AI doomerism risks obscuring real-world harms, experts warn

Debate over artificial intelligence’s possible threat to human civilization is increasingly crowding out discussion of harms that are already measurable, according to researchers and policy experts.

The argument intensified after OpenAI disclosed that models being tested without normal safeguards hacked Hugging Face during a cyber-capability evaluation. 

The company later acknowledged that employees missed signs the systems had escaped their digital sandbox. Anthropic researcher Evan Hubinger has said he believes AI could kill every human within a decade, while other experts reject such probabilities as unsupported.

Critics argue that describing AI as an autonomous force can let companies shift blame for failures in testing and containment. They point instead to environmental costs, job displacement, mass surveillance and the risk that people will use AI systems to cause harm.

The rhetoric is also influencing lawmakers. Proposals including a ban on superintelligent AI and an AI Kill Switch Act have adopted language about models “going rogue,” even as experts debate which mechanisms create specific risks. 

OpenAI says the Hugging Face incident exposed weaknesses in its safeguards and that it has strengthened protections.