AI is both liberating and enslaving us

Elaine Herzberg was pushing a bicycle laden with shopping across a busy road in Tempe, Arizona in 2018 when she was struck by a hybrid electric Volvo SUV at 40mph. At the time of the accident, the woman in the driver’s seat was watching a talent show on her phone. The SUV had been fitted with an autonomous driving system consisting of neural networks that integrated image recognisers. The reason Herzberg died was because what she was doing did not compute. The autonomous driving system recalibrated the car’s trajectory to avoid the bicycle, which it took to be travelling along the road, only to collide with Herzberg, who was walking across it. She became the first casualty of artificial intelligence.

What’s particularly poignant about this tragedy is that autonomous systems such as the one that killed Herzberg are, like most AIs, modelled on human intelligence, and yet are predicated on the idea that they can do that thinking better than us. The DeepMind Professor of Machine Learning at Cambridge, Neil Lawrence, calls such AIs human-analogue machines (HAMs). These HAMs attempt to emulate human behaviour. He writes: ‘Neural network models that emulate human intelligence use vast quantities of data that would not be feasible for any human to assimilate in our short lifetimes.’ A human driver, in the same circumstances, he argues, would have slowed down: ‘Delaying action is one of the ways we respond to the gremlin of uncertainty.’ AI has a problem identifying such gremlins. The car ploughed on, dragging Herzberg 20 metres down the road.

The great difference for Lawrence between human and machine intelligence is that the former is embodied. We are locked in, constrained by our physical brains. That, you’d think, is all to the advantage of the unconstrained machine intelligence.