The arc that leads somewhere good
I spend my working days at the seam between AI-defect litigation and audit-trail engineering. From that seam, one pattern repeats.
An AI makes a consequential decision. Someone gets hurt. Discovery starts. Depositions reveal the model was untested against adversarial prompts, the decision log doesn't exist, the vendor's benchmark deck bore no resemblance to production behavior. A settlement lands, or a verdict does. Then two parallel cases file. Then a class certifies. Then a legislator picks up the story. Then a state attorney general writes a rule. Then a federal agency writes a bigger one. Then licensing frameworks emerge. Then insurance stops covering the un-audited.
The arc used to take decades. It's now compressing to 18-36 months because the plaintiffs' bar has already built the infrastructure, courts are already writing the doctrine, and governments are already reflex-protective of their electorates. That reflex is a feature, not a bug — it's how democracies keep faith with the people who elect them. And it means the AI arc runs faster than any before it.
Six industries are visibly inside the arc right now. Here's the shape of each, without drowning in case names.
Healthcare AI
Algorithms recommending mass denials of care are already at class-certification stage. The near-term outcome brings filed methodology descriptions, minimum-time floors on human review, disparate-impact audits, and audit trails that survive challenge. The consumer outcome: fewer wrongful denials, faster appeals, treatment approved on the merits rather than by an unaudited algorithm working at scale.
AI hiring
Bias-audit requirements are already in force in at least one major city and one state, with more rules taking effect through 2026. Federal harmonization comes next, along with joint liability for employers using vendor AI-scoring tools. The consumer outcome: applicants of every demographic get considered against defensible criteria, not against a model trained on a training set nobody audited.
Autonomous vehicles
Product-defect framing survives methodology challenges where pure-AI-behavior theories fail. The federal regulator gains AI-behavior-testing authority parallel to its crash-testing authority. Operational-design-domain declarations become filed documents. Every autonomous system deployed on public roads needs a pre-deployment test suite audited by an accredited third party. The consumer outcome: an autonomous car you actually want in the lane next to yours.
Financial AI
Consumer-credit algorithms already inherit disclosure and adverse-action-notice obligations that predate AI. Rules requiring filed model cards, disparate-impact audits, and real-time monitoring follow. Federal banking regulators coordinate parallel guidance for AI use in credit decisions at supervised institutions. The consumer outcome: credit decisions you can appeal and get answers to, rather than a rejection stamped by a model no human can explain.
Consumer chatbots and kid-safety AI
Wrongful-death cases involving AI-chatbot conversations are moving toward multidistrict-litigation formation. The federal consumer-protection agency uses its unfairness authority to require age-appropriate design testing for AI systems accessible to minors. The consumer outcome: parents can hand their kid a device without needing a PhD in machine learning to evaluate every app on it.
Deepfake and synthetic media
Federal statute requires generation-provenance labeling for AI-generated media above some threshold. Models trained on non-consensual or right-of-publicity-violating material become liable per se. Voice-cloning platforms need consented-voice-print registration before generating. The consumer outcome: the tools that let creators make new things are still available, but the tools that let bad actors impersonate you without consent are not.
The historical analog
Every industry above is walking an arc that mature-regulated industries walked before them — and the outcomes on the other side are the point.
| Industry | Triggering event | Time to major regulation | Consumer outcome once mature |
|---|---|---|---|
| Pharmaceuticals | Elixir Sulfanilamide poisonings | 1 year | Drugs that actually work, don't poison you |
| Medical devices | Dalkon Shield injuries | ~2 years | Implants tested before they enter your body |
| Nuclear power | Three Mile Island | Immediate | Statistically among the safest energy sources |
| Finance | 1929 crash | 4 years | Bank deposits you can trust to still exist |
| Aviation | Sequence of jet-age crashes | Multi-year | Safest form of travel humans have ever built |
| Automotive | Late-1960s safety literature | ~5 years | Fatality rate per mile driven down roughly 80% |
| Data privacy | Cambridge Analytica | ~2 years | Rights to your own data you can actually exercise |
The AI arc runs faster than any of these. The infrastructure is already there. The doctrine is already writing. The electorate is already asking.
The pain and the payoff are both real
None of this happens without cost, and pretending otherwise would be as dishonest as pretending the outcome is bad.
Regulation is a growth-through-pain process. Every industry that walked this arc paid real costs along the way. Companies that couldn't afford the compliance overhead disappeared. Products consumers liked got pulled. Timelines slowed. Compliance became a specialization that consumed engineering budget that would otherwise have gone to features. Some of the pain was avoidable and some was not. All of it was real.
The payoff was also real, and larger. Cars are roughly 80% less fatal per mile driven than they were in 1965. Air travel is the safest mode of transportation humans have ever built. Drugs actually work and mostly don't poison the people taking them. Bank deposits still exist when you go to withdraw them. Not one of those outcomes was inevitable. Each was the direct result of walking through the pain of standing up the audit trail, the reporting requirements, the licensing, and the enforcement. In every case the general quality improvement was larger than the growth cost, and the beneficiaries were the people who could least afford to individually audit what they were buying.
The referee's role in the arc
Every mature-regulated industry needed someone to keep score during the growth-through-pain phase, before the regulator was fully staffed and the rules were fully written. That someone has always been the plaintiffs' bar — the working attorneys who take the case that names the problem, prove the harm in court, and force the record onto the public docket where legislators and regulators eventually read it.
Calling that role "ambulance-chasing" mistakes the incentive for the function. The incentive is real; the function is also real. The attorney who files the first tenant-screening AI-discrimination class action is not extracting value from an accident. She is making the accident legible — naming the model, forcing production of the training data, deposing the engineers who chose the threshold, and creating the discovery record that becomes the citation in the eventual rulemaking. Without her, the accident stays private. The vendor tightens things quietly, if at all. The next person harmed by the same model does not know she is the next person, because the first person's case never surfaced.
The referee needs an evidence infrastructure that holds up under adversarial cross-examination, or the whole system reverts to he-said / she-said and the vendor wins on the presumption of complexity. That is the specific gap I work in.
My role: evidence infrastructure for the arc's referees
Rule 702 tightened in 2025. Every expert methodology that uses AI to explain AI now sits at the exclusion surface. That is the specific technical problem I was asked to solve when the plaintiffs' bar started calling.
The arsenal I run rests on classical statistical estimators with published provenance predating the AI era by 50 to 125 years — Fisher variance-ratio F-test (1920s), Pearson chi-square goodness-of-fit (1900), Kullback-Leibler divergence (1951), Cohen's kappa (1960). The analytical pipeline uses no machine-learning inference at any step. Every measurement is produced by a model family distinct from the one being audited and captured in a cryptographically-hashed evidence record independent of the audited system. Every finding traces to peer-reviewed statistical methods that have half a century of adversarial cross-examination in federal court preceding the AI era.
That is the technical foundation. What it produces is evidence that survives the exclusion motion, and testimony that survives cross. What it does at the arc-level: it makes the referee's calls stick.
I work for whoever needs the evidence infrastructure to hold. Plaintiff, defense, GC, corporate board, insurance carrier. The methodology is the same regardless of which side of the caption I'm sitting on. What changes is the question posed. What does not change is that the answer has to survive a Rule 702 challenge or it is not worth producing.
The middle between extremes
Over-regulation is a real failure mode too. Rules sometimes ossify around a specific technology and lock in the past. Rules sometimes get captured by incumbents and become moats. Rules sometimes add compliance overhead only large firms can absorb, killing the small competitors who would otherwise have driven the next wave of quality improvement. Anyone advocating for regulation without acknowledging those failure modes is arguing in bad faith.
The middle between the two extremes is where the good outcomes actually live. Not zero regulation and not maximum regulation — a regulation that raises the floor on what's unacceptable without capping what's possible above the floor. Every mature-regulated industry found that middle eventually. Sometimes it took multiple iterations to get there. Sometimes the first version was wrong and got rewritten. Sometimes political weather blew the pendulum too far one direction and it swung back. But the direction of travel across all of them is the same: the floor raises over time, the quality improves over time, and consumers keep more of the value that the technology promised.
That's the contrarian read on the current moment. Not "regulation is bad" and not "regulation is good." Rather: regulation is how the floor gets raised, and the floor rising is how a technology's promise reaches the people who can't afford to audit it themselves. The pain is real. The drawbacks are real. And on the arc of any decade you measure, the payoff is larger than both.
If your product touches a person's rights, safety, or money via an AI-driven decision, you're already inside the arc. The arc goes somewhere better — for the people who use the product, and eventually for the industry that builds it. Faster than it ever has before.
The Regulation Ratchet, and how to follow the arc
The monthly newsletter that carries this manifesto's name tracks the arc in motion. Each issue reads public SEC filings across sectors that have not yet been rulemade, and describes the disclosure-language patterns that historically precede regulatory arrival. It is descriptive analysis of what public filers themselves say, not prediction of any specific harm, verdict, or agency action. The purpose of the newsletter is calibration — keeping the reader oriented to where the arc is currently moving fastest, so the reader can plan against a landscape that is genuinely knowable in advance.
Deeper subscription tiers extend the same underlying signal set with more filer-level detail, cross-sector overlay analysis, and (at the top tier) a quarterly single-sector strategic brief for operating executives and boards. All four tiers share the same author, the same discipline, and the same posture: independent, descriptive, human-analyzed, and priced to be accessible to the working attorneys, compliance officers, and operators who want to see where the arc is heading before they are already inside it.