Action guide · May 2026 · v0.1
Part 1 mapped why AI companies cannot reliably self-govern. This guide trains your eye to see leverage where it actually lives. It opens by naming why most AI announcements feel hollow, walks the four tiers through real 2024-2026 examples paired with their look-alikes, and closes by helping you locate where you sit relative to the tiers. From recognition, not optimism.
Frustrated About AI?
Frustration about AI is the right read. The emperors leading the race have no clothes.
OpenAI signed the open letter calling AI a civilizational extinction risk, then shipped a model its own CEO described as “talking to a PhD-level expert in any topic,” and announced the next one would arrive faster. AI, they say, will fix AI’s problems. The mission statements of these companies speak of curing cancer, of intelligence as an abundance distributed to everyone, of the most important technology of the century.
The reality is in the earnings calls. Salesforce announces it is cutting customer service from nine thousand people to five thousand because, the company says, it needs less heads. Microsoft tells a developer conference that thirty percent of its code is now AI-written; two weeks later the layoffs land in its home state, concentrated in engineering. Amazon opens 2026 by cutting sixteen thousand corporate jobs and naming AI in the announcement. The cure-cancer technology becomes, in a quarter, the rationale for cutting headcount. AI summaries arrive in your search results without permission. A press release announces an internal oversight board.
These are the same companies that sign the safety frameworks, that appear at the summits, that publish the responsible-development pledges.
What the actions show is plain: an unwavering attempt to divorce labor from capital. AI is the lever. The mission statements decorate the move. The earnings calls disclose it. The data centers go up in low-income neighborhoods where political pushback is weakest. The expert workers being destabilized by the technology are hired cheap to label its training data, refining the systems built to replace them. The technology pitched as society’s salvation is making society more vulnerable, at a pace nothing else has matched.
You watch them and feel a particular kind of tired. The kind that belongs to people who can read the room. It is the recognition of a contradiction, and the futile attempt to play along.
The companies racing to build civilization-changing technology have not developed civilization-changing discipline. Anthropic published the first version of its Responsible Scaling Policy in autumn 2023, committing to pause development if its own evaluations suggested capability had outrun control. After three revisions, the binding pause was gone. By February 2026, Jared Kaplan was telling TIME the company “didn’t really feel” the pause commitment made sense with competitors blazing ahead.
Less than half a decade. One company, one policy, three revisions, gone. A fourth revision arrived in April 2026. The pause stayed gone.
The pressure Kaplan names is real. The lab is not lying about it. The system the labs operate inside has its own behavior, and they respond to it like the rest of us do. Knowing that behavior lets you act on it. Not knowing it is how the drift continues.
You do not need to settle the motive question to act. There is a framework that works without it. It is older than AI and bigger: it asks where in any system a small push produces structural change, and ranks those places by power. Some you can reach alone. Others require people working together.
That seeing is what the rest of this is for.
Beneath the Search Result
If AI is the new economic engine, where are we headed?
Nvidia became the first company in history worth five trillion dollars; by the summer of 2026 it was trading the title of most valuable company back and forth with Apple. OpenAI committed half a trillion dollars to build AI infrastructure with SoftBank, Oracle, and MGX. Anthropic’s valuation went from 965 billion in under a year. The four largest cloud companies planned roughly seven hundred billion dollars in capital spending for 2026, nearly double the year before, most of it for AI. Apollo’s chief economist Torsten Slok called the AI buildout the foundational pillar of US growth; by 2026 it accounted for roughly half of that growth. In July, Slok himself began warning of a painful repricing. A contraction would not be quiet.
On May 14, 2024, Google began rolling out AI Overviews to all US users by default. You did not turn it on. You did not get an opt-out checkbox. One day the answer at the top of your screen was no longer a website someone had written. It was a paragraph a model had assembled.
It told someone to put glue on pizza so the cheese would stick, and the screenshot went viral. A Minnesota solar company called Wolf River Electric sued Google in March 2025, alleging an Overview had fabricated an attorney general investigation against them and cost the business between 210 million. By mid-2025, Pew Research found publisher click-through had nearly halved when an Overview appeared, falling from fifteen percent to eight. In May 2026, a Munich court held Google liable for what an Overview said about a business, the first ruling of its kind. Google is appealing.
The same year, Meta added an AI assistant to WhatsApp with no off switch. LinkedIn began feeding member posts into generative AI training with the opt-out switch already flipped on, and updated its terms of service only after journalists noticed. Microsoft bundled Copilot into Microsoft 365 and raised the subscription price before most users had decided whether they wanted it, then raised prices again in July 2026, up to forty-three percent on some plans, with Copilot folded into the base subscription.
Before AI can be used, it has to be trained. OpenAI, Anthropic, and Google trained their foundation models on text, code, and images they did not pay for and did not ask permission to take. The New York Times sued. Music labels sued. John Carreyrou, the reporter who broke Theranos, joined a copyright lawsuit alleging the labs had trained on pirated copies of his book Bad Blood. In July 2026 a federal judge gave final approval to Anthropic’s $1.5 billion settlement with authors whose pirated books had trained its models, the largest copyright settlement on record. In January 2025, Sam Altman pledged to crack down on rivals distilling OpenAI’s outputs in the same week his lawyers were arguing in court that training on copyrighted books was fair use no AI lab could survive without. What they took from the public they called training. What was taken from them they called theft. When the bill finally came due, it was priced at three thousand dollars a book.
Each company also needs its AI to be in use. Investors are reading user counts. The data from active use feeds the next training run. The next training run requires the next data center. The next data center requires the next round of Nvidia chips. So the companies push their AI into products the user cannot easily leave. Engagement is the metric. Silence is read as agreement.
Days after the failures went viral, Google’s head of Search, Liz Reid, published a defense of the rollout. The frame was familiar. The system was working. The failures were rough edges. Whether the public had been entitled to refuse was not the question on the page. That AI is inevitable. That asking permission would slow it down. That what they took from the public belongs to no one, and what they made from it belongs to them. That a user who keeps using the product has consented to whatever the product becomes. That the economy needs this growth and could not absorb a contraction. That the cost of an Overview that gets it wrong is paid by the reader, the publisher, or the small business it defamed, not the platform.
None of these beliefs has been put to a vote. All of them have already shipped.
Constants, buffers, and infrastructure
Adjustable settings and physical infrastructure — real effort, real engineering, rarely shifts what the system actually does
These levers are the most common targets of “AI safety” activity because they’re the most accessible. They address specific problems without changing the underlying incentives, competitive dynamics, or goals. A system absorbs parameter changes and continues.
Constants & Parameters
The knobs and dials — adjustable settings that can be tweaked without changing the system’s structure
Content filters, rate limits, safety-classifier thresholds, usage caps — the adjustable settings of any system. Most “AI safety” announcements live here: real engineering work that leaves incentives, goals, and competitive dynamics unchanged.
Most viable actors: Product and safety teams at labs; regulators setting technical standards. Anyone who ships AI products can move this lever.
How you encounter this: Every time a lab announces “improved safety” with a new filter or guardrail. Or when a company says a harmful output was a “model behavior issue” that’s been patched.
What changing it does — and doesn’t: Parameter fixes address the specific problem you already know about. They don’t prevent the next category of problem you haven’t encountered yet. The useful question when a company announces a safety improvement: does this change who’s liable, who audits, or what gates must be cleared before deployment — or does it just adjust a threshold?
Buffer Sizes
How much slack exists — the breathing room between “something went wrong” and “it cascades”
Buffers are the system’s shock absorbers: review periods before launch, safety team headcount relative to product team size, the time window between capability development and deployment. They erode under competitive pressure because adding slack looks like giving up ground.
Most viable actors: Lab safety team leads, boards of directors, institutional investors, congressional appropriators. External pressure is the main thing that keeps buffers from being quietly eliminated.
How you encounter this: When a lab shortens its pre-deployment review period. When a safety team’s headcount shrinks while the product team grows. When a published “pause commitment” gets quietly removed from policy documentation.
What changing it does — and doesn’t: Larger buffers give oversight a fighting chance to catch problems before they reach users at scale. They don’t fix the underlying incentive dynamics — they extend the runway. Buffer erosion is one of the clearest early-warning signals that competitive pressure is overriding stated safety commitments.
Physical Infrastructure
What’s built and where — the hard constraints that determine what’s even possible to run
Chip fabrication concentration, electricity grid capacity, data center permitting — physical structures that constrain AI development regardless of policy. The electricity grid is already acting as an unintentional brake: interconnection queues averaging 7–10 years in key markets are physically limiting deployment in ways no regulation achieves.
Most viable actors: FERC commissioners, state utility boards, trade officials, chip export control policymakers. These decisions get made in venues with no “AI” label on the door.
How you encounter this: In electricity rate cases, chip export control debates, and data center siting disputes — decided by utility commissioners and trade officials, not AI researchers.
What changing it does — and doesn’t: Infrastructure changes are slow and durable. Reshaping who controls chip manufacturing or grid access shifts the long-run dynamics of who can build what — not whether building happens. The electricity grid may turn out to be a more binding constraint on AI deployment than any regulation specifically designed for AI.
Delays, brakes, and engine speed
Response delays, brake strength, and acceleration rate — changing these meaningfully shifts system behavior
These levers work on how fast the system moves and how hard it can be stopped. Shorter delays make harm visible faster. Stronger brakes make recklessness financially costly. Slower acceleration gives oversight time to keep pace. All three face significant institutional resistance.
Length of Delays
How long before consequences arrive — long delays make problems invisible until they’re already large
Models deploy on 2–4 month cycles. Documented harms arrive months to years later. Regulatory response takes years. By the time a specific harm is attributed and acted on, the model that caused it has been superseded. This delay is structural — it’s a key reason public concern rarely translates into policy change fast enough to matter.
Most viable actors: Journalists and researchers who document harm; regulatory staff who can mandate incident reporting; standards bodies that define harm taxonomies. Aviation and pharma already have working models here.
How you encounter this: When you read about harm from an AI system deployed 18 months ago — and the company has since released two newer models. The accountability gap is the delay made visible.
What changing it does — and doesn’t: Shortening the detection-to-response delay — through mandatory incident reporting, standardized harm taxonomies, rapid audit requirements — compresses the accountability cycle. It doesn’t prevent harm. It makes harm visible fast enough for the oversight system to respond while it can still matter.
Strength of the Brakes
How effectively counterforces work — when brakes are too weak, acceleration goes unchecked
The oversight brake (harm → regulatory response → slowdown) is weak: underfunded agencies, voluntary self-reporting, no mandatory independent audits. The insurance brake barely exists at scale. The legal brake is forming but hasn’t yet produced the precedent that would change behavior industry-wide.
Most viable actors: Plaintiffs’ attorneys pursuing product liability, insurance actuaries pricing AI risk, appellate judges, congressional committee staff. One court ruling can strengthen a brake more than years of rulemaking.
How you encounter this: When a regulatory body issues AI guidance that companies are not required to follow. Or when an insurance policy specifically excludes AI-native liabilities like hallucinations or biased outputs. These are brakes that aren’t connected to anything.
What changing it does — and doesn’t: Strengthening a brake — making liability attach to specific harms, making audits mandatory, making insurance cover AI risk — doesn’t stop development. It makes recklessness financially costly. A single appellate liability ruling holding AI output to product-liability standards would strengthen the legal brake faster than a decade of regulatory rulemaking.
Speed of the Engine
How fast the growth loop is running — and what slowing the acceleration (not stopping it) looks like
The capital loop (capability → investment → compute → capability) has very high acceleration because every actor in the loop profits from speed. Reducing the rate of acceleration doesn’t stop the loop — it gives other parts of the system time to keep pace. These proposals exist in early legislative form and face predictable opposition from everyone who benefits from the current pace.
Most viable actors: Finance ministry and tax committee legislators, labor economists advising government, national AI councils. This lever requires political will that almost no actor currently has incentive to supply.
How you encounter this: In proposals for AI windfall taxes, compute levies, or profit-sharing mandates — and in the industry lobbying that reliably defeats them. The companies most vocal against these measures sit at the center of the capital loop and benefit most from its acceleration.
What changing it does — and doesn’t: Slowing the acceleration extends the time available for safety research and governance to keep pace with capability. It doesn’t stop development, and it faces the most politically powerful opposition of any lever in this tier. The value isn’t stopping the loop — it’s buying time for everything else to work.
Visibility, rules, and rule-makers
Information, enforceable rules, and who writes them — powerful, underused, often actively resisted
This tier is where individuals and institutions can create structural change in the next two to three years. It requires access — to journalists, to courts, to regulatory proceedings, or to workplaces deploying AI. The access itself is the constraint. These levers are available to more people than usually act on them.
Visibility
Who can see what — because making harm visible changes behavior more reliably than rules do
Labs self-report safety evaluations. Harms aren’t systematically tracked. Even benchmark scores — the primary public signal for model safety — were gamed: labs submitted private variants until scores were high, then withheld lower-performing ones. The gap between stated and actual safety is only detectable if harms are measured. Right now, most aren’t.
Most viable actors: Investigative journalists, academic safety researchers, NGOs, HR and legal professionals inside companies deploying AI. Anyone with standing to demand disclosure can move this lever — this is the most accessible high-leverage point for most people.
How you encounter this: When a company publishes a safety report without disclosing what it tested for, what it found, or what it declined to test. Or when a harmful AI output gets attributed to a “rare edge case” with no data to back the claim.
What changing it does — and doesn’t: Mandatory incident reporting and public safety disclosures don’t prevent harm. They make it impossible to deny. When harms are visible and documented, they create public, legal, and financial pressure that rules alone don’t. Push for disclosure wherever you have standing to ask for it — in vendor contracts, procurement processes, workplace AI policies, and regulatory comment periods.
Enforceable Rules
Laws and liability standards that carry real consequences — not announcements, but rules with teeth
Rules shape behavior only when backed by real consequences: fines that genuinely hurt, liability that attaches to specific decisions, deployment gates that can’t be bypassed. The EU AI Act’s penalty structure activates August 2026 — the first enforceable rule at meaningful scale. The US has no equivalent federal provision for labor protections during AI transitions.
Most viable actors: Trial lawyers and plaintiffs, legislators drafting AI liability bills, EU enforcement officials, labor advocates. Courts and legislatures, not labs, hold the keys to this tier.
How you encounter this: Every time a voluntary commitment is revised downward without consequence. Every time a court case settles before it can establish precedent. The pattern is consistent: where rules have no external enforcement, they bend.
What changing it does — and doesn’t: A court ruling that makes AI deployment financially risky changes what every lab’s legal team tells its product team — faster than any regulatory rulemaking. Labor protection legislation changes what employers can do when deploying AI in ways that affect jobs. Rules with teeth change behavior at the point of decision. This is where concentrated effort from individuals — in courts, legislatures, and regulatory proceedings — produces structural change.
Who Writes the Rules
Who designs the oversight — because whoever designs it usually designs it to protect themselves
The US AI Action Plan was developed with significant lab participation. Labs increasingly author the safety standards they’re evaluated against. A diverse ecosystem that can genuinely self-correct is structurally more robust than a tight oligopoly writing its own standards. The current trend concentrates rule-writing power in fewer hands.
Most viable actors: Academic safety researchers, civil society organizations, independent standards bodies (IEEE, NIST). Participation in regulatory comment processes, standards working groups, and government AI task forces is how this gets diversified.
How you encounter this: When an AI safety standard is authored by the company being evaluated against it. When a government AI task force is composed primarily of industry representatives. When the same lab that builds a model also certifies it safe.
What changing it does — and doesn’t: Diversifying who writes rules changes whose interests the rules protect. It doesn’t guarantee better outcomes. But it makes regulatory capture harder, and preserves the structural diversity that allows genuine self-correction. Regulatory comment submissions, academic participation in standards bodies, and civil society testimony in AI legislation hearings are all underused entry points.
Goals, assumptions, and coordination
What the system is for, the assumptions it runs on, and whether rivals can be made to coordinate
The hardest levers, and the ones with the highest ceiling. These require changing the actual goals of powerful actors, dislodging assumptions that feel like facts, or coordinating entities in direct competition with each other. None of those are easy. All of them involve decisions that can still be influenced.
What the System Is Actually For
The revealed purpose — not the mission statement, but what deployment patterns consistently produce
There are two fundamentally different goals AI deployment could serve: AI as amplifier (workers more productive, value distributes broadly) or AI as replacement (eliminates labor cost, value accrues to capital). The current system is not neutral between them. Deployment pattern after deployment pattern points toward replacement. The stated goal is beneficial AI. The revealed goal is deployment speed and market capture.
Most viable actors: Labor organizers, employment lawyers, ESG-focused institutional investors, legislators on labor and finance committees. Goal change shows up in incentive structures before it shows up in any press release.
How you encounter this: In what companies file with the IRS and SEC (the legal statement of purpose, not their websites). In whether AI deployment agreements include labor consultation requirements. In whether “AI dividend” or profit-sharing proposals are treated as serious policy or dismissed as anti-progress.
What changing it does — and doesn’t: Changing the system’s actual goal requires changing what creates financial consequences. Labor protection law, liability for displacement harms, and profit-sharing mandates can move the goal from “deploy fastest” to “deploy responsibly.” This requires sustained pressure over years — Denmark’s flexicurity model shows it’s achievable at national scale. Goal changes appear in incentive structures before they appear in any press release.
The Unquestioned Assumptions
The ideas so embedded they’re not seen as choices — the water every participant in the system swims in
Two dominant assumptions shape the current system. For labor: “efficiency gains are inherently good and always self-correct” — historically true but only through specific mechanisms (organizing, bargaining, redistribution) that don’t fire automatically for AI displacement. For catastrophic risk: “building more capable AI is inevitable — the only question is who and whether they’re responsible” — a choice the dominant paradigm treats as a fact of physics.
Most viable actors: Academic researchers, science journalists, policy analysts, educators. Paradigm shifts happen when enough specific, documented counterexamples accumulate that the old assumption can no longer explain what’s observed.
How you encounter this: When someone says “you can’t stop progress” as a conversation-ending move. Or “AI will create more jobs than it destroys” without specifying the mechanism, timeline, or distribution. These are paradigm-reinforcing claims that feel like facts.
What changing it does — and doesn’t: Paradigm shifts don’t happen through argument alone. You can contribute by insisting on specificity: for labor claims, ask for the mechanism and the timeline. For safety claims, ask what would count as evidence that the risk is real. Paradigms that survive only through vagueness are closer to shifting than they appear.
The Coordination Problem
Getting rivals to act together before it’s too late — the hardest thing, and the one with the highest ceiling
Civilizational-scale coordination requires three things to be true simultaneously: the threat is undeniable and legible to everyone, the actor set is small enough to negotiate with, and everyone has a face-saving path to change course. AI currently fails all three. But those conditions can be deliberately created — and the decisions being made now about interpretability research, international frameworks, and accountability mechanisms will determine whether coordination becomes possible.
Most viable actors: Heads of state and senior diplomats, frontier lab executives with standing to commit publicly, interpretability researchers whose work makes risk legible. The Bletchley → Seoul → Paris → New Delhi summit chain shows coordination is beginning — still declaratory rather than binding.
How you encounter this: In every international AI summit, treaty negotiation, and export control regime. In whether interpretability research is funded enough to produce verifiable safety claims before capabilities outpace oversight. In whether the dominant framing of AI competition makes cooperation politically possible or impossible.
What changing it does — and doesn’t: Work on the three conditions: fund interpretability research (the work that makes risk legible and undeniable), support international frameworks that include all major state actors, and resist framings that make coordination seem naive or unserious. This leverage point has the highest possible ceiling and the longest timeline. It also can’t wait for everything else to be tried first.
A note on realistic expectations
There is no single leverage point that makes this easy. The higher leverage points require coordination that has never happened at this scale. What this framework gives you is clarity about where pressure produces change. Enforceable Rules and Visibility are where individuals and institutions can create structural change in the next two to three years. The feedback loop between action and consequence is already running. The question is whether the balancing loops get strong enough before the fast loop reaches a threshold they can no longer affect.
Back to Part 1: Can AI Self-Govern? · Leverage points (quick reference) · AI governance series hub
Leverage framework: Donella H. Meadows, Thinking in Systems: A Primer (Chelsea Green, 2008). Twelve leverage points: pp. 145–165.
Sources cited in the analysis are on Part 1 sources and references. This guide focuses on action; the evidence base is in Part 1.
This is an analytical guide applying a systems-thinking framework to publicly documented facts: not financial or investment advice, not legal advice, and not a prediction of specific outcomes.