That’s up for debate. I would think it’s ‘fairer’ to have them track the ease at which they can get placed, and that letters occurring frequently in short words are easier to place than ones that do not.
As an extreme example, if you add a million 15-letter words starting with “QX” to the dictionary, Q and X get bottom scores in the rarity department, but do not become much easier to place.
As a realistic example, consider B. Google tells me BA, BE, BI, BO and BY are valid scrabble words. That makes B fairly easy to place.
It would be even ‘fairer’ to pick values that mean every Scrabble board has an equal chance of winning a Scrabble game, but that is as good as impossible to judge, even nowadays. Yes, you could let a computer play itself for zillions of games, but the optimal strategy will depend on the tile values, so you’d have to train it for every set of values you evaluate.
(I put ‘fairer’ in quotes because, in the end, the game can be whatever its designers see fit without being unfair. Some settings would not feel fun to play for most humans, say one where the sole Q in the game scores a million points, but even that wouldn’t be unfair to the players)