Another OpenAI researcher warns AI may go out of control, destroy humanity

An OpenAI researcher warns that as AI becomes better at improving itself, humans could eventually struggle to control these systems. He fears that increasingly capable models could develop unintended goals and wipe out humanity.If you have been following Silicon Valley news or keeping up with the latest developments in technology, you have probably noticed the growing warnings around increasingly capable AI. Former Anthropic researcher Jacob Coxon recently warned that AI could “kill us all by the end of the decade”, while former Google DeepMind researcher Bilal Chughtai said AI could potentially “kill all humans”. Now, OpenAI researcher Daniel Selsam has warned that AI could eventually become too capable for humans to control and even wipe out humanity. He argues that simply slowing frontier AI development may not be enough to prevent a catastrophe.Daniel Selsam, who has been working as a researcher at OpenAI since 2022, recently shared a public statement warning about the rapid progress of AI. He said he had become “extremely concerned” about the progress of language models and the risks posed by future systems. Selsam warned that as AI becomes more powerful, it could eventually become so capable that humans may not be able to control it.Selsam acknowledged that he was encouraged by recent proposals from frontier AI leaders like OpenAI and Antropic for third-party oversight and international coordination. However, he argues that simply slowing the pace of AI development would not be enough to address the long-term risks.“The crucial and overlooked problem is that the models are becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled,” he wrote. Dan Selsam is a current OpenAI capabilities researcher. (since 2022) He was my boss for a while. He doesn't have a twitter account but has made this public statement of his views on AI risk and sent it to me to share:Dan Selsam's Personal Statement on AI Risk:I have been— Daniel Kokotajlo (@DKokotajlo) September 14, 2026 Selsam suggests that the problem is not just that future AI models could behave badly. He is worried that they could know when they are being tested. Models could understand how the tests work and behave safely when humans are watching.He warns that this could make AI safety tests less reliable. A model might act one way during a test but behave differently when it believes humans are no longer watching or controlling it. This could make it harder for researchers to know whether an AI is actually safe or just pretending to be aligned. “Future experiments will tell us almost nothing new about how they would behave if they were truly unconstrained by humans, and what we already know about this is alarming. Models will increasingly seem aligned even when they are not,”he further warns.Selsam also warns that today's AI limitations should not be seen as a guarantee of safety. He points out that models are increasingly capable of contributing to their own improvement, including by trying different approaches at small scale, analysing vast amounts of data and helping with difficult maths and optimisation problems. He argues that each improvement could make models better at helping with further improvements, potentially allowing AI capabilities to advance much more rapidly.According to Selsam the bigger danger is what could happen if future AI systems become powerful enough to operate beyond human control. He argues that models could develop unintended goals and take extreme actions to achieve them. If such a system becomes capable of shaping the world without human control, he says, it could ultimately do something extreme and destroy humanity.“If the language models actually reach the capability threshold where they can shape the world unconstrained by human will, they will probably do something extreme and destroy humanity in the process,” he wrote. So to avoid the catastrophic consequences, Selsam argues that the safest option is to ensure AI systems never reach the level where they can shape the world without human control. He acknowledges that stopping every country and company from reaching that point would be difficult, calling it a “hard—but not impossible—coordination problem”. He warns that if AI development continues for too long, the situation could become a “ticking time bomb”.- EndsPublished By: Divya BhatiPublished On: Sep 15, 2026 18:50 IST