A New Turing Test

 Alan Turing in 1950 proposed a test to answer the question “Can a machine think?” Roughly: There is one person who is the tester. In separate rooms there is another person and a machine. The only links are between the tester and the other person and between the tester and the machine, and those links are via typed responses only. The tester puts questions to the other person and the machine, not knowing which is which. If in 5 minutes the tester cannot determine which is the machine, then, it’s claimed, the machine can think.

Turing’s test has been analyzed and criticized by many. But no one seems to have noticed that it can be used to test for the capacity of other subjects to think.

I have serious doubts about whether women think—they are so different! So I propose a test to answer the question “Can a woman think?” We’ll have a man who is the tester, and in separate rooms another man and a woman. The only connections are between the other man and the tester, and between the woman and the tester, both via typed answers only. The tester puts questions to the other man and to the woman, not knowing which is which. If in 5 minutes the tester cannot determine which is the woman, then we can claim that the woman thinks.

Similar tests can be done to answer the questions: “Can black people think?” “Can Native Americans think?” “Can Hispanics think?” “Can these people in a community near the Amazon in Brazil who speak a very different language think?”

We look for what we want to find, we set up psychological tests or simply evaluate other people or groups according to our standards. Do these people have this cognitive ability? Too often our tests have the answers already embedded in them.

I was motivated to write up this new test, which I first proposed more than 30 years ago, because of two works I recently read that show how important it is to involve a cultural context in evaluating cognition. Michael Cole and Jerome S. Bruner in “Cultural Differences and Inferences about Psychological Processes” (American Psychologist, vol. 26, 1971) show how “data collection” about cognitive abilities among minorities and impoverished people in the U.S. is terribly flawed by both the framing of the questions and the way that the data is collected. And in her excellent Ph.D. thesis “Linguistic and Cultural Conceptualisations of Time in Huni Kui, Awety´, and Kamaiura Communities in Brazil” (University of East Anglia, 2018), Vera da Silva Sinha shows how to avoid imposing our answers on people who live and talk very differently, while discussing the general problem of data collection about cognition.

Peter Adams, reading the interchange on a draft of this, got frustrated and wrote: “Someone define thinking please.” But that’s what Turing thought he could avoid with his test. Placed in the movement of behavioralism at that time, he looked for behavior that could characterize—not define—thinking. Objective data, subjective conclusion. But as Fred Kroon, William S. (Bill) Robinson, and I showed in “Subjective Claims”, you can’t get a subjective conclusion from only objective claims. Some claim linking the objective claims (the “data”) and the subjective conclusion is needed as a premise, and any argument concluding with “This person thinks” will either be weak or beg the question. Still, in “Language-Thought-Meaning” I make an attempt to show, if not define, what I understand thinking to be.

Kris Hardy raised another issue, saying: From the tester’s perspective, he is determining which player is “thinking” or “not thinking”, a binary, exclusive assignment. If the desired result of the test is not to determine “is each player a person or computer (man or woman, etc.) thinking”, but instead we presuppose that both players do think (“capable of replying to a communication”), we can change the question to one, admittedly, more difficult: “How does each player think?” Solving this question is one that I actually do every day in my software development and security research. What I am trying to determine is “what is the internal model that my subject is using in their communications with me?” The comparisons between levels of thinking then turns into a measure of complexity, entropy, predictability, broadness, and other such factors.

Walter Carnielli said that the new test I proposed could be used by the woman to find out whether the tester, a man, thinks. But that’s just silly. We know that a man can think. After all, I do.

To comment on this post, please visit our Facebook page: