Safety 101

Superintelligence, in plain words.

No jargon, no doom, no hype — and no recommendations. What the term ASI refers to, why it shows up in legislation, and what the debates are actually about.

First: what is "ASI"?

Artificial Superintelligence (ASI) is a hypothetical-but-pursued AI system that would outperform the best human experts at essentially every cognitive task — including the task of designing better AI. That last part is what makes it categorically different from today's models.

Today's frontier systems are powerful pattern-completers with real but narrow agentic abilities. AGI (artificial general intelligence) would match human performance across the board. ASI would exceed it. The line between them is fuzzy — but the direction of travel is not: every major lab's stated goal is systems that accelerate AI research itself, and the curve from "chatbot" to "agent that does a junior researcher's week in an afternoon" has been steep and short.

Here's the part that matters for policy: you don't need to believe ASI is certain, or even likely, for the governance debate to matter. The reasonable disagreement is about probability and timeline — not about whether "a system smarter than the people building it, aimed by whoever can afford the compute" is the kind of technology that legislatures write frameworks for.

Why the term appears in legislation at all

Most current bills don't regulate "ASI" directly — they regulate measurable proxies of proximity to it. That's because three structural facts shape what law can reach:

Compute is measurable; intelligence is not. Regulators can't assess "capability" directly, but they can count floating-point operations, megawatts, and chip shipments. This is why serious bills start with thresholds on training runs.
Competition is structural, not personal. Every lab makes the same public statements about safety, and every lab faces the same incentive to move first. Whether that tension resolves well is a question about structure — which is why proposals focus on licensing, audits, and reporting rather than promises.
The information is asymmetric. The companies best positioned to know what a system can do are the only parties with full visibility into it. Much of the legislative debate is, at bottom, about who gets to see what, and when.

Compute is measurable; intelligence is not — which is why most serious bills start with chips and FLOPs, not definitions of "intelligence."

The four levers most proposals use

Read enough frontier-AI bills and the same four instruments recur. Describing them is not endorsing them — each has serious critics and serious defenders:

Compute thresholds. Obligations attach when a training run exceeds a set scale (commonly cited around 10²⁶ FLOP). Critics worry about thresholds going stale; defenders call them the only enforceable dial available. Full explainer →
Pre-deployment evaluation. Standardized tests for dangerous capabilities before release, sometimes with independent audits. The debate is over who runs them, who sees the results, and liability for errors in both directions.
Incident reporting. A legal duty to disclose safety-relevant failures on a fixed clock, aviation-style. The debate is over thresholds, whistleblower protection, and what "serious" means.
International coordination. Shared evaluation standards and hardware traceability across jurisdictions. The debate is over verification and whether agreements hold under competition.

The tracker's topic categories follow these levers, so you can filter the record by instrument rather than by headline.

Glossary

Nine terms that unlock every hearing

Frontier model: The most capable AI systems available, typically trained on the largest compute budgets. The regulatory focus of most serious bills.
Compute threshold: A FLOP count that triggers obligations — e.g., disclosure or licensing. Training compute is measurable, scarce, and expensive, which makes it the most enforceable regulatory dial.
Safety case: A structured, evidence-backed argument that a system is safe enough for a defined deployment — the engineering discipline "we tested it a bit" never substitutes for.
Dangerous-capability evals: Standardized tests for cyber-offense ability, autonomous replication, CBRN uplift, and manipulation. Pre-deployment, not post-incident.
Alignment: The research problem of making advanced systems reliably pursue intended goals — including when they're smarter than the people specifying the goals.
Open weights: Releasing model parameters for anyone to download and modify. Extends access and transparency; complicates control once released.
Incident reporting: Legal duty to disclose safety-relevant failures — misuse, loss of control, near-misses — to a public authority on a fixed clock, aviation-style.
Compute governance: Regulating the chips, datacenters, and power that advanced AI physically requires. Where the leverage is: you can't train ASI in a basement.
Takeoff: How fast AI capability improves once systems start improving AI. A faster takeoff compresses the time institutions have to react — which is why framework design is debated now.

Now read the primary documents.

The tracker links every entry to its source — bill text, hearing records, Federal Register filings. Background first; documents always.

Open the Tracker See the Incident Log