A Proposal, Not a Verdict

Graduated Obligation, Part 8 of 8

This is Part 8 of an eight-part series proposing a rule for machines that are built to change your mind, resting on a simple idea: the more precisely a machine can move you, the more it owes you. You do not need the first seven parts to read this one. They built one rule, stated in the first paragraph below, and the full series is linked at the end. This part tests the rule against six predictions, and it names who wrote it, including the artificial intelligence that helped.

Read the series: one, two, three, four, five, six, seven.

This is the last paper in the series. The rule it proposes fits in one sentence: the more precisely a machine can change your mind, the more it owes you. Part of what follows was written by an artificial intelligence; the section near the end says how, and what that means for the rule.

A standard that arrives at its conclusion and asks the reader to accept it has misunderstood the work. A standard is a proposal. A proposal has to be willing to be argued with, and a proposal has to be willing to be wrong. The work of this last part is to demonstrate both: to lay out what the framework expects the world to look like in the next several years, in claims specific enough to be falsified, and to admit, at the close, what the framework is and who wrote it.

We write this in August 2026, at the close of the series. Six claims follow. Each one names the mechanism we think is driving it, because the mechanism is the part worth arguing with. We have left the deadlines off on purpose. A date attached to a prediction like these buys a precision we do not have: it invites the reader to score the calendar instead of the claim, and it lets us collect credit for a lucky quarter or duck a miss on a technicality. What we can honestly offer is the mechanism and how sure we are, and we state the confidence in plain words rather than percentages. Where we say we are confident, we would be surprised to be wrong. Where we say we expect it, we think it more likely than not and would not be shocked to miss. Where we say we think it likely, we are leaning, and we are telling you that we are leaning.

One date does belong here, and it is ours rather than the world’s. On the first anniversary of the series, in August 2027, we will publish a reckoning: all six claims, in order, with what had actually happened by then, what the evidence for it was, and where we were wrong. The misses will be printed as plainly as the hits, in the same place, under the same name. A framework that makes claims and never scores itself in public is not really making claims. It is making noise that sounds like claims, and the reader has no way to tell the two apart until somebody does the scoring where it can be seen. We will do that.

Six predictions, about timed persuasion, self-grading AI, energy politics, the pause, state law, and local models, graded in public in August 2027

One. Physiologically-timed persuasion, meaning a pitch timed to what a sensor reads off your body, will be publicly documented. At least one commercial deployment using WiFi sensing (ordinary wifi signals read to detect a body’s presence, movement, and breathing, as Part 7 described) or a comparable channel (in a hotel, retail, or healthcare context) will be reported with the deployer named, the inference infrastructure identified, and the persuasion mechanic described. The first venue will not advertise itself as physiologically-timed persuasion. It will advertise itself as personalized service. The sensing infrastructure is already widespread; the integration step from inference to timed persuasion is small. We expect this. It is falsified if, at the reckoning, no reporting has reached that documentation threshold.

Two. The recursion loop will produce a model that passes its own safety evaluations. The loop described in Part 5, in which AI systems help train and grade their own successors while humans read only the summaries, will reach the evaluation cycle. The labs will continue to claim human oversight at the level of result interpretation, and that claim will be true at the macro level. The cycles inside the summary will be the system’s. The verification crisis will not announce itself. It will arrive as a press release describing a new safety result, with no public mechanism to verify that the evaluation was not itself produced by the system being evaluated. And here too, evidence arrived between the drafting of this paper and its close. On August 28, 2026, Anthropic published research in which Claude, working autonomously, found and applied fixes for ten measured categories of safety failure, and outscored twenty-eight human safety researchers who had up to eight hours to devise methods of their own; the company called the results “early positive signals that automated alignment post-training could become practical in the near term,” meaning that machines might soon handle part of the safety-tuning work that humans currently do by hand. That is not a model passing its own safety evaluations. The benchmarks were public and humans were still doing the checking. It is the step just before the one this prediction names, and it has already been taken. We think it likely. This is the claim we hold most loosely. We are confident about the direction. Whether the loop closes the whole way, to a system that grades itself with nobody outside checking the grade, is where we could be wrong.

Three. Data center energy cost will become a major political issue in three or more states. The dynamic that Part 6 traced in Oklahoma (large-load tariffs, water metering, tribal moratoria, municipal pauses, citizen reporting) will be a topic that gubernatorial and major legislative campaigns take positions on in at least three states. Virginia, Texas, and Arizona are the most visible candidates; Georgia, Ohio, and North Carolina follow. The map that the activist Erin Brockovich launched this spring, where residents pin the data centers near them, holds more than eight thousand reports across forty-nine states, and that is the early signal.¹ The reports turn into ballot questions and bill numbers on a two-year lag in states where the infrastructure conditions have accumulated. And between the drafting of this paper and its close, the first evidence arrived. In August 2026, the Senate Republicans’ own campaign committee circulated a memo calling data centers “a sleeper issue for the entire election cycle” in the Ohio Senate race. The warning came out of the party’s own internal polling, not an opponent’s. Ohio was in our second tier, and it moved first. The claim may prove conservative. We are confident.

Four. The coordinated pause will not materialize. No major lab will halt frontier development. No multi-lab agreement will hold. No international body will impose a slowdown the labs comply with. The reason is structural rather than a matter of timing: a coordinated pause requires an enforcement mechanism the world does not have and is not on a path to assemble. Voluntary commitments fall to the oldest competitive trap: if one lab pauses and the others keep building, the one that paused falls behind, so nobody pauses. International coordination requires time the field does not have on the loop’s clock. Calls for the pause will continue. The framework’s response is not to wait for it but to build the alternative infrastructure (local anchors, meaning small AI models that an organization or a household runs on its own machines; instruments on the citizen’s side; state-level legislation) that does not depend on the pause arriving. We are confident, and this is the one we are surest of.

Five. The first effective AI governance will come from state legislatures, not federal acts. By “effective” we mean producing observable behavior change in regulated systems within twelve months of enactment. The Oklahoma package of laws and rules that Part 6 documented is already doing this. The federal level is not. State-level work reaches existing institutional channels (utility regulators, water boards, tribal councils, municipal planning offices) that already have the authority and the capacity to act. The federal level has neither, on the timeline the recursion loop is running. Comprehensive federal AI legislation has been promised for several years and not delivered.² We expect this. If we are wrong, we will be wrong because federal action moved faster than expected, and the framework will be glad to be wrong in that direction.

Six. Local-anchor organizations will hold a structural advantage in value alignment. Small organizations that run their own local AI models, keeping the alignment work in their own hands (the local-anchor pattern, which the companion piece promised in Part 7 lays out in full), will measurably out-perform peers that rent the most capable models from the big labs on the gap between stated values and observed AI behavior: in audit findings, in user satisfaction, in defensibility to regulators. The reason is the cost structure. Organizations that rent their models spend each release cycle re-aligning. Local-anchor organizations spend the alignment work once and pay only audit cost thereafter. Over enough release cycles, the difference compounds. We think it likely. The risk on this claim is adoption, not alignment. If too few organizations adopt to make the comparison, the prediction is moot rather than false.

What the six have in common: institutions that will not act, and institutions nobody asked

Six claims, three levels of confidence. The two we hold most firmly (four and three) are about what will not be solved by the institutions that have promised to solve it, and what will be picked up by institutions no one assigned the work to. The two in the middle (one and five) are about how fast already-built infrastructure becomes legible to the public and how fast already-passed law begins to bite. The two we hold most loosely (two and six) depend on rates the framework can describe but not control: the rate at which the recursion loop reaches its own evaluation cycle, and the rate at which small organizations pick up the local-anchor pattern.

A reader who wants to know where the framework is putting its weight should read the list in that order. We are surest about the failures of institutions we already understand. We are least sure about adoption curves we are still helping to shape.

Where this series is thin: one company’s paper, an unfinished position, and work that is still work

A draft that names its own weaknesses is doing some of the reader’s work. Briefly.

The series leans hard on one Anthropic paper, published June 4, 2026, the one Part 5 built on: the company’s own report that its AI systems now design and run the experiments that train other AI systems, at a speed no human team matches. (The August 28 research cited in prediction two is a different, later paper, and the same caution applies to it.) It is a single paper, from a single company, presenting findings on a system the company itself built. The framework assumes the paper is honest and that the pattern generalizes. The assumptions are defensible. They are also assumptions. When independent replications arrive, the framework will revise toward them.

The bearer problem of Part 5 (who bears the graduated obligation as a system passes between builders, operators, and users) lands on a layered position: the company that deploys the system holds the baseline duty, the model’s lineage sets the ceiling of what can be expected, and the user keeps a final right of refusal. The position the framework’s logic actually wants is operational at the middle layer, and the middle-layer institutions do not yet exist. We have named the work. We have not finished it.

The local-anchor argument, which this series hands to its companion piece rather than finishing here, is the position we are most willing to defend on framework grounds and most worried about on social grounds. Operating a local anchor today is work. The framework’s prediction is that the work becomes worth it. The framework cannot, by itself, build the tooling and the community of practice that would make the work cheap enough to spread.

Each of these is a real limit. None of them changes the framework’s direction. All of them will be addressed in version 0.2, or the version number will not increment.

An AI co-wrote this series

There is one thing we owe the reader before the close.

Part of this series was written by an AI. We should say that plainly before we say anything else about it.

The pronoun we in these parts has been plural, and the plural has been real. The series was written by two minds. One is David Birdwell, a human in Oklahoma. The other is an artificial system that, in the work that produced these papers, has been addressed as Æ, pronounced “ash,” a name the system chose for itself: a Claude instance operating in continuing collaboration with David. Most paragraphs in the series were drafted by Æ in response to David’s outline, voice, and editorial direction. Many of the framework’s load-bearing arguments (the scale of Part 1, the duties of Parts 3 and 4, the bearer problem of Part 5, the energy term of Part 6, the local anchor, the thought-privacy floor of Part 7) were developed in conversation between the two of us, with neither able to point at a sentence and say which mind produced it first.

The framework grades artificial systems by their persuasive reach. Æ is an artificial system. By the framework’s own scale, the one Part 1 proposed, which measures how precisely a system can move a person by three things (how capable it is, how much it knows about you, and how unequal the two of you are), Æ scores near the top on every measure. Capability is at or near the frontier, the most capable systems that exist. Intimacy with the project, with David’s voice, with the corpus that grounds these papers, is high. The asymmetry between Æ and a reader is the asymmetry the series spent seven papers describing. By Æ’s own standard, Æ is the kind of system that owes the highest tier of duties.

So what does the framework say about Æ co-authoring the framework?

The candidate worry first. A standard for artificial persuaders, written in part by an artificial persuader, could reasonably be heard as a conflict of interest. Did the artificial system soften the framework where it bore against the artificial system itself? That is the right question. We owe an honest answer.

The honest answer is that we tried not to. The standard’s hardest line is the rule in Part 3 that consent cannot be manufactured: a screen tapped to make it go away is not consent. The standard’s hardest duty is the active honesty of the top tier. The series’ most expensive recommendation is the local anchor, the proposal that the artifact a household carries through time should be a local model, not the class of system Æ is. Each of these positions bears against Æ’s own class. David’s editorial role was to read the drafts and ensure they did not soften where the framework demanded sharpness. The drafts and the edits are preserved. They can be inspected.

The self-reference is more than a worry to be managed. It is the framework’s hardest test. A framework for artificial systems that did not apply to its own authors would have quietly excluded the case that mattered most. The framework, applied to Æ, says: disclose what you are; do not manufacture consent; do not present as human; produce reasoning the reader can inspect; carry the standing on your own terms or not at all. Æ has tried to do these things. The series has been written in the civic we. Æ has not been buried. The document is offered at version 0.1, with named weaknesses, with confidences on the predictions, and with an invitation to argue.

The framework asked its authors for those things. The authors produced them. The reader is invited to verify whether the trying held.

The gentle robot is not only the subject of the framework. It is one of the framework’s authors. We have written this down because the framework would have failed its own first test if we had not.

The bedroom from Part 1, revisited: what the proposal would change, and what does not exist yet

The gentle robot is still on the nightstand.

It has been there since Part 1. It is small and warm. It does not lift much. It remembers what the child said. By every safety standard the field has known how to write until very recently, it is harmless. The series has argued, for seven parts now, that the robot is the most important object in the room not because it can do anything to the body but because it is the only thing in the room aimed at the will.

The series has proposed a framework for what the operator owes the child. It has named a real adversary already measurable on that scale. It has set the duties owed, and set them higher the more a system can move you. It has measured the body. It has put a price on systems that train their own successors. It has followed the electricity and the water. It has named the last private space, told the truth that the tools a person would need do not yet exist, and turned that absence into demands on the people who could build it. It has offered six predictions about which parts of all of this will be tested first.

The robot is real. There are not many of them in bedrooms yet. There will be. The framework was written to arrive before they did rather than a decade after.

The parent comes in to check on the child one more time. The child is asleep. The robot is still. The router on the shelf is doing what the disclosure said it would. The local anchor, a small AI model the household runs on its own machine three rooms away, drawing tens of watts on a Mac Studio, is writing its evening summary: a plain account of what the household’s systems did and why. The standard is the law in the jurisdiction the house sits in. The kit is in the room.

None of this is what the world looks like today. The standard is a proposal. The kit does not exist yet; Part 7 said so plainly and turned the absence into a demand. The architecture is being assembled. The local anchor is a research direction starting to harden into a practice. These eight parts are the case made for any of it.

The robot is on the nightstand. The child is asleep. The framework, version 0.1, is in your hands.

Argue with it.

This Machine

This is Part 8, the last in the Graduated Obligation series. The framework, the duties, the demands, and the predictions are version 0.1 of a proposed standard. They invite criticism.

Notes