AI (still) doesn’t reason? Yes, but…

Do you remember my June 2024 article, “AI (still) doesn’t reason: Alice’s experiment“? I started with a seemingly elementary question: “Alice has 6 brothers and 10 sisters. How many sisters does Alice’s brother have?” A fifth-grader answers in three seconds: 11 sisters. Many AI models, however, struggle with whether or not to include Alice herself.

The experiment (again)

Recently, I presented the same riddle to my digital friend “Sam,” the ChatGPT Pro version, to see if it had learned anything. The result?

• Answer: Correct, 11 sisters.
• Response time: 3 minutes and 45 seconds.

One wonders if it was truly “reasoning,” or if it simply simulated human logic more effectively.

Why does AI stumble over a fifth-grade problem?

1. Probabilistic estimation, not thought: AI generates text based on a probabilistic inferential process, not on true “conscious reasoning.”
2. Data bias: If there is confusion or errors in the training data, the model may repeat them.
3. Context and memory: AI does not always retain previous corrections; occasionally it falls back into the same errors, especially when starting a new session.
4. The human factor remains crucial: We must “stay in the loop,” because AI alone, at the moment, does not guarantee infallibility (as Mollick teaches).

An evolutionary leap?

In the June 2024 experiment, the AI got bogged down. Today, “Sam” (ChatGPT Pro) gets it right, but takes nearly 4 minutes to parse the logic. The question remains:

• More computing power = greater capacity to “reason”?
• Or is it just a simulation of reasoning, more refined and with fewer slips?

The moral of the story

We experiment continuously with AI, as Ethan Mollick suggests when speaking of Co-Intelligence: “Invite it to the table,” but keep in mind that:

• AI does not replace human reasoning.
• It needs our feedback (repeatedly!) to improve.
• Every version and every “digital twin” can behave slightly differently.

So what? Well, whether it takes 3 seconds or 3 minutes, the answer to Alice’s riddle remains the same – 11 sisters. The issue is not speed, but rather understanding the why. And the most honest answer, so far, is that AI does not perform logic like we do, but it is quite good at simulating it.