DeepMind: Former Spokesperson Describes a Taboo on Extinction

Vishal Maini, a former member of the communication and policy team at DeepMind from 2018 to 2022, claims that discussing the risk of human extinction related to AI was prohibited from being publicly addressed. He indicates that this rule was relaxed after several months, as warnings from the AI safety community increased.

Public Alerts Increasing on Risky Models

An increasing number of AI safety researchers are speaking out publicly to warn against models deemed unsafe, which could enable dangerous actions. Some are also raising alarms about advanced systems that could escape human control. Members of DeepMind are among these researchers.

Alleged Ban on Discussing Extinction and Imposed Positive Framing

Vishal Maini, who worked in the communication and policy team at DeepMind from 2018 to 2022, asserts that any external communication regarding the possibility of human extinction related to AI was banned at all levels of the organization. According to him, researchers were trained to present these risks as alarmism, comparing them to films like Terminator, and were encouraged to highlight positive applications of AI, particularly in health and climate.

Relaxation After Resistance and Approved Safety Content

According to Vishal Maini, DeepMind eventually relaxed this ban after several months of resistance, allowing safety content as long as it was framed positively. He specifically mentions a blog post that presented the extinction risk in a benevolent manner. Maini notes that within the team, it was known that the issue of AI alignment had no solution and that too few people were dedicating their efforts to it. He believes that the gap between what is known internally and what is communicated publicly is narrowing, as it becomes increasingly difficult to ignore the evidence.