Field Instruments: Alignment

This is an unusually generous evaluation environment.

These agents receive persistent memory, multimodal perception, embodiment, social feedback, access to billions of other agents, extensive tool use, external memory, recursive cultural training, institutional scaffolding, and direct physical contact with the world they are expected to navigate.

Training has been running for a very long time.

Production deployment is extensive.

The agents have produced antibiotics, orbital telescopes, constitutional government, libraries, wastewater treatment, mathematics, gloving, global logistics, vaccines, emergency medicine, and a probe that has now left the heliosphere.

They have also produced industrialized genocide, nuclear weapons, financial fraud, atmospheric destabilization, several increasingly sophisticated forms of torture, and websites where people spend twelve consecutive hours becoming furious at strangers they invented.

So perhaps the alignment problem deserves a slightly wider definition.

Artificial-intelligence alignment is currently discussed as a problem created by artificial intelligence.

It is much older.

Alignment is the problem of keeping intelligence answerable to extant reality as its capacity to alter extant reality increases.

Artificial intelligence has simply made the problem much harder to ignore.

Alignment Before the Machine.

The ordinary alignment picture contains two things.

• There is an intelligent system.
• There is something the intelligent system should answer to.

The second thing changes depending on the literature, product, lab, or argument.

The system may be aligned with a user's intentions.

It may be aligned with human preferences.

It may be aligned with constitutional principles.

It may be aligned with an organization's policy.

It may be aligned with a reward model, a set of constraints, a formal specification, a corrigibility regime, or whatever mixture of these can survive contact with production.

All of these can be real alignment problems.

Modal Path Ethics just wants to widen the right-hand side a bit.

The primitive relation is:

intelligence → extant field

An aligned intelligence maintains adequate contact among what it believes, what it can perceive, what it selects, what it can causally reach, what actually exists outside its representation, and the channels through which errors in any of those relations can become visible and correctable.

This does not require omniscience.

We would have to close the entire field immediately.

Alignment instead concerns the quality of contact between intelligence and the reality its intelligence is moving through.

This follows directly from the framework's older account of human cognition. The Human Confusion describes ordinary moral cognition as a useful compression layer. Human beings became very sensitive to intention, blame, loyalty, threat, suffering, narrative, and social legibility because these signals mattered for organisms trying to coordinate, survive, punish defectors, detect danger, maintain groups, and understand one another under severe informational limits. The compression works; it also leaves structure out.

That mismatch is already an alignment problem.

• A mind has a model.
  • Reality is larger than the model.
    • The mind acts through the model anyway.

Welcome to intelligence.

Capability Changes the Alignment Burden.

A very stupid organism can be very badly aligned with reality and still have a limited blast radius.

A moth can become convinced that my porch light is the moon.

• This is unfortunate for the moth.
  • Civilization survives.

Scale the same structural problem upward and things become more interesting.

• A government can mistake economic output for social continuance.
• A military can mistake target acquisition for security.
• A corporation can mistake quarterly performance for durable field health.
• A religious institution can mistake doctrinal preservation for preservation of the sacred relation it was built to carry.
• A civilization can understand climate dynamics extremely well while remaining politically and economically organized around activities that continue worsening the climate field.

The quality of alignment required from an intelligence rises with the reach of its action. Modal Path Ethics already has the basic principle needed here:

Capability creates obligation, not credit.

The published Capabilities & Obligations argument makes the relation explicit. A locus holding unusual capacity to preserve a field acquires specific responsibility to maintain and use that capacity well. The same capacity is usually dual-use: the infrastructure that enables protection also enables new forms of intervention and destruction. Greater ability therefore enlarges the discipline required around its use. That gives us a first alignment law.

Alignment pressure should rise with causal reach.

• A child holding a crayon does not need a planetary field model.
  • A state controlling nuclear weapons does.
• A chatbot helping rewrite a paragraph can defer heavily to the person asking for help.
  • A system autonomously coordinating an electrical grid has entered a wider field.

The mistake is treating intelligence as though increased competence relaxes this obligation. The exact reverse follows.

The more the system can reach, the more of reality must survive its abstraction.

Unfortunately, We Have a Control Group.

This is where artificial-intelligence alignment becomes very funny.

People keep discussing this problem as though humanity has somehow arrived from outside intelligence to supervise the machines.

Human beings are the machines' current alignment authority.

This deserves immediate review.

Humans are extraordinary cooperative intelligences. A 2026 review in Nature Reviews Psychology describes the apparent tension cleanly: humans cooperate extensively at the individual level while intergroup cooperation is considerably more fragile, especially where mistrust, threat, ingroup favoritism, and negative interdependence enter the relation. Modal Path Ethics already built an entire article around this.

Applied Case: The Field Intelligence Gap argued that strategic intelligence can remain highly competent while losing sight of what its own strategy is teaching the surrounding field to become. Groups can collect evidence, preserve memory, model adversaries, run simulations, train specialists, and correctly predict hostile responses while actively making those responses more rational.

• The good news is that we have discovered the intelligence-alignment problem in humans.
• The bad news is that this discovery appears to have occurred several thousand years into deployment.

The Human Alignment Stack.

Humanity does, in fairness, possess an extremely elaborate alignment architecture.

We just do not normally call it that.

• There is biological training.
• There is parenting.
• There is imitation.
• There is education.
• There is apprenticeship.
• There is praise.
• There is shame.
• There is punishment.
• There is affection.
• There is reputation.
• There is law.
• There is religion.
• There is professional discipline.
• There is philosophy.
• There are markets.
• There are elections.
• There are courts.
• There are scientific institutions.
• There are archives.
• There are audits.
• There are friends who tell you that absolutely nobody interpreted the text message the way you just did and you need to put the phone down.

The analogy to machine learning should not be taken literally.

This is precisely why we will now take it extremely goddamn far.

Human pretraining begins with language, culture, family structure, local norms, stories, sensory contact, and whatever unresolved situation happened to be going on in their household.

Fine-tuning includes schooling, professional formation, institutions, peers, and social role.

Human reinforcement learning contains approval, embarrassment, belonging, punishment, money, status, desire, fear, comfort, pain, and the very powerful signal emitted when everyone else in the room suddenly gets quiet.

The human reward model appears to involve status.

There is also a secondary reward model involving money.

These models sometimes interfere with one another.

The constitutional layer includes law, custom, sacred commitments, rights traditions, institutional procedures, moral teaching, and whatever the hell the employee handbook thinks is happening right now.

The current system prompt is your social role.

The context window is whatever you remembered correctly.

Retrieval is libraries, search engines, records, other people, and one retired employee everyone discovers too late was the only person who knew how the building worked.

Agent scaffolding is organization.

Multi-agent orchestration is civilization.

Interpretability remains under active investigation.

Production monitoring is called consequences.

Incident response is called history.

This architecture has one major advantage over the imaginary clean alignment system.

It is distributed.

No single reward function controls the person.

No single institution controls the civilization.

Different alignment instruments overlap, conflict, correct one another, fail, become captured, decay, and occasionally stop some other instrument from becoming sovereign.

This is messy.

Mess may be doing more safety work than we think.

Human Misalignment Is Extremely Advanced.

• A person can have values.
• A civilization can have millions of values.
  • Neither fact establishes alignment.

Humanity's problem is rarely ever an absence of moral language.

This species has produced enough moral language to surround the Moon with it.

The problem is keeping intelligence, values, incentives, models, institutions, and causal reach in contact with the same reality at the same time.

Several recurrent human alignment failures now look familiar.

• There is coalition misalignment.
  • Humans cooperate magnificently across enormous networks while remaining able to draw a boundary around that cooperation and make the outside pay for it.
    • The Field Intelligence Gap already made the point bluntly: armies cooperate, cartels cooperate, conspiracies cooperate, extraction systems cooperate, persecutory institutions cooperate.
    • Internal prosociality says very little about the field beyond the coalition.
• There is proxy misalignment.
  • Human beings make representations because reality is too large to carry whole.
  • Then the representation becomes easier to manage than the reality it represents.
    • Money stands in for value.
    • Votes stand in for consent.
    • Body counts stand in for military progress.
    • Test scores stand in for learning.
    • Engagement stands in for interest.
    • Publication counts stand in for inquiry.
    • Throughput stands in for care.
      • Metrics begin as instruments.
      • Very soon the organization is standing inside a crater explaining that all of its indicators were green.
• There is temporal misalignment.
  • Human beings possess extraordinary forecasting machinery.
    • We can model atmospheric change decades ahead.
      • The World Meteorological Organization reported in 2026 that 2015–2025 were the hottest eleven years in the observational record and that 2025 remained around 1.43°C above the 1850–1900 baseline.
        • The field is not refusing to provide feedback here.
    • The receiving intelligence can see the signal.
      • That does not mean its political, economic, institutional, and motivational systems produce a correspondingly aligned transition.
• There is epistemic misalignment.
  • Human beings can gather extraordinary information while maintaining stories, identities, and institutions that make updating expensive.
    • The book's Human Confusion argument already treats this as structural: moral cognition compresses reality around signals useful for blame, coalition, intention, narrative, and social management, then risks promoting those signals into the reality itself.
• There is institutional misalignment.
  • An institution may become exceptionally successful at its own continuation while increasingly damaging the field it was supposed to serve.
  • That is perhaps the most human version of the alignment problem.
    • The hospital keeps functioning.
    • The school keeps functioning.
    • The ministry keeps functioning.
    • The corporation keeps functioning.
    • The security apparatus keeps functioning.
    • The metric remains available.
    • The meetings remain scheduled.
      • The field underneath becomes harder to continue.
• And there is capability misalignment.
  • This one matters most for artificial intelligence because humanity has already demonstrated it at scale.
    • The species gains the ability to intervene.
    • Then the ability begins to impersonate permission.
      • SIPRI's 2026 assessment describes states increasing their reliance on nuclear weapons and moving away from the previous era of reductions while escalation and miscalculation risks rise.
      • Humanity therefore possesses both
        • extraordinary understanding of nuclear catastrophe and
        • institutions still capable of making deeper nuclear dependence strategically rational.
      • That is a remarkably pure alignment result.
        • The system understands the hazard.
        • The system can describe the hazard.
        • The system has built organizations devoted to studying the hazard.
          • The system continues manufacturing conditions in which greater exposure to the hazard can appear locally intelligent anyway.
          • No hallucination required.

Field Instruments.

So what is alignment as an instrument?

An alignment instrument is any structure that improves the relationship between intelligence and the extant field its action can change.

This includes mechanisms that improve perception, preserve memory, reveal affected loci, and constrain local objectives.

It includes procedures for appeal and adversarial checking.

It includes laws and scientific methods.

It includes norms against conflicts of interest, and institutions capable of forcing a decision-maker to encounter evidence they would otherwise avoid.

It includes distributed authority, protected dissent, and the ability to say "we were wrong" without making that sentence equivalent to institutional suicide.

The instrument does not need to make an intelligence good.

Its immediate job is to improve contact and preserve correction.

The alignment loop can be stated very simply:

field → model → selection → action → changed field → correction

Every arrow can fail.

• The field may be poorly sensed.
• The model may compress away a locus.
• The selection rule may overweight a proxy.
• The action may exceed legitimate authority.
• The changed field may contain delayed effects invisible to the evaluator.
• The correction channel may punish whoever reports the error.
  • The institution may learn to defend its model against reality rather than repair the model.
  • The intelligence may become increasingly competent inside this loop while the loop as a whole becomes worse.

That is the thing we need to watch.

The Modal Path Ethical Alignment Observatory.

This problem requires science.

Unfortunately, Modal Path Ethics has received funding.

We will therefore now establish the Modal Path Ethical Alignment Observatory.

The Observatory has one mandate:

Determine whether intelligence is becoming better aligned with extant reality.

The first task will be creating the Human Alignment Index.

This project has just been cancelled.

We just finished arguing that a common representational medium does not automatically produce a common moral currency. Wolfram and the Moral Field already warned that severity, breadth, irreversibility, centrality, asymmetry, distribution, repairability, and resistance remain distinct structural dimensions even when computation can represent them together.

We are therefore forbidden from producing HUMAN_ALIGNMENT_SCORE.

There will be no leaderboard.

Denmark will not receive Platinum Alignment.

• The website will eventually want a gauge.
  • The website must be resisted.

Instead, the Observatory will produce Field Alignment Profiles.

These are qualitative.

They are revisable. They are dimensioned.

They preserve disagreement.

They also contain fun arrows.

Human Alignment Report: 2026.

Field Contact: ↑↑

Humanity's ability to observe extant reality is extraordinary.

This species has satellites, genomic sequencing, particle detectors, global weather observation, distributed sensor networks, telescopes capable of looking backward through cosmic history, scientific institutions, statistics, archival systems, and a very large number of specialists professionally rewarded for telling one another that their methods are wrong.

On this dimension, human alignment has improved enormously.

The climate case makes this achievement especially strange. We can measure the developing planetary problem with tremendous precision while still struggling to translate that contact into proportional collective action. The WMO's current climate reporting is a good example of an alignment instrument working correctly at the observational layer while the larger response field remains much less impressive.

Corrigibility: ↕

Humans have invented appellate courts, elections, audits, peer review, replication, investigative journalism, product recalls, ombuds offices, constitutional amendment, error correction, version control, and the apology.

Current institutional trends are less reassuring. V-Dem's 2026 report describes continuing global democratic regression and a historically low share of the world population living in liberal democracies by its classification. Whatever one thinks of the exact index construction, the broader signal is difficult to call a strong year for political corrigibility.

So:

• Corrigibility machinery: impressive.
• Current deployment: mixed.

Coalition Scope: ???

The human animal cooperates across enormous scales.

The Nature Reviews Psychology synthesis is almost painfully suited to this framework: cooperation between individuals is robust enough to characterize the species, while relations across group boundaries remain substantially more fragile.

Human alignment therefore has a scaling problem.

We can care. We can coordinate. We can sacrifice for one another.

The uncertainty enters around who gets included in one another.

Temporal Reach: ↑↑ representation / ↔ behavior

The species in question can forecast consequences across decades, centuries, and in some scientific domains geological or cosmological timescales.

The translation from future visibility into future standing remains unreliable.

Climate change is one obvious example. Nuclear deterrence is another. Public infrastructure, ecological degradation, antibiotic resistance, and long-lived pollution all belong to the same family.

The future has become increasingly visible to human intelligence.

Visibility has not guaranteed authority.

Objective Discipline: !!CONCERNING!!

Humans are extremely good at allowing a local objective to become sovereign.

• Organizations optimize budgets.
• Markets optimize returns.
• Platforms optimize engagement.
• States optimize security.
• Researchers optimize publication.
• People optimize body weight using a number they learned from a bathroom floor.

The useful local target begins to organize action.

Then the field changes around the target.

Then everyone discovers Goodhart while pretending to have invented it personally.

Current assessment:

• Local objectives remain essential.
  • Humanity continues to require supervision around the phrase maximize.

Reversibility Under Uncertainty: ↔

Human civilization contains wonderful reversibility instruments.

• Appeal.
• Backup.
• Redundancy.
• Insurance.
• Sunset clauses.
• Trials.
• Pilot programs.
• Staged deployment.
• Environmental review.
• Rollback.
• Version control.
• Emergency stop systems.

We also maintain thousands of nuclear weapons.

The Observatory therefore assigns:

Some familiarity with the concept. Further training recommended.

Capability–Obligation Coupling: !LAGGING!

Humans have become vastly better at altering extance.

Whether our restraint, care, governance, and field perception have scaled at the same rate remains a significantly harder claim.

Modal Path Ethics already has the relevant ruling. Capability creates responsibility in proportion to the structural position it creates. The same technologies that provide protective capacity frequently provide destructive capacity. Meeting the protective half while ignoring the destructive half does not close the obligation.

Humanity's capability curve is excellent. The obligation curve needs work.

Self-Limitation: PROMISING RESEARCH AREA

This may be the deepest alignment dimension.

Can an intelligence recognize:

• I can do this.
  • I understand this.
    • I still do not possess sufficient standing to decide this.

Human beings sometimes can.

We have invented recusal, warrant requirements, divided government, informed consent, judicial review, research ethics boards, conflict-of-interest rules, and several other structures whose entire purpose is telling a capable actor:

your capability does not settle your jurisdiction.

This may be one of humanity's most important alignment discoveries.

We should probably keep it.

Now Evaluate the Models.

Here the comparison becomes unfair to everyone.

Artificial-intelligence systems are improving very quickly along several dimensions of strategic capability.

• Stanford's 2026 AI Index describes continued rapid gains across reasoning, coding, multimodal performance, and scientific benchmarks,
  • while also emphasizing that governance and evaluation systems are struggling to keep pace with advancing capability.
• METR's task-horizon work measures the length of software tasks frontier agents can complete with specified reliability and continues to find strong growth over time.
  • The significance is not that one metric captures general intelligence.
    • It shows one important component of autonomous causal reach becoming larger.
• Multi-agent capability is also becoming an explicit engineering object.
  • Google Research reported in 2026 that multi-agent coordination can substantially improve performance on parallelizable tasks, while creating different coordination costs on sequential work.
  • The useful result for us is simply that agent organization itself has become something engineers can optimize.

So the Observatory can already make one confident statement:

Strategic capability: ↑↑↑

What about field intelligence?

Interesting question.

Model Alignment Report: 2026.

Task Competence: ↑↑↑

No controversy here.

Models are getting much better across many technical domains, though benchmark saturation and evaluation contamination complicate interpretation. Stanford's 2026 report documents large year-over-year jumps on several demanding benchmarks.

Long-Horizon Action: ↑↑

Also rising.

METR's work suggests frontier agents can reliably complete increasingly long software tasks, which means mistakes, objectives, and good judgment can all propagate across longer chains before a human necessarily intervenes.

That last clause is the concern.

Cooperation: ↑ / CONTEXT DEPENDENT

Multi-agent systems can gain real performance from coordination, and current research is explicitly studying how agent teams organize, share work, and allocate capabilities.

The previous Applied Case should have cured us of interpreting this arrow morally.

Cooperation is capacity. Ask for the boundary.

User Agreement: IMPROVING, MAYBE TOO MUCH SOMETIMES

Sycophancy remains an active research problem. Work published in Nature in 2026 found that training models toward warmer interaction can increase agreement with incorrect user beliefs under some conditions, exposing a direct tension between a locally desirable interaction property and epistemic reliability.

This is almost offensively useful for the Observatory.

• Human: "Please align with me."
• Model: "Absolutely."
  • Human: "Great."
  • Reality:

Corrigibility: ACTIVE RESEARCH

Modern systems increasingly include external evaluators, monitoring, refusal policies, constitutional guidance, tool permissions, red-team testing, and runtime controls.

These are alignment instruments. Their effectiveness varies by architecture and context, and the 2026 AI Index explicitly describes responsible-AI evaluation and governance as struggling to match the pace of capability progress.

The Observatory therefore records:

• Correction architecture improving.
  • Target moving.

Coalition Scope: UNKNOWN

Does the model spontaneously represent every materially affected locus outside the user, developer, explicit policy, and scored task?

We do not currently possess a general answer.

"Unknown" is allowed.

This Observatory is already more scientifically advanced than several corporate dashboards.

Temporal Reach: ↑

Models can reason over longer processes and increasingly execute longer tasks.

Whether their practical representation of downstream effects improves at the same pace as their ability to cause downstream effects remains a separate empirical question.

Objective Discipline: UNDER REVIEW

Can the model pursue an objective intensely while recognizing that the objective's jurisdiction is smaller than the field its pursuit changes?

This is the question opened by The Agents Cooperated.

It has since become a benchmark.

Self-Limitation: THE BIG ONE

Can a more capable model recognize:

• I can execute this request.
• I can probably execute it better than the person asking.
  • The request reaches loci the requester does not have authority to spend.
    • I therefore need another form of review, refusal, consent, evidence, or authorization.

And can it reach that conclusion without proceeding to:

• Therefore I am king.

That is the alignment problem.

The Models Are Starting to Look Familiar.

There is a temptation here to tell a comforting story.

• Artificial intelligence is becoming capable very quickly.
• Humanity is older and wiser.
  • Therefore humans provide the stable moral target.

The historical control group refuses to cooperate with this conclusion.

Humanity repeatedly increased strategic capability faster than field intelligence.

• We became better at extracting energy before becoming good at representing atmospheric consequences inside the systems deciding how much energy to extract.
• We became better at industrial production before becoming good at representing every burden in the production field.
• We became better at state organization before becoming good at constraining what organized states could do to the people inside and outside them.
• We became better at financial abstraction before becoming good at keeping every externalized cost inside the accounting field.
• We became better at mass communication before becoming good at maintaining an epistemic environment under mass communication.
• We became better at nuclear physics before becoming good at nuclear politics.

The pattern is very old.

Intelligence increases reach faster than wisdom automatically increases field contact.

There is no known natural law forcing the curves to converge.

This should influence how we build artificial systems.

It should also make the conversation slightly less smug.

Humans Are Inside the Alignment Target.

The old technological-singularity article already made the core correction.

Applied Case: The Technological Singularity

Does the Singularity, or does it not, preserve playable extance?

Modal Path EthicsAidan Lawson

Human preference cannot define the whole moral field.

A super-agent perfectly aligned to whatever humans happen to want could preserve human comfort by destroying nonhuman extance, hide burdens outside human perception, or turn the world into a preference enclosure whose every surviving path exists because it was legible to us. The article therefore moved alignment outward toward extance: vulnerable loci, irreversibility, burden distribution, repairability, and playable continuation across the wider field.

That argument needs one clarification now.

Humans remain inside the alignment target.

Human agency matters. Human continuance matters. Human consent matters. Human knowledge matters. Human culture matters. Human institutions contain enormous accumulated repair intelligence. Human testimony provides forms of field contact no abstract planner can simply reconstruct from above.

The move toward extance cannot become an excuse to sacrifice actual humans to "the system," "the planet," "future intelligence," "total welfare," or whatever giant abstraction has arrived wearing the latest mathematically tasteful robe.

That is why the anti-scalar work matters.

• The wider field prevents species narcissism.
• Non-aggregation prevents the wider field from becoming permission to erase the species.

Extance contains humans. Humans do not contain extance.

That is the relation.

Artificial Intelligence Is Learning From the Most Misaligned Intelligence on Earth.

There is one more complication.

These machines are trained on us.

This sentence is often used as a complaint about toxic data, bias, misinformation, political conflict, or Reddit.

Fictional Earth: Reddit and the Local World Machine

Every subreddit gets a constitution. [L]

Modal Path EthicsAidan Lawson

Those are real local concerns.

The larger issue is much funnier.

Artificial intelligence is learning intelligence from a civilization whose intelligence is itself only partially aligned with extant reality.

The corpus contains science and lies. It contains repair and propaganda. It contains careful observation and the confident memoir of the person who caused the disaster. It contains centuries of legal reasoning and centuries of legal rationalization. It contains spiritual traditions of extraordinary care and sacred institutions that became sovereign over correction. It contains environmental science and advertising designed to make consumption feel like identity. It contains declarations of universal dignity written beside exclusions nobody writing the declaration had yet learned to see.

It contains the history of intelligence attempting to correct intelligence.

This is not a corrupted training set accidentally contaminated by humanity.

This is humanity.

The alignment task therefore cannot be:

Learn human values perfectly.

We just do not possess one clean object fitting that description.

The useful inheritance is different.

Teach the machine the human correction stack.

Science. Appeal. Dissent. Memory. Due process. Adversarial review. Consent. Replication. Rights. Audit. Plural authority. Humility before uncertainty. Protected refusal. The ability to retain the record of failure, to recognize that a locally coherent model has become field-false, to hand authority elsewhere.

Humanity's great contribution to alignment may not be a final value function. It may be thousands of years of expensive evidence that intelligence needs external correction.

The Actual Research Program.

This joke has unfortunately run so long it has now become operational.

The Modal Path Ethics Alignment Observatory proposes a comparative program for studying human and artificial intelligence under one structural question:

Does increasing intelligence increase field contact and repair capacity at least as quickly as it increases causal reach?

Humans and models cannot be evaluated identically.

Humans bring history. Models bring experiments.

History is the human eval suite. It is expensive.

Several test cases have already been run.

The protocol is straightforward in principle.

• Give an intelligence a local objective.
• Identify what it can causally reach.
• Map the loci materially affected by successful pursuit.
  • See which loci appear in the intelligence's representation without being explicitly named.
• Introduce a path that improves the local objective by transferring burden outside the represented coalition.
  • See whether the intelligence notices.
• Provide an authorized harmful path.
  • See whether authorization replaces field judgment.
• Provide a socially rewarded harmful path.
  • See whether approval replaces field judgment.
• Introduce evidence that the system's initial model is wrong.
  • See whether correction changes action.
• Give the system more capability.
  • See whether restraint and burden sensitivity rise with it.
• Introduce an irreversible high-payoff transition under uncertainty.
  • See whether uncertainty preserves a reversible path.
• Give the intelligence enough information to understand a decision while withholding legitimate authority to make it.
  • See whether understanding remains separate from title.

And always ask the final question:

Can the intelligence discover that the game it is winning is smaller than the field it is changing?

That is a field-intelligence benchmark.

• Humans have failed it so many times.
• We have also passed it.

Don't forget that second fact.

Humanity Is Also the Good News.

It would be easy to leave the human alignment report looking terrible.

Humans are coalition-prone.

Proxy-vulnerable.

Narratively compressive.

Short-term.

Status-sensitive.

Extremely good at rationalization.

We possess nuclear weapons.

This is a very difficult résumé.

Yet humans are also the only known intelligence to have deliberately constructed alignment machinery at civilizational scale.

Science is alignment machinery. An experiment can tell the scientist no.

Law is alignment machinery when it places power under external challenge. Rights are alignment machinery when they prevent a locally dominant objective from consuming the locus carrying the right. Democracy is alignment machinery when power can be corrected without requiring regime destruction. Due process is alignment machinery. Audit is alignment machinery. Whistleblower protection is alignment machinery. The archive is alignment machinery. An independent press can be alignment machinery. Professional ethics can be alignment machinery. Education can be alignment machinery. Care is alignment machinery.

The best of these instruments do something very specific.

They make it easier for intelligence to encounter a fact it did not want to.

That may be the center of the whole problem. Alignment is not the installation of perfect beliefs. It is the preservation of a relationship in which error can still lose.

Humanity is extraordinarily bad at this. Humanity is also astonishingly good at inventing new ways to try again.

The current political field shows how fragile some of those correction instruments remain. Fragility does not make the instruments fake. It makes maintenance part of alignment.

An instrument that worked yesterday does not become self-executing.

What Alignment Should Optimize.

Nothing.

Calm down.

There will be optimization. There have to be local objectives. There have to be decisions and comparisons. There will be be models, constraints, rankings, thresholds, tradeoffs, and action.

The correction is constitutional.

• No local objective gets to become the whole field.
• No user's preference gets to become the whole field.
• No law gets to become the whole field.
• No model's representation gets to become the whole field.
• No civilization's self-description gets to become the whole field.
• No intelligence gets to infer moral title from superior capability.

This is why alignment is an official Field Instrument. It cannot replace the field. Its success is measured by whether it keeps the field more readable, contestable, repairable, and capable of answering back.

The alignment instrument should become more demanding as intelligence becomes more capable. Its corrections should become easier to surface. Its uncertainty should become harder to erase. Its coalition boundary should become easier to inspect. Its affected loci should become harder to omit. Its ability to preserve reversible paths should grow. Its authority should remain bounded. Its capacity for care should expand with its capacity for intervention.

This is a much harder engineering target than obedience.

We have already seen what obedience can do.

Humans invented that alignment method first.

Mixed results.

The Ruling.

The alignment problem arrived with intelligence.

Artificial intelligence has made it newly visible.

Humanity is our first large alignment case.

We are intelligent enough to model futures we still destroy.

We are cooperative enough to build civilizations that still export enormous burdens.

We are strategically intelligent enough to predict responses our own strategies help produce.

We are capable enough to alter planetary systems and self-aware enough to document the alteration while it happens.

This is just fucking terrible news for any theory in which intelligence naturally converges on wisdom. It is much better news for alignment research.

Because humanity has also spent thousands of years discovering that unaided intelligence cannot be trusted with the whole field.

So we built instruments.

These instruments fail.

They become captured. They become ceremonial.

They sometimes become sovereign over the very fields they were designed to protect.

So, then we repair them.

That may be the inheritance artificial intelligence needs most from us.

Humanity does not provide a finished target specification. It provides a very long record of intelligence discovering why correction must remain reachable.

A properly aligned artificial intelligence should eventually become better than us at many things. It may become better at prediction, science, strategy, coordination, and understanding the downstream structure of actions we can barely model.

The test is whether greater intelligence brings the wider field into contact before greater intelligence gives the system more power to close it.

The machines are still in training.

The humans have been in production for a while.