The letters humans chose—and why an AI agent chose differently
Human play spread across a wider set of letters, while a controlled agent concentrated on N, R, and S. The difference points to two distinct ways of reading a LimiLetter board.
The letters they chose
The clearest difference was concentration. Human selections were distributed across more of the alphabet. The agent repeatedly returned to a smaller group of high-coverage letters.
| Letter | Human play | Agent play | Agent gap |
|---|---|---|---|
| N | 12.5% | 18.8% | +6.3 pts |
| A | 11.3% | 6.3% | -5.0 pts |
| R | 7.5% | 12.5% | +5.0 pts |
| S | 7.5% | 12.5% | +5.0 pts |
| I | 8.8% | 6.3% | -2.5 pts |
| T | 7.5% | 6.3% | -1.2 pts |
Human opening choices were also more varied. E and T each accounted for 20.0% of human openings, and N accounted for 13.3%; six other letters each appeared in 6.7% of openings. The agent split its openings evenly among A, E, and R at 33.3% each.
Who scored more points?
Human play scored slightly more. The average completed human Daily earned 18,383 points, compared with 17,745 for the controlled agent. That is a 3.6% human advantage.
| Measure | Human play | Agent play | Result |
|---|---|---|---|
| Daily score | 18,383 | 17,745 | Humans +3.6% |
| LIMIs used | 16.0 | 16.0 | Tie |
Score and efficiency measure different things. Efficiency asks how few LIMIs completed the board. Score also depends on letter value, how many live words each choice removes, and the order of those choices. The human advantage came from higher-scoring eliminations, not from using fewer LIMIs overall.
Who was more efficient?
Neither side was more efficient across the full Daily: human play and the agent both averaged 16 LIMIs. The route to that tie differed by puzzle. Humans averaged 3.8 LIMIs on the medium puzzle while the agent used 5. The agent used 5 on the first hard puzzle while humans averaged 6.2. Both used 6 on the final hard puzzle.
| Puzzle | Human average | Agent | More efficient |
|---|---|---|---|
| 1 · medium | 3.8 | 5.0 | Humans |
| 2 · hard | 6.2 | 5.0 | Agent |
| 3 · hard | 6.0 | 6.0 | Tie |
Why the choices probably diverged
The agent optimized visible coverage
At each turn, inspect the ten visible leaders and choose a target-safe letter appearing in the most of them. Break ties by the summed rank of matching leaders, then alphabetically. That rule naturally favors letters shared by several prominent decoys, so its choices cluster around a smaller set of reliable coverage letters.
Humans appeared to change tactics with the board
The wider human distribution is consistent with players switching between broad elimination and targeted cleanup. A letter can be valuable even when it appears in fewer words if it removes the specific decoy blocking the target. Human pattern recognition may also make unusual letter combinations stand out sooner than a fixed coverage rule would.
Consistency is not the same as efficiency
The fixed policy helped on one board, hurt on another, and tied on the third. Its concentrated letter strategy was consistent, but it did not produce a lower overall LIMI count or a higher score.
What the percentages mean
We compared de-identified data from people’s play with a controlled agent replay of the same Daily. Both groups completed 100% of the comparison puzzles. The agent saw the same visible target and leading words, used the same game engine, and selected one target-safe letter at a time. The percentages describe this benchmark only; they do not represent every LimiLetter player or every kind of AI agent.