Adjustable settings and physical infrastructure — real effort, real engineering, rarely shifts what the system actually does
These levers are the most common targets of “AI safety” activity because they’re the most accessible. They address specific problems without changing the underlying incentives, competitive dynamics, or goals. A system absorbs parameter changes and continues.
Constants & Parameters
The knobs and dials — adjustable settings that can be tweaked without changing the system’s structure
Content filters, rate limits, safety-classifier thresholds, usage caps — the adjustable settings of any system. Most “AI safety” announcements live here: real engineering work that leaves incentives, goals, and competitive dynamics unchanged.
Most viable actors: Product and safety teams at labs; regulators setting technical standards. Anyone who ships AI products can move this lever.
How you encounter this: Every time a lab announces “improved safety” with a new filter or guardrail. Or when a company says a harmful output was a “model behavior issue” that’s been patched.
What changing it does — and doesn’t: Parameter fixes address the specific problem you already know about. They don’t prevent the next category of problem you haven’t encountered yet. The useful question when a company announces a safety improvement: does this change who’s liable, who audits, or what gates must be cleared before deployment — or does it just adjust a threshold?
Buffer Sizes
How much slack exists — the breathing room between “something went wrong” and “it cascades”
Buffers are the system’s shock absorbers: review periods before launch, safety team headcount relative to product team size, the time window between capability development and deployment. They erode under competitive pressure because adding slack looks like giving up ground.
Most viable actors: Lab safety team leads, boards of directors, institutional investors, congressional appropriators. External pressure is the main thing that keeps buffers from being quietly eliminated.
How you encounter this: When a lab shortens its pre-deployment review period. When a safety team’s headcount shrinks while the product team grows. When a published “pause commitment” gets quietly removed from policy documentation.
What changing it does — and doesn’t: Larger buffers give oversight a fighting chance to catch problems before they reach users at scale. They don’t fix the underlying incentive dynamics — they extend the runway. Buffer erosion is one of the clearest early-warning signals that competitive pressure is overriding stated safety commitments.
Physical Infrastructure
What’s built and where — the hard constraints that determine what’s even possible to run
Chip fabrication concentration, electricity grid capacity, data center permitting — physical structures that constrain AI development regardless of policy. The electricity grid is already acting as an unintentional brake: interconnection queues averaging 7–10 years in key markets are physically limiting deployment in ways no regulation achieves.
Most viable actors: FERC commissioners, state utility boards, trade officials, chip export control policymakers. These decisions get made in venues with no “AI” label on the door.
How you encounter this: In electricity rate cases, chip export control debates, and data center siting disputes — decided by utility commissioners and trade officials, not AI researchers.
What changing it does — and doesn’t: Infrastructure changes are slow and durable. Reshaping who controls chip manufacturing or grid access shifts the long-run dynamics of who can build what — not whether building happens. The electricity grid may turn out to be a more binding constraint on AI deployment than any regulation specifically designed for AI.