AI Logic: Not the Logic You Recognize from Geometry, Folks
March 3, 2026
Remember that one science fiction movie where a super computer deemed humans as inferior and took over a space ship or society? In order to regain control of the computer, an intelligent, debonair space hero asked one simple question: “Why?” The computer encounters a paradox and it shuts itself down or explodes. Apparently real life AI experience similar situations with simple logic problems says Popular Mechanics: “Scientists Found AI’s Fatal Flaw—The Most Advanced Models Are Failing Basic Logic Tests.”
Carleton College, Cal Tech, and Stanford combined their research on LLMs, such as Claude and ChatGPT. They learned that LLMs are as prone to error as humans and might even perform worse than their creators. Here are their errors with cognitive reasoning:
-
“LLMs perpetuate human errors like bias, and they make other human-like errors because they don’t have the intuitive scaffolding that helps us learn not to make those mistakes.
-
LLMs lack core executive functions (working memory, cognitive flexibility, and inhibitory control) that help humans succeed in reasoning, leading to systemic failures in LLMs.
-
LLMs are poor at abstract reasoning, like understanding relationships between intangible concepts (e.g. knowledge, trust, security) and picking out rules affecting small sets.
-
LLMs show human-like confirmation bias toward information they already parse well.
-
LLMs show order and anchoring biases, like overweighting the first example in a list of items.”
LLMs are even worse at social reasoning. They lack morals, can’t understand social rules, are easily manipulated, and given to hallucinations. The LLMs can’t “comprehend” basic logic such as two-hop reasoning, basic yes and no questions, and casual inference. That’s only the tip of the iceberg in LLMs
It’s good to know that LLMs are not yet perfect. Each improvement opens the door for more hyperbolic marketing. The question I have is, “Will humans learn how to deal in a constructive manner with AI logic?” Some humans have that doom scrolling nailed. But AI logic?
Whitney Grace, March 3, 2026
Comments
Got something to say?
-
- Subscribe to Beyond Search
Feature archive
News archive
-
