Vishal Maini, a former communications and policy staffer at DeepMind, reveals that the lab strictly prohibited public discussion of AI extinction risks between 2018 and 2022. According to Maini, this gag order applied to everyone, regardless of seniority. Researchers were actively coached to dismiss existential threats as mere alarmism. They were told to compare these fears to Hollywood tropes like “Terminator” and steer conversations toward beneficial applications in healthcare or climate science instead.
The internal reality starkly contrasted with this polished public face. Maini notes that the team privately acknowledged the AI alignment problem remained unsolved. They also recognized that far too few resources were dedicated to solving it. After months of internal pushback, leadership finally relaxed the strict ban. The compromise allowed for safety content, provided it was framed positively. One notable example was a blog post that discussed extinction risks using friendly, accessible language rather than apocalyptic tones.
This disclosure highlights a significant gap between private technical concerns and public corporate messaging during a critical period of AI development. Maini argues that this disconnect is now shrinking simply because the evidence of potential risk has become harder to ignore. His account aligns with a broader trend of AI safety researchers breaking their silence. Many are now warning publicly about unsecured models and systems that could slip beyond human control.
Even researchers within DeepMind are joining this chorus of caution. The shift suggests that early suppression of safety debates may have delayed necessary public scrutiny. It also raises questions about how other major labs manage similar internal dissent. As models grow more capable, the pressure to maintain a purely optimistic narrative becomes increasingly difficult to sustain. The industry can no longer afford to treat existential risk as a cinematic plot device.
