Ivory TowerSpiegeloog 446: Revolution

Ivory Tower: Regression Is All You Need

By November 16, 2026No Comments

The Turing test used to be a symbol of human uniqueness. According to one popular interpretation of the test, proposed by Alan Turing in 1950, a computer should be assigned consciousness if we fail to distinguish between the computer and a human. Even though criticisms of the test have existed from its conception, the Turing test was universally considered a tall order, and the evident failure of anything artificial to beat it provided a comfortable reassurance of the special place of human consciousness in the order of things. When the philosopher John Searle, in his infamous Chinese room experiment, proposed that a system could pass the Turing test without having any sort of conscious experience, he was mocked and ridiculed. One reason for the mockery was that everyone “knew” that to pass the Turing test would require not just one but several revolutions in our understanding of the brain, the mind, and their intricate relationship. That a system without intelligence and consciousness would be able to pass the Turing test at all was widely considered a purely academic thought experiment; many commentators pointed to the incredible sophistication of the brain and how advanced a system would need to be to engage in human conversation.

With the arrival of ChatGPT that proved to be complete and utter nonsense. Large Language Models (LLMs) are incredible computing machines, but they are most stunning because their amazing performance results from an extremely simple internal design. LLMs are basically very large regression models that have been fitted to very large datasets. Evidently, John Searle was right in this respect; GPT actually makes the Chinese Room look complicated. Conscious experience, understanding, or intelligence are not necessary to beat the Turing test. No scientific revolution required: Regression is all you need.

While we are still recovering from the realization that human linguistic behavior can be produced by such shockingly simple statistical models, the implications of this fact are slowly beginning to dawn on the academic community. If language is so easy to implement, then maybe thinking and reasoning aren’t that difficult either. And, lo and behold, consciousness is actually considered an option. A leading AI developer suggested as early as 2022 that Large Language Models like GPT and Claude may already be “slightly conscious”. Currently, surveys suggest that a sizeable minority of users are starting to believe that GPT has conscious states.

The traditional blockade against assigning consciousness to robots was the Turing test, but clearly that blockade doesn’t work anymore: Current AI systems are already able to fool some humans some of the time into believing that they are conscious, and before long they will fool most of the humans most of the time. So is there anything else that we could deploy to block the ascription of consciousness to AI? No, says Henry Shevlin in an excellent blogpost that reviews the scientific state of the field.

If that is true, then we’re bound to enter into a sort of evidential limbo with respect to the question whether AI has some sort of conscious experience, similar to the one we have with other species that we are unsure about. Do mice have consciousness? Frogs? Cockroaches? If that sounds ridiculous, some very small fish recently passed the dot test that used to be the domain of great apes and dolphins. GPT will follow soon: as my colleague Hannes Rosenbusch noted, the only thing it really lacks is a paintable forehead.

Denny Borsboom

Author Denny Borsboom

More posts by Denny Borsboom
Share