Why handing over total control to AI agents would be a huge mistake

And when the systems can control multiple information sources at one time, the capabilities of the damage explode. For example, an agent with access to both private communications and public platforms can share personal information on social media. This information may not be correct, but it will fly under the radar of traditional fact examination mechanisms and can be amplified with more participation to cause heavy damage. We imagine that “I was not – it was my agent !!” It will soon be a common refusal to exempt bad results.

Keep the human in the episode

A historical precedent shows the reason that preserving human oversight is very important. In 1980, computer systems falsely indicated that more than 2000 Soviet missiles were heading towards North America. This error sparked the emergency procedures that seriously brought us from the disaster. What is the disaster that has been avoided is human motivation between different warning systems. If the decisions are fully authorized for the self -government systems that determine the priority of speed on certainty, the result may be disastrous.

Some will face that the benefits deserve risks, but we say that the awareness of these benefits does not require the surrender of complete human control. Instead, the development of artificial intelligence agents should occur as well as the development of guaranteed human control in a way that limits the scope of what artificial intelligence factors can do.

The open source agent systems are one of the ways to treat risk, because these systems allow more human supervision of systems that you cannot do. In Huging Face, we develop Smolgents, a frame that provides safe sand environments, and allows developers to build agents with transparency in their core so that any independent group can check if there is suitable human control.

This approach stands in a flagrant contradiction with the prevailing trend towards increasingly complex artificial intelligence systems that block decision -making processes behind layers of royal technology, making it impossible to guarantee safety.

While moving to developing increasingly advanced artificial intelligence agents, we must realize that the most important feature of any technology does not increase efficiency but enhancing human welfare.

This means creating systems that remain tools instead of decision makers and assistants instead of alternatives. Human rule, with all its disadvantages, is the main component in ensuring that these systems serve instead of sabotaging our interests.

Margaret Mitchell, Avigit Ghosh, Sasha Lakkoni, Jada Pistley all the work for the embrace of the faceand An international start -up company in artificial intelligence official open source.