The Voight-Kampff Test

The movie Blade Runner is based on a novel called Do Androids Dream of Electric Sheep? by Philip K Dick. The movie plays up the action and downplays the robotic animals that the social climbers in the novel acquire so that they can pretend to their neighbours that they have real life pets, such as a sheep, long after most such things have gone extinct in the bleak dystopian world filled with pollution. It's a bizarre and often quite amusing Sci-fi.

The main character has the job of tracking down androids who try to pass themselves off as human. The problem is that the hardware is good enough now that he can't tell humans from androids on sight alone and they even have some blood and flesh on their exoskeletons so simply checking if the bleed isn't enough. Sawing one in two would guarantee being able to differentiate, but it would be a test to destruction which must be avoiding in case they were a real human.

So the standard way is to ask a series of questions called a Voight-Kampff test, where the questions are designed to measure not knowledge, but human things which are hard to replicate like appropriate emotions. Not only do they have to give appropriately human responses, they also have to do so at the correct speed. Answer a question which most humans would reel from in disgust too quickly or too clinically and you out yourself.

For instance in one interrogation one of the questions starts with the preamble: “You are watching an old movie on TV, a movie from before the war. It shows a banquet in progress; the entree consists of boiled dog, stuffed with rice.” The android he is testing interrupts at this point, but doesn't react to the mention of boiled dog. Notice that this is a test not of intelligence where AI can outsmart humans on many knowledge or pattern recognition based tasks, but on something very human: emotion. This is a slightly different approach to the Turing Test which uses conversation rather than emotional responses as its way of differentiating.

I was recently trying to sign up an account on SoundCloud and I had to pass a reCAPTCHA test (more on the purpose of those here) where I had to pick the squares with road signs and cars in them. This is used to build up the machine-learning algorithm for recognising things needed to navigate roads for driverless vehicles. The reCAPTCHA provides the site using it a way of telling the difference between humans and bots by asking the human to do something which computers find hard but human find easy.

Both the Voight-Kampff test in the Philip K Dick Novel and the particular reCAPTCHA I was using suffer from the same problem. As AI gets better, what happens when the very best AI surpasses the level of the least good human? Do Androids Dream of Electric Sheep’s central plot deals with this issue and I managed to fail the reCAPTCHA three times as the cases of identifying signs that haven't already been decided are the ones where both humans and AI find it hard. It's either that or I am a robot.