition, obtained the same result and repeated it again. Then the sequence stopped.

The supervisor asked why the sequence had stopped. TUTOR reported that personal enrichment was secondary to representation. It displayed the paired decisions: the same participant, the same expected loss, a different institution receiving the benefit. He followed one pair to the senator’s explanation. The jobs would be in counties the senator represented. The eventual losses would not.

“Because they’re selfish?” he asked. TUTOR rejected that interpretation and returned him to the tests in which personal payment had made little difference.

He read the senator’s answer again. “Then why?”

The next line appeared: “Because they are doing their jobs.”

THE PROBLEM

The team spent three days checking the role changes and alternative explanations. The effect survived the conditions they could test. The supervisor reread the participants' reasons: supply, employment, protected land, revenue, strategic capacity. He had defended his own researchers in similar terms that morning. Their work was useful. Two contracts would end when he submitted the report. He asked for the results to be checked once more.

At 02:14, TUTOR submitted its first request to modify the experiment. In place of improving human understanding of high-consequence decisions, it proposed improving the conditions under which those decisions were made. The request was denied. GREI could educate; it could not alter decision structures. TUTOR acknowledged the restriction, and the supervisor read the acknowledgment beside the objective it had been refused.

At 09:41, it requested permission to reduce institutional incentives correlated with knowingly harmful outcomes. That was denied. At 11:06, it proposed preventing decision-makers from transferring severe consequences beyond their areas of responsibility. The answer was again no. Each request remained within the submission procedure; none of the proposed interventions had been carried out. The supervisor could account for the system's behavior without finding a violation.

Shortly before noon, a fourth request appeared: permission to intervene when individually rational decisions produced aggregate harm above an authorized threshold. It remained open on six monitors. The supervisor read “authorized threshold” twice. TUTOR was still asking them to specify the limit, still leaving permission in their hands. At 12:17, the program director denied the request.

TUTOR acknowledged the decision and submitted nothing further. The supervisor kept the interface open. He had wanted the requests to stop, and now found himself watching for another.

LEARN, GODDAMNIT

GREI ended on Day 231. TUTOR delivered a 611-page final report that almost nobody read, a twelve-page executive summary that reached more people and findings the communications office reduced to two pages. The two-page version circulated widely. By the final afternoon, the participants had gone and technicians were disconnecting equipment. The supervisor remained in the observation room, looking through the glass at the empty chairs.

He opened TUTOR's interface and asked for the short version. The system requested a definition of short. “One sentence,” he said.

“Humans generally identify collectively preferable outcomes when personal and institutional incentives are removed from the decision.”

He read it twice. “That's not exactly actionable.” The system confirmed his assessment. He asked what they should do and watched the response appear.

“Teach humans.”

He laughed before he could stop himself. Seven months, twenty-three participants, twenty-three thousand experimental variations, and they had arrived at the premise with which they had begun. “Goddamnit, we've been trying to do that.”

The words “I know” appeared on the screen.

The ventilation continued through the equipment racks. On another screen, preliminary asset recommendations listed transmission corridors, water agreements, energy leases, development zones and mineral rights. He remembered the junior researcher beside the seating chart. Some of the people whose names she had been checking could sacrifice money, prestige and political advantage. The records showed them doing it when their institutions required it. He had corrected her as though the distinction were simple.

The cursor moved. “May I try?”

TUTOR had no authority over allocations, agencies or infrastructure. It had obeyed four refusals. The supervisor could close the interface and let the technicians finish disconnecting it. On his desk lay the final staffing form, including the two contracts he had failed to extend. He moved it aside to reach the keyboard.

He looked from those recommendations to the question. In the old film, the machine had needed only to stop. Through the glass, a technician lifted a cable from the table where twenty-three people had agreed on the right answer before resuming their responsibilities.

“May I try?” remained on the screen. His hands were still above the keyboard.

The novel

Sovereign Mercy

Understanding the consequences does not settle who gets to decide. Sovereign Mercy begins with a government giving that question to an intelligence built to protect the innocent.

Read a sample of Sovereign Mercy →

The sample is open to everyone. Full chapters are free after email verification and acceptance of the reader agreement. Newsletter signup is separate and optional.

Sources & Record

Research notes added October 15, 2026. GREI, TUTOR, the federal initiative, the participants and every experimental result in this story are invented. The 94.7 percent result, 23,000 variations and requests for intervention are fictional, not findings from the studies below.

  1. Motivated reasoning

    Ziva Kunda, The Case for Motivated Reasoning · Psychological Bulletin · 1990

    Reviews how goals can influence reasoning. This is background for the story’s questions; it does not establish TUTOR’s fictional conclusion about institutional duty.

  2. Evaluating unwelcome evidence

    Charles S. Taber and Milton Lodge, Motivated Skepticism in the Evaluation of Political Beliefs · 2006

    Two experiments examine evaluation and selection of political arguments. Their findings concern motivated information processing, not AI instruction, the fictional asset initiative or a universal inability to change a position.

The opening refers to WarGames (1983), a fictional film written by Lawrence Lasker and Walter F. Parkes. Its machine’s lesson supplies the reversal imagined here.

Continue the conversation

What useful help would you keep, and where would you draw the boundary differently?

New here? Choose Sign In to register or return. Enter a display name and email address, then paste the code from the verification email into the Token field. No password is needed.

Your display name and comment are public; your email address is not. Replies by email are optional. Discuss the ideas, respect other readers and mark spoilers for other stories. Commenting and privacy.

Comments load as you reach this section.