Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
AI earns trust in critical systems through control¶
Google DeepMind and its research partners report on September 4 that an artificial intelligence (AI) controller has undergone testing at the Laser Interferometer Gravitational-Wave Observatory (LIGO) in Livingston, Louisiana. Their Deep Loop Shaping method reduces control noise in a difficult mirror-control loop by a reported factor of 30 to 100.
The achievement is technically specific, and that is precisely why it offers a useful lesson for AI in critical systems.
LIGO measures distortions smaller than most of us can reasonably imagine. Distant waves, traffic, and other vibrations can disturb its mirrors. The observatory depends on thousands of feedback controls to maintain alignment without adding noise that obscures the signal it is built to detect.
This is not a setting where “generally helpful” is a meaningful requirement. Performance must be defined against physics.
Trust comes from a bounded job¶
Deep Loop Shaping uses reinforcement learning (RL), first in simulation and then on hardware, to suppress noise in a particular control problem. The controller does not run the observatory or decide what a detected event means. It has a defined function, measurable constraints, and a surrounding control architecture.
That boundedness is not a limitation to apologize for. It is a design advantage. Teams could specify acceptable behavior, observe the consequences, compare performance, and retain the rest of the system's safeguards.
For critical applications, confidence grows when the role of AI can be stated with similar precision: what variable it influences, over what range, with which inputs, under which conditions, and with what fallback.
Simulation is evidence, not permission¶
The team first trains and evaluates the controller in a simulated environment, then tests it on the physical system. The result reportedly carries over to hardware, but that transfer requires empirical demonstration.
This distinction matters across defense, aviation, energy, health, and industrial control. A simulation encodes assumptions about disturbances, sensors, timing, equipment, and users. Good performance demonstrates what happened inside those assumptions. It does not prove the assumptions complete.
The evidence chain should therefore preserve model versions, simulator conditions, test configurations, hardware results, observed limits, and operator actions. When the environment changes, the team needs to know which evidence still applies.
Human-centered control is more than a stop button¶
In a consequential feedback system, human oversight includes the ability to understand system state, recognize a departure from the operating envelope, intervene at a useful time, and restore a known condition. A nominal override that arrives after the process has become unrecoverable is not meaningful control.
Nancy Leveson's systems-theoretic approach to safety treats accidents as control problems shaped by interactions, inadequate constraints, and flawed feedback. That view fits AI-enabled critical systems better than asking whether the model is safe in isolation.
The LIGO work points toward a disciplined pattern: choose a consequential but bounded function, connect performance to physical measures, move deliberately from simulation to hardware, keep the broader control structure visible, and preserve evidence for continued operation.
AI does not earn trust in critical systems through eloquence. It earns trust by remaining controllable when the world pushes back.
Sources and research trail¶
- Google DeepMind, “Using AI to Perceive the Universe in Greater Depth” (September 4, 2025).
- Tracey et al., “Improving Cosmological Reach of LIGO Using Deep Loop Shaping” (2025).
- Leveson, Engineering a Safer World (2011).
- Sargent, “Verification and Validation of Simulation Models” (2013).