Take a moment to salute the majesty of human conversation. When we talk to each other, whether it’s about last night’s TV or the wisdom of a military strike on Syria, we are doing something far harder than sending a rocket to the moon. We did the moonshot decades ago but we still can’t make a machine that will hold a decent conversation.
On 14 September, researchers will gathered in Derry, Northern Ireland, to demonstrate their latest efforts. If any of them has created a machine that successfully mimics a human, they will leave $100,000 richer.
The money is being put up by Hugh Loebner, a New York based philanthropist. His goal, he says, is total unemployment for all human beings throughout the world. He wants robots to do all the work. And the first step towards that is apparently to develop computers that seem human when you chat to them.
It’s not a new idea. Alan Turing is credited with the first explicit outline of what is now called the Turing test. A human judge sits down at a computer and has a typed conversation with an entity that responds to whatever the judge types. If that entity is a computer, but the judge thinks it’s a person, the conversational computer program passes the test.
At the Derry event, the programs won’t compete directly. Instead, the judges will enter a conversation at two terminals, one of which conveys the thoughts of a human being, the other one being controlled by a program. The judge will decide which seems more human; if it’s the computer, that program goes through to the next round, where the challenges get harder.
So far, no one has won the big prize but every year the most convincing program wins a smaller amount. The creator of the last program to be rumbled this year will walk away with 4,000 of Loebner’s dollars.
Many people in this research field think the competition is a waste of time. The founder of MIT’s artificial intelligence (AI) laboratory, Marvin Minsky, once offered to pay $100 to anyone who can convince Loebner to withdraw his prize fund. Minsky’s problem is that the Loebner Prize gives AI a bad name. The programs are not convincing for long – steer the conversation the right way and you can unseat them fairly easily (you can see last year’s conversations here). Yet AI is in fact becoming rather useful.
Computers may not be able to hold a conversation with human beings, but algorithms that adapt “intelligently” to circumstances are starting to hit the streets: Google’s self-driving cars run on AI. The way phone calls are routed through a network relies on other autonomous, flexible programs. Email spam filters, speech-recognition software, stock-market trades and even some medical diagnoses routinely employ machines that seem to think for themselves.
Where the Loebner Prize is most useful is probably in providing a check on our enthusiasm. Researchers have created AI programs designed to look at CCTV footage and decide whether a crime is about to be committed. A rapidly moving limb suggests an assault taking place. Spotting a gait associated with fast running can be interpreted as someone fleeing a crime scene.
Similar innovations have been tried on the London Underground – a program looks for “suspicious” patterns of movement which indicate that someone might be preparing a terrorist attack or be about to jump under a train. Once the program has decided there is a risk, it will alert the authorities.
Though AI programs remain as flawed as those attempting to hold a conversation, let’s hope we won’t be tempted to cede all our liberties to them.