Enlarge image

Virtual personality: chatbot (symbolic image)

Photo: Teera Konakan / Getty Images

Chatbots with artificial intelligence behave in such a way that they can hardly be distinguished from humans.

This is the conclusion reached by a US research team led by computer scientist Qiaozhu Mei from the University of Michigan and economist Matthew Jackson from Stanford University in a new study that appeared in the “Proceedings of the National Academy of Sciences”.

The authors refer to the Turing test formulated by Alan Turing in 1950: You can tell whether artificial intelligence exists because a questioner cannot distinguish whether a person or a machine is formulating the answers.

At the time, Turing predicted that this would happen around the year 2000.

As it turned out, development took a little longer.

For the new study, the researchers had the software fill out a psychological survey and tested it with various games.

How they behaved should indicate characteristics such as trust, fairness, risk aversion, willingness to help or cooperation.

The bots' actions were anonymously compared with those of humans who were randomly selected from tens of thousands of test subjects from 50 different countries.

“Statistically indistinguishable from humans”

The ChatGPT-4 language model published last year showed “behavioral and personality traits that are statistically indistinguishable from those of humans.”

His features tended to be viewed as more human than those of real people.

The predecessor model ChatGPT-3 could be identified as artificial in some respects, but it was also close.

According to the study, both variants learned from experience during the test and adjusted their behavior as if they were humane.

ChatGPT-4 appeared particularly human in the game “Ultimatum,” in which a player chooses whether he accepts how his counterpart divides a sum of money between them, or both come away empty-handed.

Even in a public goods game in which players either keep money for themselves or give it away for shared use, the software was largely considered human.

The moves were viewed as indicators of fair, selfless or cooperative behavior.

People aren't that selfless

As an exception, ChatGPT-4 was not very convincing in the game “Prisoner's Dilemma”.

In it, both players have to decide at the same time in several rounds whether they will work together with the other and share a reward or whether they will each strive for the full amount of a smaller sum.

Here, too, it was about cooperation, reciprocity and strategic thinking.

Such games are often used for analysis in behavioral economics because they can simplify how people decide between different options in everyday life.

The researchers derived a programmed attitude from the behavior: the software strives to maximize both its own benefit and that of its counterpart, and on average in exactly equal proportions.

Real people are a little less altruistic - although not completely selfish, they usually give self-interest a slightly higher weight than the benefit of others.

There are also larger differences between individuals, while the software behaves a little more predictably.