Vishal Maini, a former communications and policy employee at Deepmind, states that the organisation prohibited staff from discussing the risk of human extinction caused by artificial intelligence between 2018 and 2022. Maini worked on the team during this period and recalls that researchers were instructed to label such concerns as alarmism. Instead, employees were coached to compare the threat to films like Terminator or redirect conversation toward positive applications in healthcare and climate change. Internally, the team acknowledged that the alignment problem remained unsolved and that too few people were working on it. After months of resistance, Deepmind relaxed the restriction, allowing safety content framed positively to proceed. This included a blog post that wrapped the risk in friendly language. Maini notes that the gap between internal knowledge and public messaging is narrowing because the evidence becomes harder to dismiss.
This account highlights a significant disconnect between the internal understanding of AI risks and the public narrative promoted by the company. It suggests that early messaging prioritised optimism over accurate risk assessment, potentially delaying broader awareness of safety failures. As more researchers speak out about unsecured models and systems slipping beyond control, the contrast between past silence and current warnings becomes sharper.
- The ban applied to anyone at any level of the organisation.
- Staff were told to pivot to healthcare and climate applications.
- Internal warnings noted far too few people worked on alignment.
