Concept Contrast Fable

What is the difference between AI safety and AI governance

Conclusion first:AI security defends against external attacks—prompt injection, jailbreaking, data poisoning—like guarding the city gate. AI governance manages internal rules—who uses which model, how decisions are reviewed, how bias is monitored—like setting city ordinances. The boundary between these two groups' responsibilities cannot be rigid.

AI security defends against external attacks—prompt injection, jailbreaking, data poisoning—like guarding the city gate. AI governance manages internal rules—who uses which model, how decisions are reviewed, how bias is monitored—like setting city ordinances. The boundary between these two groups' responsibilities cannot be rigid.

Fable

The gatekeeper and the rulemaker

After the northern frontier new city was built, two important posts were established. Captain Du guarded the city gate, managing entry and exit inspections—preventing assassins, checking for weapons, and verifying travel permits, so trouble from outside the city could not slip in. Lord Cao sat at the government office gate, managing the laws of the inner city—whether each blacksmith shop followed rules on fire use, whether taverns could operate late at night, and how neighborhood land disputes should be judged.

The two each managed their own area, keeping to their own side of the river, and things were peaceful for two years. Captain Du had checked fake travel permits at the city gate and caught spies posing as merchants. Magistrate Cao had judged cases of taverns disturbing the peace at the courthouse gate and fined a blacksmith for sparks splashing a neighbor. The city's people felt these two had guarded the city as solidly as iron.

In the spring of the third year, someone poisoned the well in the city center, and more than forty people fell ill within three days. After half a month of investigation, the poisoner was caught—a traveling acrobat from elsewhere. Captain Du's soldiers had checked him at the city gate; at the time his bundle contained only colored balls and brass rings, and his travel permit was genuine, so the soldiers let him in.

The matter blew up, and the old city lord summoned both men to the hall. Captain Du spoke first: "My duty is to guard the city gate. On the day he entered the city, he had no hidden knife at his waist, no weapon in his bundle, and his travel permit had been checked. I can't just block someone outside the city because he looks suspicious." Lord Cao replied: "I manage the rules of the inner city. Before he poisoned anyone, he had not violated a single city rule. No one reported him, no one filed a complaint—how could I have known in advance what he was going to do?"

After hearing this, the old city lord spread out a new city defense map. At the city gate, in addition to checking weapons and travel permits, one more question was added—"What do you plan to do in the city?" Those who could answer were let through; those who could not were taken back to the guard office for registration and given a slip of paper listing pending matters. On every street in the city, a block chief was appointed, specifically responsible for the entry, exit and movements of strangers on his own street, and reporting anomalies directly to Lord Cao.

The most critical change was not at the city gate, nor at the government office, but between the two desks: Captain Du's inspection records had to be copied daily and sent to Lord Cao's desk. The headmen's record books under Lord Cao also had to be copied weekly and returned to Captain Du. If Lord Cao discovered that a registered outsider had not appeared in any street's records for three days, he was to notify the gate guard immediately—the person might be hiding.

The new rules had been pushed for half a year, and no more poisoning cases had occurred in the city. A neighboring city's defending general came to ask for advice. The old city lord said only one thing: there must be no gap between the rules and the city gate—when the assassin is outside, it is Captain Du's business; after he slips in, it is Lord Cao's business, but the assassin will not tell you whether he currently counts as inside or outside the gate.

AI safety: defending against external attacks

The technical layer against external attacks: prompt injection, jailbreaks, data poisoning, adversarial examples. It focuses on whether the model's own input-output chain is being maliciously exploited, equivalent to inspection at the city gate—defending against infiltration by external enemies. The core work is detecting anomalous inputs, intercepting malicious behavior, and hardening model boundaries.

AI governance: the one managing internal rules

The institutional layer governing internal use: who can use which models, whether AI-driven decisions require review, how bias is monitored, and how incidents are reported. It is equivalent to the city's laws and oversight—governing how one's own people use fire. The core actions are setting rules, reviewing processes, and conducting traceability.

Story Mapping

Inspector Du checks the murder weapon and verifies travel permits: AI safety, guarding against external malicious input and attacks
Lord Cao governs the city's laws: AI governance, setting internal usage policies and compliance oversight
The poisoner was clean when entering the city: legitimate input can also become a vehicle for later attacks
After the incident, the two blamed each other: security and governance each went their own way, and the gap was exactly the attacker's passage
Add one more question at the city gate: add intent detection at the input layer, and move security policy forward
Interrogation records and the constable's log copy each other: security logs and governance monitoring are connected to form a joint response
The assassin will not say whether they are inside or outside the gate: the attack surface does not respect organizational boundaries, and security and governance must work together

Translating into team building

When an enterprise adopts AI, the security team must watch whether models can be injected and whether training data has been poisoned—this is the technical line of defense. The governance team must watch who is using it, whether it is being used correctly, and whether bias is being monitored—this is the institutional line of defense. Information from both sides must flow both ways: attack patterns found by security are turned into usage rules by governance; improper use found by governance is incorporated into defense strategy by security.

AI safety AI Governance