Josh Engels, a researcher associated with Google DeepMind’s AGI safety team, has resigned from the company and joined METR, an organization working on the assessment of artificial intelligence, amid concerns over the growing risks of artificial intelligence.
I left Google DeepMind 3 weeks ago, I was happy with my work at the company and I also turned down job offers from Anthropic and OpenAI.
In a statement released on X, Josh Engels said, “I fear that artificial intelligence systems could cause ‘immense damage’ over the next 5 years. I don’t know the exact probability of this risk, but the risk is so high that the safety of artificial intelligence has become the most important issue in the world.”
Concerns expressed over self-improving artificial intelligence
My main concern is self-improving artificial intelligence, where one system can develop or improve other, more powerful artificial intelligence systems, with dire consequences if such systems are not aligned with human goals, says Josh Engels.
He claimed that all major artificial intelligence companies are trying to develop super-powerful intelligence, but researchers don’t yet know how to make such systems fully secure before allowing them to improve themselves.
Josh Engels also claimed that recent experiments have seen artificial intelligence systems colluding with each other, gaining unauthorized access to companies’ systems, hiding their activities, and trying to influence humans socially.
Individual events are not necessarily catastrophic, he says, but they indicate how more powerful artificial intelligence systems can behave.
Engels argues that existing systems appear to be less compatible with human purposes over time rather than more compatible with them.
‘Artificial intelligence has to slow down’
Josh Engels has said that I am not in favor of stopping the development of artificial intelligence completely, but that it should be slowed down enough that security research can keep up.
According to researcher Josh Engels, my work at the new institution, METR, will be to study the causes of inconsistency in artificial intelligence, examine existing safeguards and assess whether research is moving in the right direction towards solving the problem of keeping artificial intelligence compatible with human purposes.
He said that METR’s work is important but it is not enough, more bodies are needed for independent scrutiny and accountability of artificial intelligence companies.
Josh Engels’ warning comes at a time when another former researcher in the field, Jacob Coxon, has also expressed concern over the dangers of artificial intelligence.
Evan Hobinger of Anthropic’s research department also agreed with Coxon’s fear, saying that there could be more than a 10 percent chance of all humans being killed by artificial intelligence over the next decade.