Difficulty should change observable decisions
Word-game AI difficulty should vary candidate quality, consistency, and disclosed randomness. Longer response delay alone does not create a harder or more human opponent.
| Level | Selection pattern | Useful practice role |
|---|---|---|
| Easy | More forgiving legal choices | Learn the board and scoring |
| Medium | Mixed candidate quality | Practise consistent play |
| Hard | Stronger choices more often | Pressure-test decisions |
| Expert | Most consistent checked choices | Review narrow margins |
Difficulty needs observable dimensions
Labels such as Easy and Expert are useful only when they describe behavior. A word-game opponent can vary its chance of solving a clue, the quality of its legal placement, the time before it moves, and how often it passes. Publishing those dimensions lets players choose a challenge instead of guessing what a label hides.
Wordmap's tile-word game uses four disclosed profiles. Easy solves less often and chooses an early legal crossing. Medium solves more often and varies placements. Hard usually solves and seeks efficient crossings. Expert rarely passes and favors the highest-value legal crossing. These profiles are game parameters, not personalities.
Knowledge
Which legal candidates exist in the checked data?
Selection
How strongly does the computer rank those candidates?
Variation
How often does it choose below its strongest option?
Separate knowledge from move selection
An opponent can know the answer but make a weak placement, or fail the clue before placement begins. Modeling these stages separately creates clearer difficulty. It also reveals why a round changed: clue accuracy, crossing strategy, or timing. A single hidden strength number cannot provide that explanation.
For a small curated game, the AI should use the same legal board state as the player. It should not place through occupied cells, invent an answer outside the puzzle, or inspect a future random event. Strong play can be fast without receiving secret rules.
Delay is presentation, not intelligence
A response delay gives players time to follow the turn, but waiting longer does not make a weak move thoughtful. Keep delay modest and disclose that it is interface timing. Do not claim that the system is pondering, learning a player's psychology, or feeling pressure unless those capabilities genuinely exist.
Accessibility matters here too. Players should not be forced to wait through theatrical pauses, and reduced-motion preferences should not remove essential state changes. A clear status message is more useful than an animation pretending to reveal inner thought.
Probability creates variation
A solve probability prevents an easier opponent from becoming perfectly predictable. It should be sampled once at the correct decision stage, not repeatedly until the desired outcome appears. Repeated hidden rolls can make a nominal probability misleading and difficult to test.
Variation does not excuse impossible behavior. When the AI passes, the game should say it passed. When it solves, the placement must satisfy the same board rules. The randomness changes outcome frequency, not legality.
Choose a level for practice
Use Easy when learning placement rules, Medium when you want room to recover, Hard when inefficient crossings should matter, and Expert when the board should punish most missed opportunities. Switch levels when the current one no longer changes your decisions. Winning alone is not the only signal.
After a match, compare one player move with the strongest legal crossing visible at that moment. That review creates a concrete practice target. The AI is useful as a consistent source of pressure, not as an authority on language or a simulation of a human opponent.
Test whether the labels mean anything
- Give every level the same checked positions.
- Repeat enough trials to see selection patterns.
- Compare move quality separately from response time.
- Publish what changes and what does not.
Test whether levels are meaningfully different
Run a fixed set of reviewed rounds against each level and record solve, pass, legal placement, and score. The sample should be large enough to reveal obvious configuration mistakes, but the site should not invent a scientific benchmark from a few games. The practical question is whether the observed behavior follows the published profile.
If Easy and Expert produce the same choices across many fixed states, inspect the parameters and decision stages. If Expert wins only because it receives information the player cannot see, redesign the rule. Difficulty should emerge from disclosed accuracy and strategy differences while legality, puzzle data, and board access remain common.