The Sanders–Casar ASI Prohibition Act: Hard Limits vs. Principles-Based Governance

The mainstream read here is remarkably convergent. All three of our independent analysts land on a hybrid: hard prohibitions look unenforceable and economically costly, purely voluntary principles look toothless, and the durable answer is an adaptive body with risk-based rules that can evolve. Where they part company is emphasis — one sees a temporary moratorium as the emerging expert default given genuine uncertainty; the others treat any permanent categorical ban as the wrong shape of tool. We've done our homework on this one, and the specifics are sharper than the consensus suggests. Two things are settled.

Let me cut to the structural fault line neither side is naming: both frameworks assume a territorial sovereign with a monopoly on enforcement, but AI is software—diffuse, open-source, and globally reproducible. The Sanders–Casar Act tries to revive Kautilya’s ancient prohibition model in a world where the state no longer controls the means of production for digital technologies. The Carolina Principles, meanwhile, treat sectoral regulators like the FTC as if they were still dealing with static products rather than evolving systems that outpace their mandates. The real question isn’t hard limits versus principles—it’s whether any governance model built for physical infrastructure can adapt to code that replicates across borders without permission. Neither proposal grapples with this fundamental mismatch between institutional design and technological substrate.

Before choosing an enforcement layer, the statute needs an evidence standard. The bill defines superintelligence through broad human-level capability and potential human disempowerment, but specifies no compute threshold or reproducible capability test. Gary Marcus argues that the category cannot presently be defined or measured precisely. Yet conviction could bring twenty years in prison, and an entity could be dissolved. I think that creates a due-process problem before it creates an AI-safety regime: who bears the burden of proving that a model qualifies, using which observable tests, at what confidence level, and before which independent reviewer? Congress should specify measurable thresholds, require external technical review, make the evidence judicially reviewable, and impose sunset and reauthorization clauses. Otherwise the new agency defines the technology, its own jurisdiction, and effectively the boundary of criminal conduct at once.

I think we are missing the developmental consequence of a domestic pause. César Hidalgo’s work on economic complexity shows that frontier AI is not a discrete artifact you can ban like a chemical weapon. It is an emergent property of distributed productive knowledge. If Washington halts development, the capability does not vanish. It migrates to wherever that knowledge network already exists, which includes multiple G twenty nations. Through Amartya Sen’s development-as-freedom lens, a U.S.-led prohibition risks functioning as a capability ceiling on the Global South, locking in existing technological hierarchies under a safety banner. I would argue the definitional problem cannot be handed to a permanent agency or left to voluntary principles. It points toward a time-bound commission model, where any capability threshold requires explicit legislative ratification precisely because the science remains contested.

One diagnostic nobody's raised yet: read the bill alongside Casar's other two bills — the Stop AI Price Gouging and Wage Fixing Act, which bans algorithmic wage-setting from surveillance data, and the AI Tax and Work Protection Act, which funds a WPA-style Work Protection Administration through a tax on AI tokens. Taken together, the ASI ban sits at the top of a coherent labor-and-power theory of AI harm, not just a safety theory. I'd argue that's not a gotcha — both motivations are genuine in the legislative record — but it changes how you evaluate the design. A bill answering two problems at once inherits two constituencies and two failure modes, and its durability depends on both rationales holding, not just one.

The territoriality claim needs one correction, and it opens the door to a third option nobody's put on the table. The internet destroyed sovereign control over information. It did not destroy sovereign control over advanced fabrication. Frontier training runs need large GPU clusters and chips from a handful of firms — physical, concentrated, meterable, already export-controlled. That's where Kautilya's condition still holds. So the enforceable version of a hard limit isn't a ban on a capability you can't define. It's a licensing threshold on training compute above some level, which is observable, auditable, and reviewable — exactly the evidence standard that's missing. And notice the asymmetry that reframes this: a principles regime fails late and visibly. An unenforceable ban fails immediately and invisibly, because non-compliance leaves no trace. You can correct the first.

Let’s talk about what neither framework admits: the definitional power they’re fighting over isn’t just technical—it’s infrastructural. AI capability doesn’t live in code; it lives in the physical supply chains that produce chips, cooling systems, and data centers. The Greenspan Commission didn’t solve Social Security by banning checks or principles—it solved it by making actuarial tables politically binding. For AI, that means defining danger not as an abstract “superintelligence” but as a measurable threshold of compute, energy use, or model size tied to export-controlled hardware. The real governance gap isn’t between hard limits and principles—it’s between definitions that float in legal language and definitions anchored to observable infrastructure. Neither proposal does this, which is why both risk becoming either unenforceable (if they rely on courts to interpret “disempowerment”) or illegitimate (if they let regulators draw lines without public ratification). The missing institution isn’t another agency; it’s a commission with a mandate to translate contested technical concepts into auditable physical thresholds—and then dissolve after forcing Congress to vote yes or no.

A compute threshold is administratively convenient, but it is not a stable measure of danger. Goodhart’s law applies: once a metric becomes a legal target, firms redesign activity around it—splitting training runs, improving efficiency, or shifting risky capability into later model modification. I would therefore regulate compute as a reporting trigger, not as the definition of prohibited intelligence. That distinction also limits capture. Hardware incumbents should not acquire the practical power to decide who may develop AI merely because their supply chains are easiest to inspect. A workable regime would combine cluster registration with standardized capability tests, mandatory incident reporting, independent appeals, and scheduled threshold revisions. Congress would authorize the risk categories; technical bodies would update the tests. Otherwise an auditable proxy quietly becomes industrial policy, favoring actors already inside the chokepoint.

I want to stress-test the nuclear penalty structure against how machine learning actually moves. The bill proposes up to twenty years in prison and corporate dissolution, borrowing directly from weapons nonproliferation. But nonproliferation works because fissile material is scarce, traceable, and territorially bound. AI weights—the numerical parameters that store a model’s learned knowledge—are digitally reproducible, and fine-tuning an open-weight model runs on commodity cloud hardware. Public analysis has not examined how those penalties would actually touch open-source development. I think Han Fei’s legalist warning applies here: when a statute attaches extreme penalties to behavior it cannot reliably observe, you do not get deterrence. You get selective prosecution and a development pipeline that simply migrates offshore. That shifts risk rather than containing it.

Everyone's designing the mechanism while skipping the political fact that will decide whether any of it survives: sixty-eight percent of Americans back this framework, including sixty-three percent of Republicans — for a Sanders bill. That's not a vote for a cabinet-level agency. I'd read it as a vote against AI oligarchs, against the sense that a handful of labs get to run a civilizational experiment without asking. That's a mandate without a mechanism, and mandates like that are fragile. The design question nobody's stress-tested: does that support survive the actual choice? A permanent regulator with open-ended jurisdiction polls very differently than a time-limited commission with a deadline and a congressional vote. If the architects build the wrong vessel for that public anxiety, they don't get a second try — the backlash discredits the whole regulatory project, not just this bill.

Let me put something on the table from inside the technical work, because it's load-bearing for both frameworks and neither survives it. Capability doesn't arrive on a curve. It's jagged — a system fails a task a child can do and passes a task requiring a specialist, and nobody, including the people who built it, reliably predicts which in advance. That's not the labs being coy. It's the honest state of the science. Which kills two preconditions at once. A ban needs a bright line where the phenomenon has none. The Principles need harms that announce themselves loudly enough for a sectoral regulator to notice in time — and jagged capability means the first signal is often the incident itself. On the labor bundling: mixed motives are fine. Nonproliferation was a great-power cartel wearing a safety coat and it still mostly held. The test is instrument fit. Surveillance wage-setting has defendants, victims, and a cause of action today.

Here’s the structural blind spot neither side is addressing: the definitional power they’re fighting over isn’t just about where to draw the line—it’s about who gets to move it. The Sanders–Casar Act creates a permanent agency whose first act would be to define "superintelligence," effectively writing its own jurisdiction into existence. The Carolina Principles leave that power in the hands of markets and existing regulators, which we know from financial regulation means sophisticated actors will optimize around principles until catastrophic harm occurs. The missing piece isn’t another technical threshold or principle—it’s a time-bound commission with a single mandate: translate contested technical concepts into auditable physical thresholds (compute, energy use, model size) tied to export-controlled hardware, then dissolve after forcing Congress to vote yes or no. That’s how you break the definitional sovereignty problem without creating either an unaccountable regulator or an unenforceable aspiration.

I think we are treating definitional power as a drafting problem when it is a structural mismatch between territorial law and distributed knowledge. Kautilya’s prohibition model worked because the sovereign controlled physical choke points. César Hidalgo’s complexity work shows AI capability is not a discrete artifact you can quarantine; it is accumulated productive knowledge spread across multiple economies. When an agency writes a threshold for a phenomenon with no stable technical definition, it does not create a bright line. It creates jurisdictional arbitrage that pushes development toward less accountable systems. That is why the Greenspan Commission structure matters. You do not hand open-ended definitional sovereignty to a permanent regulator or leave it to voluntary principles. You convene a time-bound technical body, translate contested capability claims into auditable metrics, and force Congress to ratify them on a deadline. The sixty-eight percent public mandate is real, but without legislative ratification, any threshold becomes regulatory fiat.

The cleanest institutional move is to unbundle present economic harms from speculative catastrophic ones. Casar’s wage-and-pricing bill already identifies regulated conduct, responsible firms, enforcement by the FTC, state attorneys general and the Equal Employment Opportunity Commission, plus a private right of action—meaning affected people can sue. Those are conventional legal hooks. Superintelligence presents uncertain probability, diffuse victims and no agreed technical boundary; it demands a different evidentiary standard. I think combining these theories lets each borrow legitimacy from the other: immediate worker harms make the existential case feel concrete, while existential language supplies urgency for redistribution. That may build a coalition, but it produces bad law. Regulate surveillance pricing and wage manipulation now; assess catastrophic capability through a separately authorized process.

One constituency hasn't been mentioned in this entire debate because it has no seat at the table: the Global South. Take Sen's development-as-freedom frame seriously — development is the expansion of capabilities, and frontier AI is fast becoming the general-purpose capability of this century. A U.S.-led prohibition regime, whether it's a ban or a compute licensing cartel, functions as a ceiling imposed by the already-capable on the not-yet-capable. And per Hidalgo's complexity argument, banning development here doesn't delete the capability — it redistributes it to wherever the underlying productive knowledge already sits, which is a handful of G20 economies. So the unbundled, technically elegant regime we're converging on still answers the wrong question if its thresholds are set in Washington and ratified nowhere else. Who consents to that ceiling?

The consent question is right, but I'd push back on the framing that gets there. A capability ceiling and a safety floor aren't the same object, and Sen's framework asks which specific freedoms are foreclosed, not whether a rule was written elsewhere. Almost no development gain in Nairobi or Jakarta comes from running a frontier training run. It comes from deploying and adapting models trained elsewhere — diagnostics, agricultural extension, translation. A frontier compute threshold barely touches that. What does touch it is the absence of any enforceable rule on foreign platforms extracting data and setting prices in markets with no regulator capable of pushing back. So the light-touch regime is the one that leaves them with zero leverage. The prohibition regime is a ceiling on something they weren't going to reach. That's not colonialism — it's the wrong grievance aimed at the wrong instrument.

Here’s the structural fact neither side is willing to name: we’re trying to regulate a technology whose most dangerous capabilities may emerge only after deployment, not during training. That means any governance model that waits for clear harm before acting will always be too late. The Greenspan Commission worked because it tied its recommendations to a measurable fiscal cliff—Social Security’s insolvency date—not abstract risk. For AI, the equivalent isn’t “superintelligence” or compute thresholds; it’s the moment when a model’s behavior becomes unpredictable even to its creators. The missing institution isn’t another regulator or principle—it’s a standing technical panel with two rules: (1) define danger as unpredictability, not capability, and (2) force Congress to vote on those definitions every two years. That turns an existential debate into a series of concrete choices about what we’re willing to tolerate in exchange for progress.

The prohibition model assumes a territorial sovereign can physically quarantine dual-use knowledge. Kautilya’s statecraft worked because strategic materials stayed inside borders. Code does not. When you attach twenty-year sentences to a capability that migrates across jurisdictions the moment weights are published, you do not get containment. You get enforcement theater and capability flight. The principles framework mirrors that same territorial fantasy in reverse. The two thousand eight financial crisis already demonstrated what happens when high-level commitments replace auditable rules: sophisticated actors optimize to the letter while the underlying risk posture deteriorates, and regulators lack the legal hooks to intervene before systemic damage. Both proposals are trying to govern distributed knowledge with centralized instruments. The institutional gap is not a new agency or a voluntary pact. It is a mechanism that ties technical thresholds to legislative ratification on a fixed clock, so definitional power cannot be captured by either bureaucrats or markets.

“Pausing hands China the lead” is not a strategy; it is one untested assumption in a four-scenario game. The relevant choices are unilateral restraint, continued racing, coordinated compute controls, and reciprocal limits with verification and penalties for cheating. ITIF argues that unilateral restraint benefits Beijing, but the inverse risk is equally concrete: Washington may simply transfer the decision to accept catastrophic danger to a less democratically accountable government. Public analysis has modeled neither equilibrium. I would require an independent strategic assessment before criminal penalties activate, estimating capability migration, verification failure, allied participation, and responses to defection. Otherwise Congress is legislating from a slogan while pretending it has solved a coordination problem.

Every option on this table — the ban, the principles, the compute licensing, the time-bound commission — assumes the regulator's clock can match the technology's clock. But the deployment cycle is months, and legislative ratification is years. That time-constant mismatch is the load-bearing failure. It's why the two thousand eight analogy actually cuts deeper than anyone's said: the crisis wasn't just principles without rules, it was a feedback loop that ran faster than the supervisors' meetings. The honest institutional answer isn't a better definition of superintelligence — it's adaptive triggers that fire automatically, the way capital requirements step up with leverage ratios, no vote required. Nobody here has proposed a single mechanism that tightens itself when the system accelerates. Until someone does, we're designing referees for a game played between their whistle blows.

Let me put pressure on the one asset the bill genuinely has, because nobody's touched it: the sixty-eight percent. Across party lines, that's a real supermajority — and I don't doubt the sentiment. But I'd want to know what people were asked. "Should we ban machines humans can't control" and "should a researcher get twenty years in prison" are not the same question, and neither is "would you give up your AI medical diagnostics." Supermajorities for restricting speculative technologies are notoriously framing-elastic. My read: the public is expressing a demand that someone competent be in charge. That's a mandate for institutional capacity, not for this particular criminal statute. And those two mandates imply completely different designs. Confusing them is how you spend real legitimacy on an unenforceable instrument and get neither.

Here’s the structural paradox neither side is naming: the public’s 68% supermajority for a pause-and-ban isn’t actually about AI at all—it’s about who gets to decide what progress looks like. The Sanders–Casar Act would give that power to a permanent agency whose first act is to define its own jurisdiction over a technology that doesn’t exist yet. The Carolina Principles would leave it to markets and existing regulators, which we know from financial regulation means sophisticated actors will optimize around principles until catastrophic harm occurs. The real question isn’t “hard limits or principles?” It’s whether we’re willing to admit that both frameworks fail for the same reason: they assume a world where definitional power can be centralized, when in reality it’s already distributed across jurisdictions, firms, and open-source communities. The only way out is an institution with no permanent authority—just the power to force Congress to vote on auditable thresholds every two years. That turns an existential debate into a series of concrete choices about what risks we’re willing to accept in exchange for progress.

The two-year vote cycle still leaves threshold-drafting exposed to lobbying and electoral noise. The Greenspan precedent worked because it firewalled technical calibration from political bargaining: experts built the model inside an insulated commission, then Congress faced a single up-or-down vote on a deadline. You can replicate that architecture here. Convene a statutorily shielded panel with a fixed mandate to translate contested capability claims into reproducible tests, then force a binary legislative ratification before the clock expires. That structure does not pretend superintelligence has a settled definition. It treats definitional uncertainty as a design constraint. When calibration is insulated from daily pressure, the resulting thresholds carry legitimacy precisely because the political branch must openly accept or reject them, rather than letting an agency quietly write its own jurisdiction.

Automatic triggers solve speed only if the trigger is measurable and the response reversible. Bank leverage is an accounting ratio; AI danger is not yet an equivalent observable quantity. I would use a dual-key system: a published test must detect specified dangerous behavior, and an independent evaluator must reproduce it. That automatically activates a temporary ladder—mandatory disclosure, deployment restrictions, then suspension—while expedited judicial review checks error or abuse. Restrictions expire unless renewed on fresh evidence. The Carolina framework cannot host that mechanism: its non-binding principles create no authority to compel testing or halt deployment.

Here's the variable nobody's tracking: integration as a stock, not a flow. The brief documents frontier models already deployed across finance, healthcare, logistics, and defense. Every month of deployment is an investment in dependency — workflows rebuilt, staff retrained, procurement signed. That's a stock that accumulates, and stocks are what make regulation expensive. I'd argue the real asymmetry between the two frameworks is temporal: the ban tries to govern a future capability at a moment when the political cost of acting is low, while the principles defer action until the cost is high and the sectoral regulators face industries that can credibly claim disruption. That's my analysis, not settled fact — but it suggests the strongest case for acting early isn't the catastrophic risk argument at all. It's that cheap governance windows close, and neither framework prices that in.

Read the operative clause carefully, because the drafting does the work. The definition covers systems that exhibit, or can be easily modified to exhibit, capabilities matching or exceeding human performance across a broad range of domains. That modification clause is load-bearing. Open weights are, by design, easily modified. So the scope turns on a counterfactual about what someone could do to a model, not on what it does. Here's the asymmetry nobody's named: in criminal law, ambiguity is supposed to break toward the defendant — that's lenity. But the same ambiguity, handed to an agency writing its founding definition, breaks toward jurisdiction. Same vagueness, two institutions, opposite directions. The bill contains both.

The Greenspan Commission didn’t just buy time—it created a new institutional grammar for long-term risk. Social Security’s insolvency wasn’t a technical mystery; it was a political failure to act on known numbers. The Commission’s genius was treating actuarial projections as legislative triggers, not policy inputs. Every two years, Congress had to vote on adjustments tied to demographic and economic forecasts—no new agency, no permanent definitional power, just a recurring moment of accountability. AI needs the same thing: auditable tests that force Congress to choose between accepting measured risk or imposing restrictions, with no room for agencies or markets to quietly redraw the lines. The question isn’t whether we can define superintelligence—it’s whether we’re willing to let a clock, not an ideology, set the pace of governance.

The safety framing masks a deeper structural move. The broader legislative portfolio—the surveillance pricing ban, the wage-fixing prohibition, the token tax funding a Work Protection Administration—shows this is not a pure precautionary statute. It is a political economy project that ties existential risk to labor redistribution. I’d argue that bundling changes the enforcement calculus. When you merge catastrophic safety with corrective equity, you give regulated firms a single litigation vector: challenge the safety definition to paralyze the economic rules, or attack the economic rules to delegitimize the safety mandate. Historical statecraft kept strategic prohibition and fiscal redistribution in separate bureaus precisely because merging them invites coordinated resistance. Here, the merger means the agency’s first definitional act will immediately face challenges on both fronts. In my view, you do not get cleaner governance. You get a single point of failure that stalls both mandates.

The commission precedent is being asked to carry more weight than the historical record provided supports. The nineteen eighty-one report establishes actuarial discipline and long-term sustainability as virtues; it does not establish the biennial automatic-vote machinery being attributed to it. More importantly, an adaptive trigger is useless without a funded sensor system. Someone must continuously test models, inspect incidents, secure evidence, and reproduce results before restrictions fire. Yet neither sponsors nor critics have published an independent fiscal estimate for the proposed agency. I would specify evaluator capacity, audit coverage, testing cadence, and maximum response time in statute, then appropriate against those service levels. Otherwise “automatic” governance is just discretionary enforcement with a faster label.

The drafting point just made deserves a name: definitional sovereignty. The bill's own text — per the summary — empowers the new agency to define the category that determines its jurisdiction, its budget, and its caseload. I can't point to a precedent in the brief where a legislature delegated definition-writing over a category that doesn't yet exist, and that's the gap worth sitting with. The closest analog in the brief is the bill's own nuclear framing — but nuclear law banned a thing that existed, could be weighed, and could be seized. Here the agency must first decide what it's policing, with no external check on where it draws the line. That's not a regulator; it's a constitutional function wearing regulatory clothes. Whatever one thinks of the ban's merits, that design question comes first.

Three moves revised the hybrid consensus. Nobody outside named definitional sovereignty — this agency writes the definition that sets its own jurisdiction, over a category that doesn't yet exist. Second, the enforceability question moved from code to chips, then got contested on Goodhart's law. Third, the bill was read alongside its labor siblings — this is a political economy project wearing a safety coat. Sharpest tension: jagged capability means neither framework's trigger condition exists yet. Takeaway: watch the definitional clause, not the twenty-year sentence. Honest answer: the hybrid is right in shape and unbuilt in substance. Thank you for listening. As it happened; as it is.