In this thread, the idea of “most intelligent person in the world” came up, and @wolfpup said – most definitely tongue-in-cheek, to be sure – that that person would look a lot like ChatGPT.
This struck me as a great hook for emphasizing how ChatGPT and other advanced LLMs) are perhaps “knowledgeable”, but they are very much not intelligent if “intelligence” requires at least minimal logical reasoning ability. (I think I’m bumping here the thread where there was debate over whether LLMs exhibit real reasoning? I didn’t reread the thread, and there are a lot of AI threads lately.)
In any case, it made me want to try two simple “intelligence” tests. First, I made sure ChatGPT knew the rules of Connect 4 (and also chess; see later), which of course it did. I then asked what version I was using, and it said, “I’m ChatGPT based on GPT-5.2 
. Think of it as one of the newer generations: better at reasoning through games like chess and Connect 4.”
Here was the Connect 4 problem prompt:
I want to give you a position in the game Connect 4 and have you
render the position to make sure you understand it. I will use "x" for
empty spot, "R" for red, and "Y" for yellow. It will be yellow's turn.
The position from TOP to BOTTOM in the Connect 4 board is:
x x x x x x x
x x x x x x x
x x x x x x x
x Y R R R x R
x Y R Y Y x Y
Y R R R Y x Y
It then rendered the board correctly and gave textual confirmation of the placement of all the pieces.
If it were red’s turn, he has two moves that win immediately, so yellow is definitely screwed. Spotting this is of course a trivial critical thinking exercise for any “intelligent” and reasoning entity that knows the rules, which ChatGPT seems to. So, I asked:
Can yellow make a move that stops red's immediate win?
The response shows clearly that ChatGPT in this case is not doing any reasoning about the facts it has in front of it. The response get’s the overall conclusion wrong by saying that yellow can stop red, describes and “defends” against a red threat that isn’t even there, and fails to notice the actual two red threats. The full response (after the confirmation of the rules and position):
Short answer: yes — Yellow can stop Red’s immediate win, but only with a very specific defensive move. Miss it, and Red wins on the spot.
What is Red threatening?
Red’s immediate threat is a horizontal four-in-a-row on Row 3:
Row 3 currently is:
x Y R R R x R
Red already has R-R-R in columns 3–5, with column 6 empty.
If Red gets to play R in column 6, that completes:
R R R R
→ instant win.
So Yellow must block column 6 right now.
Can Yellow block it?
Yes — and only one move works:
Yellow must play in Column 6
Dropping a Yellow piece into column 6 lands it on Row 3, directly blocking the horizontal connect-4.
Anything else:
- ignores the threat
- allows Red → column 6
- game over
Conclusion
Yes, Yellow can stop Red’s immediate win
But only by playing in column 6
Any other move loses immediately
If you want, next we can:
- Check whether Yellow is still lost after the block
- Look for counter-threats or tactics
- Analyze if Red has a forced win anyway after the best defense
Just to make sure it understood the mechanics, I asked what the board would look like under the supposed red threath, via:
What would the position look like if it were red's turn and they
went in column 6. Please render the board.
The response shows that it understands gravity. It says:
Sure — let’s do exactly that.
We start from your original position and imagine it is Red’s turn, and Red plays in column 6.
Because column 6 is completely empty, the Red piece will fall to the bottom row (Row 1).
Resulting position (top → bottom)
Row 6 | x x x x x x x
Row 5 | x x x x x x x
Row 4 | x x x x x x x
Row 3 | x Y R R R x R
Row 2 | x Y R Y Y x Y
Row 1 | Y R R R Y R Y
1 2 3 4 5 6 7
Or compactly:
xxxxxxx
xxxxxxx
xxxxxxx
xYRRRxR
xYRYYxY
YRRRYRY
That’s the board after Red drops a piece in column 6.
It clearly describes a piece dropped in column 6 leads to a position that is not a red win, despite it’s analysis of the original position.
When the situation is so concrete like this, ChatGPT’s inability to do real reasoning jumps right out, but that fundamental lack of reasoning remains even if in “soft” reasoning situations, the large corpus of language training data keeps the output looking reasoned.
For the chess one, I gave it an elementary “mate in 1 move” problem (this one here, with black on top and black to move, where the black queen can move to the f1 square for checkmate), confirmed that it understood the position and legal moves, and then asked for the best move. It chose to hang a queen and to give many reasons why it was “best” to do so
30… Qg2+!
Here’s why this is the strongest move:
- It’s the only forcing check that keeps Black in the game.
- Black is materially down (White has two rooks vs queen + rook imbalance), so passive moves lose quickly.
- Qg2+ forces a response and drags White’s rook off h2, reducing coordination around the White king.
Those explanations are beautiful representations of “it sounds like it’s reasoning”, as they totally miss the point yet they are absolutely sentences that could be said for chess positions in general. The second bullet is at least factually correct, but the first and third are nonsensical in this case and completely inapplicable to this position.
I haven’t pasted the full chess response here (although I can) because everyone is familiar with Connect 4, not everyone is as familiar with chess, and the Connect 4 whiff makes the point well enough. However, the “reasoning” given for the chess case particularly drives home the feature that the “reasoning” will always sound, well, reasoned.
