We need a Weizenbaum test for AI.

Abstract

Alan Turing introduced his 1950 paper on Computing Machinery and Intelligence with the question "Can machines think?" But rather than engaging in what he regarded as never-ending subjective debate about definitions of intelligence, he instead proposed a thought experiment. His "imitation game" offered a test in which an evaluator held conversations with a human and a computer. If the evaluator failed to tell them apart, the computer could be said to have exhibited artificial intelligence (AI). In the decades since Turing's paper, AI has gone from being a fountain of scientific hype to an academic backwater to a gold rush. Throughout, the Turing test has given computer scientists a sense of direction: a quest for what Turing called a "universal machine." Although the debate continues about whether the Turing test is a reasonable measure of artificial intelligence, the real problem is that it asks the wrong question. AI is no longer an academic debate. It is a technological reality. For society to make good decisions about AI, we should instead look to another great late 20th-century computer scientist, Joseph Weizenbaum. In a paper "On the impact of the computer on society," in Science in 1972, Weizenbaum argued that his fellow computer scientists should try to view their activities from the standpoint of a member of the public. Whereas computer scientists wonder how to get their technology to work and use "electronic wizardry" to make it safe, Weizenbaum argued that ordinary people would ask "is it good?" and "do we need these things?" As excitement builds about the possibilities of generative AI, rather than asking whether these machines are intelligent, we should instead ask whether they are useful.