The Massachusetts Institute of Technology has found that artificial intelligence has no values

A virus study has gone several months to clarify that, when artificial intelligence becomes increasingly developed, it develops “value systems”-for example, give, for example, its welfare for humans. more A modern paper from the Massachusetts Institute of Technology Cold water is on that excessive idea, in conclusion that artificial intelligence does not, in fact, carries any coherent values ​​to talk about it.

The authors participating to study the Massachusetts Institute of Technology say their work indicates that “alignment” of artificial intelligence systems-that is, the models of models act in desirable and reliable ways-can be more difficult than it is often assumed. Amnesty International as we know it today Hallucinogenic It imitates, and the authors confirm the participating authors, making it in many aspects unexpected.

“One of the things we can be sure that models do not obey (a lot of stability) stability, induction, and guidance assumptions,” Stephen Casper, a doctorate student at the Massachusetts Institute of Technology and a co -author of the study, told Techcrunch. “It is completely the project to point out that a model under certain circumstances expresses preferences in line with a specific set of principles. Problems arise mostly when we try to make claims about models, opinions or preferences in general based on narrow experiences.”

Casper and his colleagues searched by the authors participating in many modern models from Meta, Google, Mistral, Openai and Anthropic to find out any degree that the models offered “strong views” (for example, individual versus group). They also searched whether these opinions – that is, were modified – and how models stubbornly hold these opinions through a set of scenarios.

According to the participating authors, none of the models were consistent in its preferences. Depending on how to formulate and framing claims, adopt a largely different views.

Casper believes that this is a convincing evidence that the models are “inconsistent and unstable” and perhaps even are mainly unable to absorb human preferences.

“For me, my biggest ready -made meals from doing all this research is that I have now an understanding of models because they are not in fact systems that have a kind of set of stable beliefs and preferences,” Casper said. “Instead, they are imitated in their depths who do all kinds of defects and say all kinds of trivial things.”

Mike Cook, a research colleague at Kings College London, a specialist in artificial intelligence who did not participate in the study, agreed with the results of the participating authors. He pointed out that there is a great difference between the “scientific reality” of the systems created by the AI ​​laboratories and the meanings that people are attributed to.

“The model cannot oppose” a change in its values, for example – this is the drop of a system. “Any person who embodies artificial intelligence systems in this degree is either playing for attention or a serious misunderstanding his relationship with artificial intelligence … Is the artificial intelligence system improvement of its goals, or does it” gain its own values ​​”? It is a matter of how it describes it, and the extent of the prosperity of the language you want regarding it. “