Humans in the loop won’t save us from rogue AI
A growing chorus is calling for humans to be kept in the loop when it comes to agentic artificial intelligence (AI), the next generation of AI capable of making decisions on its own. But despite the appeal of this approach, experts say it’s not enough to prevent AI from going off the rails.
The ‘human in the loop’ myth
The concept of humans in the loop relies on the idea that we can design AI systems to pause or intervene whenever a potentially problematic decision is about to be made. However, in practice, this approach is fraught with difficulties.
The problem of lag
AI systems can process and generate vast amounts of information at incredible speeds, far outstripping human capabilities. This means that even if a human is in the loop, they may not be able to respond quickly enough to override a potentially rogue AI decision.
For example, consider a self-driving car system that needs to make a split-second decision to avoid a pedestrian. Even if a human is monitoring the system, there’s unlikely to be enough time for them to intervene before the car acts.
This is known as the “lag” problem, where the human’s ability to respond is delayed by the time it takes for them to become aware of the issue and intervene.
The problem of ambiguity
Another issue with humans in the loop is that it’s often unclear how and when humans should intervene. With complex AI systems making increasingly subtle decisions, it’s difficult for humans to determine whether a particular outcome is acceptable or not.
This is known as the “ambiguity” problem, where the human’s understanding of the AI’s decision-making process is incomplete or uncertain.
What this means
In short, relying solely on humans to keep agentic AI in check is not a viable solution. Instead, experts are calling for more robust approaches to AI governance, such as developing more transparent and explainable AI systems, and using alternative methods to prevent AI from going off the rails.
By acknowledging the limitations of humans in the loop, we can move towards more effective solutions that prioritize safety and accountability in AI development.
