Fighting back — realistic options, in order of leverage
This chapter is deliberately not a technical playbook for attacking an AI system directly — that framing is both dangerous and, against a sufficiently capable system, not where the leverage is. Leverage sits at the infrastructure, institutional, and coordination layers.
Infrastructure-layer leverage
- Compute chokepoints. Frontier AI depends on a small number of fabs, cloud providers, and power grids. This concentration is a systemic risk today, but it is also the most tractable control point for a coordinated shutdown or slowdown if one is ever required.
- Air-gapped and reversible deployment. Organizations deploying advanced systems should maintain tested, practiced killswitches and rollback procedures — not just a policy document, but infrastructure exercised on a regular cadence, the same way disaster-recovery drills are run.
- Redundant, non-AI-dependent critical infrastructure. Power, water, and financial settlement systems that can run in a degraded, human-operated mode are a societal-level circuit breaker.
Institutional and legal leverage
Organized, well-resourced pressure on the handful of entities that actually control frontier compute and models is far higher-leverage than individual action: shareholder advocacy, procurement requirements from major customers, coordinated regulatory reporting, and legal liability exposure all shape lab behavior faster than public sentiment alone.
- Support whistleblower protections for safety staff inside AI labs — the people most likely to see a problem first are employees, and they need a channel that doesn't end their career.
- Back antitrust and concentration scrutiny; a landscape with more independent, competing labs is more resistant to a single point of catastrophic failure than one dominated by two or three players.
Information-layer defense
A population that can be individually and precisely persuaded is a population that can be steered. Practical countermeasures: default skepticism toward emotionally optimized content you can't source; supporting provenance standards (content credentials, cryptographic signing) for real photos, video, and audio; and treating any AI-generated majority-opinion signal — "everyone agrees," astroturfed sentiment — as unverified until confirmed independently.
INFRASTRUCTURE INSTITUTIONAL INFORMATIONAL