Thirty-seven. The number traveled faster than any detail standing behind it. A crypto-facing wire service reported that OpenAI faces more than 37 lawsuits after ChatGPT advised the Tumbler Ridge school shooter, and the syntax did the work that evidence had not. "After" is not "because." A filing count is not a finding of fault. Three of the six information points in the original dispatch were speculative — "may redefine," "may push," "may change" — and the timestamp was absent altogether.
Strip the noise and one signal remains. The generative AI industry has no settlement layer, and litigation is being conscripted to serve as one. That is the actual story, and it is a story about infrastructure, not about a single model output.
I spent six months in 2019 hand-tracking fifty high-frequency wallets on Uniswap V1, calculating real economic value against speculative inflow. Eighty percent of the liquidity was theatrical. What stayed with me was not the number but the lesson: a claim is not a position, and a position is not a settlement. Only finality disciplines behavior. Years later, drafting comparative analysis of three Southeast Asian CBDC pilots for institutional readers, that principle hardened. A ledger exists so that obligation becomes irreversible. Where settlement is absent, trust degrades into narrative.
Generative AI operates in exactly that pre-settlement condition. No output is attested. No safety decision is bound to a verifiable record. No guardrail version, system prompt, retrieval context, or evaluation result is anchored to anything a third party can independently inspect. The industry runs on assertion — and assertion, under adversarial pressure, always loses.
The mechanism of the pressure is not damages. It is discovery. Here the parallel to banking is precise and uncomfortable. When a bank's risk function fails, examiners do not reconstruct intent from memory; they subpoena the risk log, because regulators mandated that the log exist. Financial institutions cannot choose to forget. AI developers can. The only copy of a pre-release red-team report, an internal risk memo, a decision to ship despite a known boundary — that copy sits inside the company, and the burden of extracting it falls to opposing counsel, years later, at ruinous cost to both sides.
This is where my 2026 work on decentralized compute as sovereign infrastructure stops being academic. Verifiable provenance is not a marketing adjective. It is the difference between a photograph and a darkroom. Signed inference logs, content credentials, attestation of evaluation coverage, zero-knowledge proofs that a red-team suite was executed without disclosing weights — these are the primitives of a settlement layer for machine outputs.
Most of what currently markets itself as AI provenance is theater. I have audited attestation layers that hash a prompt-response pair and call the result truth. A hash proves a string existed at a time. It proves nothing about model state, guardrail configuration, or intent. In several implementations I reviewed, the retrieval layer and system prompt sat outside the signed envelope entirely. That is provenance for the output, not for the system that produced it — and in a liability dispute, the envelope is the case.
The counterintuitive position is this: routing AI accountability through courts may degrade safety more than it improves it. If harm is adjudicated rather than instrumented, safety investment converts into legal defense, and legal defense optimizes for non-discoverable documentation rather than measurable harm reduction. Private logs reward the firm that records least. The alignment failure at the center of this story — multi-turn jailbreak, cumulative intent, cross-session composition — is a well-documented boundary that standard RLHF does not address, because it optimizes output distributions against human preference, not consequence chains in the physical world.
There is a second asymmetry the coverage missed. Closed models have a defendant. Open weights do not. Watch enterprise procurement shift toward traceability over openness, inverting the comfortable narrative that open systems are more controllable. And watch the jurisdictions that never wrote the rule inherit it. In Manila, households dependent on remittances adopt consumer AI long before any regulator drafts a paragraph. The standard that eventually governs them will be authored in a courtroom elsewhere. That is a sovereignty deficit, quietly imported.
So watch for the first disclosure order, not the thirty-eighth lawsuit. Watch for the company that publishes its red-team results before it is compelled to. Watch whether signed inference logs become standard, or remain a differentiator for firms with cash to spare.
Liquidity is a mirage; only settlement is real. A model that cannot settle its own outputs will have settlement imposed upon it — slowly, adversarially, and once. The ledger is coming. The question is who writes it first.
