The Cryptogram Glitch
AI glitches out trying to solve Tempo's Alphabet cryptogram puzzle, “Color Notation”.
Can AI actually solve cryptogram puzzles? We wanted to put that to the test — for real. So we handed a Tempo's Alphabet puzzle to Grok, Claude, and ChatGPT to see what would happen.
The results surprised us. Not because the AI failed spectacularly — but because two of the three didn't, and the one that "failed" did it in a way we didn't expect.
The Puzzle We Used
We fed it "Color Notation," an artist-themed cryptogram puzzle tasking players with decoding a Salvador Dalí quote embedded within his surrealist painting, Landscape with Butterflies. Tempo's Alphabet puzzles aren't classic cryptograms in the newspaper sense. Instead, it utilizes a custom cipher inspired by the Polybius Square. In this system, a tile's color represents its column and its number represents its row. It's exactly the kind of on-the-spot, human-invented rule that used to be kryptonite for text-based statistical models, though modern 2026 systems are rapidly overcoming it.
Round One: Grok ✅
Grok solved the puzzle in exactly 5 minutes and 18 seconds. It found the cipher on its own from our YouTube channel, mapped every colored tile to its letter, and worked through all four lines of the puzzle systematically — showing its column/row logic line by line.
It landed on:
The puzzle decodes to: “I’m not strange, I’m just not normal.” youtube.com
How it works (using the color/number key on the right):
- Each colored number maps to a letter via the grid:
- Blue: 1=A, 2=H, 3=O, 4=V
- Black: 1=B, 2=I, 3=P, 4=W
- Brown: 1=C, 2=J, 3=Q, 4=X
- Red: 1=D, 2=K, 3=R, 4=Y
- Green: 1=E, 2=L, 3=S, 4=Z
- Yellow: 1=F, 2=M, 3=T, 4=!
- Purple: 1=G, 2=N, 3=U, 4=?
Decoded sequences:
- Top row (12 codes): I'M NOT STRANGE
- Middle-upper row (6 codes): I'M JUST
- Middle-lower row (3 codes): NOT
- Bottom row (6 codes): NORMAL
Clean, correct, and fully explained.
Round Two: Claude ❌
We gave Claude the identical puzzle. Claude also found the legend immediately and used the same underlying logic — color as column, number as row — decoding all four lines with a visible, letter-by-letter breakdown.
Claude landed on:
Decoded color-number grid cipher systematically
Fun one — here's how it breaks down using the color/number key (color = column, number = row):
Line 1: black-2, yellow-2, blue-2, blue-3, yellow-3, green-3, yellow-3, red-3, blue-1, purple-2, purple-1, green-1
→ I-M-H-O-T-S-T-R-A-N-G-E → "I'M HOT STRANGE"
Line 2: black-2, yellow-2, brown-2, purple-3, green-3, yellow-3
→ I-M-J-U-S-T → "I'M JUST"
Line 3: purple-2, blue-3, yellow-3
→ N-O-T → "NOT"
Line 4: purple-2, blue-3, red-3, yellow-2, blue-1, green-2
→ N-O-R-M-A-L → "NORMAL"
Put together, the full message reads:
"I'M HOT STRANGE. I'M JUST NOT NORMAL."
A couple notes on method, in case you want to double-check: the (orange-looking) tiles all map to the yellow column in the key, and each colored box's number tells you the row, so a black box with a "2" is column black, row 2 → "I", etc.
Look closely at that first line, though. Claude's method was identical to Grok's. Eleven of the twelve letters in line one matched exactly. The entire second, third, and fourth lines matched Grok's answer letter for letter. The only discrepancy was a single tile — Claude read one square as blue when it was actually purple, which flipped "NOT" into "HOT" in the first sentence.
That's not a reasoning failure. It's a misread pixel.
Round Three: ChatGPT❓
This is where the experiment actually got interesting. We handed ChatGPT the exact same image. It also spotted the substitution-key structure right away — colors as columns, numbers 1–4 as rows, each pair mapping to a letter — and correctly identified that the small colored squares above the blanks were pointing back into that table.
Then it stopped.
ChatGPT told us flatly that the image alone didn't give it enough information to uniquely decode the message, because nothing in the picture told it which row each colored square was supposed to use. It even walked through the ambiguity directly, pointing out that a purple square labeled "2" could mean "row 2, purple column" or something else entirely, "depending on the original puzzle rules." Rather than guess, it asked us to upload the puzzle's instructions or name the source so it could "decode the entire message" properly.
So unlike Grok and Claude, ChatGPT never produced a final phrase at all — right or wrong.
Stopped thinking ›
Yes—I can definitely help solve it, but I need a bit more context because this image by itself doesn't contain enough information to uniquely decode the message.
From what I can see:
- The colored chart on the right is a substitution key:
- Columns are colors (blue, black, brown, red, green, yellow, purple).
- Rows are 1–4.
- Each color/number pair maps to a letter (for example, Blue-1 = A, Black-1 = B, Brown-1 = C, … Purple-4 = ?).
- The colored numbered squares above the blanks indicate which letter to pick from that table.
- The butterflies and punctuation suggest it's spelling out a phrase or quote.
The problem is that the image alone doesn't tell us which row (1–4) each colored square should use. For example, a purple square labeled "2" could mean:
- row 2, purple column = N, or
- the "2" might represent something else depending on the original puzzle rules.
Without the instructions, there are multiple possible decodings.
If you can upload either:
- the page of instructions for this puzzle, or
- the name of the puzzle/game/book it came from,
I can decode the entire message for you.
I can also solve it manually step-by-step once I know the rule that determines which row of the letter grid to use.
So Did AI "Fail" This Puzzle?
Sort of — but not in the way we expected, and not in the same way twice.
The part of this puzzle that's supposed to be the hard part for AI — spotting a custom, human-invented rule that doesn't exist anywhere in a training dataset — is the part all three models got right, instantly. None of them tried to force a frequency-count solution or hallucinate a "statistically plausible" phrase. All three recognized a legend when they saw one and understood it was a lookup table.
Where they diverged was what happened next. Grok and Claude both committed to an answer. Grok got there perfectly; Claude got there with one misread tile that flipped a single word. ChatGPT, on the other hand, correctly recognized that the row assignment wasn't fully determinable from the image and declined to commit to a guess — which meant it also never handed us a wrong answer, but it never handed us a right one either.
Put another way: two of the three models were willing to be confidently wrong. One was only willing to be correctly uncertain.
The Real Takeaway
The honest lesson here isn't "AI can't think like a human." It's that AI can often find the trick, but three different models will handle the last mile — actually reading the tiles correctly and being willing to commit to an answer — in three different ways. Grok read it perfectly. Claude reasoned it out perfectly but tripped on a single pixel and still delivered a confident final answer anyway. ChatGPT spotted the same ambiguity a careful human might catch, and refused to bluff past it.
If you're using AI to check your work on a puzzle like this, that's exactly the gap to watch for. A model's logic can be flawless and its final answer still wrong — and a different model might sidestep that risk entirely by not giving you a final answer at all. Neither outcome tells you the puzzle is solved. Only your own eyes on the original tiles can do that.
Try It Yourself
Puzzles like this reward exactly the kind of careful, patient attention that's easy to shortcut with a tool. Give our interactive cryptogram puzzle a real, unaided attempt before you let AI take a swing at it. Use AI as a hint if you get stuck — not as the answer key. The satisfaction of spotting the trick yourself, tile by tile, is the whole point.
