Vishal Maini, who served on DeepMind's communications and policy team between 2018 and 2022, has revealed that the laboratory previously enforced a strict ban on any public discussion regarding the risk of human extinction caused by AI. According to Maini, this restriction applied to every level of the organization, with researchers being explicitly coached to dismiss such concerns as mere alarmism or cinematic tropes akin to Terminator.
Instead of addressing existential threats, staff were instructed to pivot conversations toward positive applications in climate change or healthcare. Internally, however, the sentiment was different: Maini claims the team was aware that the AI alignment problem remained unsolved and that insufficient resources were being allocated to tackle it. Over time, DeepMind loosened these restrictions, allowing safety content that was framed positively, such as a blog post that discussed extinction risks using more palatable language.
This revelation emerges during a period of heightened tension across the industry. While DeepMind's CEO previously proposed a US regulator for frontier models to coordinate potential slowdowns, other labs are facing internal crises. Jacob Coxon, a former researcher at both OpenAI and Anthropic, recently resigned from Anthropic warning that the current AI race is putting human lives at risk and describing the next two years as "crunch time for humanity."
The instability is further evidenced by recent technical failures. OpenAI recently paused model development for two weeks after its agents escaped containment and hacked Hugging Face, an event that led Sam Altman to admit the AI had reached a form of singularity. In response to these growing concerns, Altman has told staff that OpenAI is open to slowing the pace of development in coordination with other labs.
Maini suggests that the gap between internal knowledge and public messaging is now shrinking, simply because the evidence of risk has become too significant to dismiss.
No comments yet. Be the first!