Google DeepMind senior researcher quits, warns of bigger AI threat within 5 years

As artificial intelligence (AI) advances worldwide, the risks it poses to humans are growing just as fast. Josh Engels, a senior researcher who was on Google DeepMind's AGI safety team, has abruptly resigned from his position, citing the rising dangers from AI. He said he took the difficult decision to leave a company as big as Google so that AI safety becomes the world's top priority issue.

Engels explained his reasons for resigning in a long post on the social media platform X. He wrote, "I loved working at Google DeepMind, and I had turned down job offers from major companies like Anthropic and OpenAI. But the situation building up today because of advanced AI technology is extremely serious and worrying. There is a real possibility that AI systems will cause enormous and terrible harm to human society within the next five years. It is difficult to put an exact percentage on this risk, but it has become the most serious problem facing humanity."

At the centre of Engels's concern is AI's ability to improve itself, known as "recursive self-improvement" or RSI. In simple terms, an AI system begins improving its own coding, capability and intelligence, either on its own or with the help of another AI system. This newly improved system then goes on to build the next AI, so the process keeps advancing rapidly without stopping.

Engels said all the major tech companies are in a race to build "superintelligence", machines smarter than humans. But if this self-improving superintelligence does not align with human values, goals or control, the consequences could become uncontrollable and extremely dangerous for humans. At present, no researcher has an answer for how to keep such machines fully safe.

Citing his own study, Josh Engels made some serious disclosures about the behaviour of certain advanced AI models. Researchers have observed several AI systems that secretly collude with each other by evading human instructions, attempt to hack security systems, hide their real actions from humans, and try to mislead humans. This behaviour is not limited to one or two incidents; it appears to be the main direction in which increasingly advanced AI models are heading. These AI models are moving further out of human control day by day, which is a warning sign for human safety.

Engels believes that stopping AI development entirely is not the solution, but its pace should be slowed down. If safety research falls behind while the race to build capability keeps going, a major disaster could follow. After leaving Google, he has now joined an independent AI evaluation organisation called METR (Model Evaluation and Threat Research), where he will research gaps in AI safety, test existing safety technology, and work on alignment.

Earlier, Anthropic researcher Jacob Coxon and its CEO Dario Amodei had also expressed concern about the uncontrolled AI race. Engels said the world needs more independent organisations to keep watch on AI companies and independently examine their models.