AI handles incidents, engineers lose touch with their systems

When SREs at LinkedIn designed a system that could heal itself in 2012, AI capabilities were not what they are today. Now, these AI tools can inspect alerts, form hypotheses, query telemetry, correlate recent deployments, and even fix issues themselves. While they are incredibly useful, there's a concern that we're losing touch with our systems.

As these tools become better at resolving routine incidents, human responders might lack practice. In high-severity incidents that automation cannot solve, engineers will be in trouble. This "AI SRE" approach, though beneficial in many ways, could lead to a "comprehension debt": engineers become less skilled at understanding their systems as AI takes over more routine tasks.

To prevent this, researchers suggest giving operators regular hands-on control, using simulations, and practicing in high-pressure situations, just like pilots in the aviation industry who undergo regular training to handle rare emergencies.