# Plumbing for Brains in Jars: How AI Regulation Actually Works

> AI regulation polices the silicon, the outputs and the liability, not the model.

- Published: 2026-11-14
- Reading time: 6 min
- Source: original
- Canonical: https://reverb256.ca/posts/plumbing-for-brains-in-jars/
- Markdown: https://reverb256.ca/posts/plumbing-for-brains-in-jars.md

---

For years, the public debate over artificial intelligence governance sounded like a midnight philosophy seminar colliding with a disaster movie. Commentators argued over whether we could legislate "intelligence," ban rogue thoughts, or teach silicon to love humanity before it learned to play chess on a planetary scale. Meanwhile, the engineers and lawyers tasked with writing rules that actually survive contact with reality quietly arrived at a much more practical realization: you cannot regulate a digital mind in the abstract, because once a model's weights exist, they can be copy-pasted across fiber-optic cables faster than any legislator can draft a preamble. So instead of trying to read the brain's mind, modern regulation polices the **jar**—the industrial silicon, electrical grids, API airlocks, and corporate balance sheets keeping that brain alive.

Seen through that lens, concrete AI regulation stops looking like sci-fi handwringing and starts looking like familiar, cheerful industrial plumbing. Society has spent centuries figuring out how to live happily alongside volatile, brilliant inventions—from high-voltage wiring to commercial aviation—not by outlawing physics, but by installing circuit breakers, pressure gauges, and clear rules about who pays when a boiler leaks. Today's global governance stack applies that exact playbook across five physical and legal chokepoints.

## Metering the Vat

The first chokepoint is the industrial vat required to grow a frontier brain in the first place. While finished model weights are weightless and infinitely reproducible, training a frontier system demands staggering concentrations of specialized GPUs, gigawatts of power, and high-speed interconnects. Laws like the European Union's AI Act and California's SB 53 seize on this physical bottleneck by setting **compute thresholds**—at 10²⁵ and 10²⁶ floating-point operations, respectively—above which a model is legally presumed to carry systemic risk, triggering mandatory red-teaming, cybersecurity hardening, and incident reporting. Backing up those thresholds are proposals for **Cloud "Know Your Customer" rules**, requiring hyperscalers to verify who is renting massive clusters, alongside hardware-level telemetry that embeds tamper-evident odometers directly into datacenter chips. There is, of course, a delicious cat-and-mouse comedy built into this layer: every few months, algorithmic breakthroughs and ternary or low-bit quantization squeeze frontier-grade cleverness into smaller and smaller jars, forcing regulators to chase a moving target as efficiency turns yesterday's supercomputer feat into tomorrow's workstation hobby.

## Which Rooms It Enters

Once a brain is grown, the second layer of governance simply decides **which rooms the jar is allowed to enter**. Rather than regulating a spam filter with the same heavy hand as an intensive-care diagnostic tool, risk-tiered frameworks draw bright lines around context. Certain rooms are locked altogether: social scoring, workplace emotion-reading, and untargeted facial scraping are banned outright under European rules and targeted by anti-manipulation statutes like Texas's TRAIGA. High-stakes rooms—hiring, credit underwriting, housing, and medical triage—remain open, provided the jar arrives wearing a safety inspection sticker. Pre-market conformity assessments, Algorithmic Impact Assessments in the public sector, and anti-discrimination audits in jurisdictions like Illinois and Colorado require builders to prove their training data was not poisoned, install black-box event recorders, and wire up a prominent red human-override switch before the system ever touches a live decision.

## Dyeing the Spout

Because a paragraph, voice clip, or photograph emitted by a synthetic brain looks indistinguishable from human craft, the third chokepoint **dyes whatever comes out of the spout**. Regulations like EU AI Act Article 50 and California's AI Transparency Act pair cryptographic provenance metadata—tamper-evident digital wax seals such as C2PA Content Credentials—with statistical watermarks that subtly bias token or pixel selection, giving machine outputs a mathematically detectable accent. Alongside those output markers, laws like California's AB 2013 slap a standardized nutrition label onto the side of the glass, requiring developers to publish structured summaries of their training sources and copyright compliance. It is an elegant transparency mechanism, even if open-weight ecosystems and basic paraphrasing cheerfully remind policymakers that watermarks work best on cooperative commercial spouts rather than unsealed jars running in someone's basement.

## The Jar-Maker Pays

Yet the sharpest enforcement tool in the entire regulatory arsenal requires zero new AI agencies, zero FLOP counters, and zero inspectors peering into neural weights. It simply **rewires tort law so that if the brain bites someone, the jar-maker pays**. Under the revised EU Product Liability Directive, standalone software and AI systems are explicitly classified as products subject to strict, no-fault liability, barring vendors from hiding behind clever terms-of-service disclaimers when a defective system causes harm. Better still, because an ordinary citizen cannot dissect billions of proprietary parameters to prove *why* an algorithm misfired, courts can shift the burden of proof—presuming a defect once plausible harm is shown unless the vendor produces internal audit logs proving the system behaved safely. Even without specialized AI statutes, tribunals are already enforcing this common-sense rule: as Air Canada discovered in *Moffatt v. Air Canada* when its customer-support chatbot invented a nonexistent bereavement fare policy, a company owns every word its digital brain speaks.

## Deterministic Runtime Guardrails

The final layer addresses what happens when we give these brains mechanical hands—API keys, browser sessions, payment rails, and autonomous agentic loops. Once an agent can execute database writes or send emails on its own, reviewing an audit log next Tuesday is cold comfort. Enterprise governance and public-sector standards, including Canada's federal guide for agentic AI and ISO/IEC 42001 frameworks, therefore bolt **deterministic runtime guardrails** onto the jar's output pipe. Non-AI policy proxies sit inline between the model and the outside world like an incorruptible bouncer, scrubbing sensitive data, verifying cryptographic agent IDs, blocking unauthorized tool calls in milliseconds, and keeping an external kill switch wired directly to the power supply.

Nowhere is the practical reality of this multi-layered stack clearer than in the split screen between **Canada** and the **United States**. Canada initially reached for an omnibus European-style statute with its Artificial Intelligence and Data Act, only to watch the bill lapse when Parliament prorogued in early 2025. Rather than rebuilding a monolithic AI bureaucracy from scratch, Ottawa pivoted in 2026 toward surgical precision: weaving automated-decision explanations and human-review rights into consumer privacy reform under Bill C-36, pulling conversational AI services under online duty-of-care rules in Bill C-34, and letting sectoral heavyweights do the heavy lifting—from banking regulator OSFI's strict model-risk rules under Guideline E-23 to Quebec's Law 25, which quietly serves as the de facto private-sector sheriff across the country. South of the border, Washington took the opposite tack: Executive Order 14409 barred federal agencies from imposing pre-release licensing or permitting hurdles on frontier models while directing prosecutors to crack down on criminal agent misuse, leaving an energetic patchwork of state laboratories—led by California, New York, Colorado, Illinois, and Texas—to enforce incident reporting, watermarking, and civil-rights guardrails on the ground.

Strip away the apocalyptic rhetoric, and a surprisingly optimistic picture emerges. We do not need to solve the metaphysics of machine consciousness to govern artificial intelligence well. By metering the industrial vats, labeling the outputs, bolting deterministic airlocks onto autonomous agents, and keeping tort liability squarely on the shoulders of the builders, democratic legal systems are doing what they do best: turning a wild, dazzling frontier technology into safe, accountable infrastructure—so the rest of us can get back to building cool things with it.

Related: [Open Weights Are Data](/posts/open-weights-are-data/) · [Lawful Access Is a Backdoor](/posts/lawful-access-is-a-backdoor/) · [The Ideological Sandbox](/posts/the-ideological-sandbox/) · [Running a Company of Agents](/posts/running-a-company-of-agents/) · [The Configuration Grows](/posts/canada-config-grows/) · [State as Operating System](/posts/state-as-operating-system/) · [The Great Acid](/posts/the-great-acid/)
