QQuestion: Does AI really understand what it says?

Nobody has a settled answer yet. The records show that people have disagreed about this for a long time.

A good starting point is Alan Turing's paper from October 1950. Instead of asking "Can machines think?", Turing proposed the imitation game: judge a machine by how it behaves in conversation, not by what is inside it¹. Today's habit of comparing AI models by their test scores follows the same idea.

There are strong voices on the other side. The "stochastic parrots" paper, presented in March 2021, described a language model as a system that stitches together the forms of language it has seen, by probability, without any reference to meaning². One of its authors argued in a co-written book, published in May 2025, that today's AI systems are not thinking beings but text and media generators shaped by training data and human labor³.

Others have tried to test it. In June 2025, Apple researchers reported that AI accuracy collapses once problems get complex enough⁴. The same month a rebuttal argued that much of that collapse came from how the experiment was set up — limits on how long an answer could be, and puzzles that had no solution at all⁵. On ARC-AGI-3, a test released in March 2026, people solved 100% of the tasks at launch while frontier AI managed 0.51%⁶. In September 2026, OpenAI claimed its GPT-6 Astra scored 99.9% on the same test — a score the company measured itself⁷.

So here is where things stand. AI can talk very naturally, but whether that means it understands the way people do is still argued over by researchers. The more convincing an AI sounds, the more it's worth checking.