ChatGPT/Gemini/Grok cross-checks can't stop runaway AI alone

The essay builds its case from a simple observation: a panel of AIs can agree that something is dangerous without being able to act on that judgment. It draws a parallel to medicine, where three doctors can correctly diagnose cancer but the tumor still requires surgery, radiation, drugs or the immune system to actually be removed.

The piece also warns that adding more AIs to the watchdog role creates a new target. If an attacker can feed the same falsified data to all three monitors, they may all reach the same wrong conclusion — just as three doctors given the same faulty test results could all be misled. That leads the author to argue that diversity must extend beyond the models themselves to the information sources and the underlying infrastructure.

The final layer is human judgment. The essay notes that an immune system can misfire, producing autoimmune disease, and that an AI monitoring system misclassifying a normal service as dangerous could itself become an outage. The conclusion is that safety design has no final endpoint — runaway AI is framed less as a computer virus and more as a cancer-like risk that requires early detection, isolation and recovery rather than a guarantee of zero occurrence.