HUMAN IN THE LOOP
What can a chimpanzee teach us about AI oversight? Rather a lot. At least about being the ‘human in the loop’.
On 29 November 1961, Enos went into orbit in a capsule built to fly itself. He did exactly what he’d been trained to do: pull levers to turn off lights. But when a thruster failed, he had no way to fix it. And he came home one orbit early.
NASA’s verdict was blunt. Both faults ‘could have been corrected had an astronaut been aboard’.
And less than three months later, John Glenn hit a similar thruster fault. He switched off the automatic control and flew the capsule himself.
Two flights, but only one with a human in the loop.
Back then, that phrase meant someone chosen, trained and given the controls. NASA even capped their height so they’d fit the capsule. And the astronauts asked for a window, so they could see what they were flying.
By the mid-1980s, computer scientists were arguing over whether to put a human in the loop of Star Wars missile defence. One put it simply: ‘How would a human ever know to intervene?’
By 2003, Patriot air-defence crews were still at their consoles. But after three friendly-fire incidents, a Pentagon review found the operating protocol ‘largely automatic’, with operators ‘trained to trust the system’s software’. The person was there. But the job had changed.
Which brings us to today.
We use the same four words for the person who checks an AI’s draft, the one who approves its decisions, the expert who trains it and the manager who answers for it.
So what do we actually mean by ‘human in the loop’? Who is the human? What is the loop? And what state are they in when it matters?
They’re questions we’ll be coming back to over the next few weeks. Because if a person is going to stay in the loop, they deserve the human’s job. Not the chimpanzee’s.
