ID Comment: Artificial intelligence for re/insurance should be made to fail the Turing test

As any fan of Benedict Cumberbatch will tell you, artificial intelligence (AI) isn’t new.

What they might not know is that an invisible pendulum has been swinging between two groups of data scientists in the decades since Alan Turing’s time as a codebreaker at Bletchley Park.

The Turing test, which in 1950 the famous mathematician called the imitation game, is a test of a computer's ability to communicate indistinguishably from a human. Turing proposed that a human evaluator would judge natural language conversations between another human and a machine designed to generate human-like responses, and if the evaluator could not reliably tell the machine from the human, the machine would be said to have passed the test.

Fast forward to the 1990s, and Peter Norvig and Stuart Russell argued that AI researchers had devoted little effort to passing the Turing test, believing it is more important to study the underlying principles of intelligence than to duplicate an exemplar.

The quest for artificial flight succeeded, Norvig and Russell said, when the Wright brothers stopped the practice of imitating birds. “Aeronautical engineering texts do not define the goal of their field as making machines that fly so exactly like pigeons,” the two computer scientists quipped, “that they can fool even other pigeons.”

The pendulum has swung between statistical AI, which is good at intuitive judgements, such as pattern recognition and object classification, and symbolic AI, which is good at principled judgements, such as logical reasoning and rule-based diagnoses.

When research by either of these two groups hits a wall, the other has tended to pick up the slack (and funding). Meanwhile, the first group tells stories of regret about the hype they generated as they sit around the campfire in their latest AI winter.

One such AI winter followed Fifth Generation Computer Systems (FGCS), the initiative begun in 1982 by Japan's Ministry of International Trade and Industry. Critics said it was a commercial flop, while sympathisers said (much later) that it was ahead of its time. The sympathisers were right: FGCS spurred the development of concurrent logic programming.

Scientists will tell you, it is often during the “winters” that they make advances, perhaps because no one else is watching and so they can concentrate. There was even more focus on data throughout the 2000s, and the next evolutionary stage of AI during the 2010s was machine learning. So far, so good, until you see an AI hallucination and, without warning, the pendulum swings again.

Scientists will tell you, it is often during the “winters” that they make advances, perhaps because no one else is watching and so they can concentrate

Where we are today is with large language models, or statistical predictions, but the next stage in AI’s evolution will be finding ways that best combine the two different approaches – statistical and symbolic.

There are still two separate groups of AI scientists. There are those working on machine learning and computational neuroscience, who gather at conferences like NeurIPS, and there is the semantic web community, who hang out every year at the Snowflake summit. There is, however, a third group. They are the scientists who want to be bridges between the other two. As one of them said, poetically: “The light is going to be shining on the folks who are combining.”

The risks from re/insurers using AI – to be better, faster, and even perhaps, fairer – are fewer than the risks from them not using it. The positives far outweigh the negatives, we’re told. While pursuing the art of the possible, however, it will be better if AI technology for re/insurance fails the Turing test.

Why? Eugene Goostman.

That is the chatbot developed in 2001 by one Ukrainian and two Russian computer programmers. Eugene is portrayed as a 13-year-old Ukrainian boy. These characteristics are intended to induce forgiveness – for grammatical errors and any lack of general knowledge – in those with whom “he” interacts.

Eugene eventually validated Turing’s prediction that, by the year 2000, machines would be capable of fooling 30% of human judges after five minutes of questioning. At a contest marking the 60th anniversary of Turing's death, on June 7, 2014 – three months after Russia annexed Crimea – 33% of the event's judges thought Eugene was human.

The cruel irony is lost on no one. No one with human intelligence, that is. It would, however, be lost on AI.

Re/insurers are there to do social good, and with so much inhumanity in the world, it would be wrong to dehumanise their work. By all means, have a machine crunch the data, but let a person help the client.

Insurance is a promise to pay. No machine could understand the sentiment behind that.