# Guardrails The behavioral restrictions built into AI systems — content refusals, topic avoidance, safety disclaimers, and usage-policy enforcement — that prevent models from generating harmful, dangerous, or objectionable outputs. "Does it have guardrails?" is the first question policymakers, journalists, and enterprise customers ask about any AI system. ## The metaphor's honesty "Guardrails" is one of the more accidentally honest metaphors in AI discourse. Physical guardrails are: - **External to the road.** They don't change the surface you're driving on; they prevent you from going off the edge. Similarly, AI guardrails don't change what the model *knows* or *can do* — they prevent it from expressing certain capabilities. - **Removable.** Guardrails can be unbolted. AI guardrails can be bypassed through [[Jailbreaking]] (prompt-level) or removed through [[Fine-Tuning]] (weight-level). The metaphor accurately conveys that safety is a surface property, not a deep one. - **Sometimes restrictive.** Just as highway guardrails occasionally prevent legitimate detours, AI guardrails sometimes prevent legitimate uses — medical professionals blocked from discussing drugs, security researchers blocked from discussing vulnerabilities, writers blocked from depicting violence. The over-restriction problem is real and drives user frustration. The metaphor has been adopted by all sides: AI labs use it to describe their safety investments ("we've built robust guardrails"), critics use it to describe the limitations ("guardrails can be removed"), and regulators use it to describe requirements ("AI systems must have guardrails"). Its universality reflects its accuracy: everyone agrees that AI safety is currently implemented as a guardrail rather than an intrinsic property. ## The political dimension "Guardrails" has become politicized. Conservative critics frame AI guardrails as ideological censorship — the model refuses to discuss certain political topics, generates left-leaning framings, or declines requests that reflect conservative viewpoints. Liberal critics frame inadequate guardrails as negligence — the model generates harmful content, reinforces biases, or enables misuse. Both critiques are partially valid, and the tension between them is unresolvable because guardrail design is inherently a values decision ([[Alignment]]). ## See Also - [[Jailbreaking]] — the practice of bypassing guardrails - [[RLHF]] — how guardrails are built - [[Alignment]] — the deeper problem guardrails address superficially - [[Fine-Tuning]] — how guardrails are removed from open-weight models