do humans have an edge over scrabble engines?
It seems to be folklore in the Scrabble community that Nigel Richards still outperforms the best Scrabble engines.
Arjun Panickssery@panicksseryAnyone into competitive Scrabble?
I saw the other day the claim that the best player, Nigel Richards, is actually still better than SOTA Scrabble AI engines, I guess because of the branching factor and predicting a range of racks the opponent could have?
@grok can you explain
I am not sure whether this is actually true or not: I think it is arguable. We haven’t seen Nigel Richards play hundreds of games against Macondo or anything like that. So it seems entirely plausible that Macondo or another current Scrabble bot would win more than 50% of the time. (The win rate probably wouldn’t be much above 50%, because Scrabble is simply a rather luck-based game at the end of the day.)
There is at least a pretty strong argument that there were some Nigel moves that computers couldn’t replicate until earlier this year: the blog post Bayesian Rack Inference in Macondo: Teaching a Scrabble Bot to Read Minds by César Del Solar from a couple months ago goes over one. The post starts with a summary of the move:
At the 2018 National Scrabble Championship in Buffalo, NY, the GOAT of all time Nigel Richards was playing against another top player and grandmaster Joel Sherman in a de-facto three-game finals for the $10K top prize. Joel needed to beat him 3 in a row to win the whole thing. In the second game (Joel won the first one), Joel opened with the rack of AENSTUU. He thought briefly before exchanging UU, his best play. Nigel had the rack AELNOQT and the commentators stated the obvious - play QAT for 24 points. There was no question in anyone’s mind about this basically automatic play.
After a minute or so, Nigel plays -OQ (exchange OQ) and one of the commentators gasps audibly. You can see the moment in this video on YouTube:
Until a new feature added in April of this year, Macondo preferred QAT. The new feature tries to run Bayesian inference to figure out which tiles your opponent has on their rack after each turn, and then makes a decision for a play in light of this fact. The idea behind the move is that after the exchange Joel probably kept good bingo tiles, and might be able to use the A or T in QAT to play a bingo — and with the rack inference feature, Macondo says the exchange is very slightly better. César Del Solar’s post goes into more detail on the maths the inference uses, though I haven’t had time to read through it yet.
Macondo is also (arguably) weak in closed boards, at least against certain sorts of opponents. It tends to rely on hoping that its opponent will open the position and give it a chance to catch back up, rather than setting up lanes itself. Opening the position can often be a bad idea — your opponent gets a turn to try to take advantage of it before you do — but if you’re already losing, then it is often necessary to take risks like that.
One possible recent example of this dynamic is Mark Meller playing MOO in the 2026 North American final. Mark describes his reasoning like this:
Meller: The more I think about my position, the more I don’t like it. Fifty points might not seem like a massive deficit, but with tiles and bingo lanes rapidly disappearing, and without any good spots for my Q-U, I no longer feel being patient is an option. TOQUE scores weakly, and then if Orry blocks the D in DITZ, I have nowhere to bingo and I just don’t see myself making up the deficit. MONO has the same problem: If Orry blocks the D, I’m left with little chance for a comeback, and he’s going to be careful not to open a big Q spot.
Fatsis: You decide to play MOO below ChOWSES, making EM and SO and “forking the board,” that is, opening two potentially high-scoring lanes, one on the bottom row and one on the rightmost column. Gutsy!
Meller: MOO looks a bit crazy, and obviously there are some cases where Orry slams the door shut with a 100-point bingo there. But when you’re already in a bad position, you can’t live in fear—I don’t want to die a slow death on a closed board without giving myself a chance. After MOO, my leave of AEQU suddenly becomes way more potent, with a strong threat of 50-point plays like QUARE or QUATE, either on the bottom making EME or on the right making MOOR or MOOT.
Also, from my perspective, bingos like ACQUIRE or BAROQUE on the bottom aren’t out of the question, and if Orry blocks there I have some draws at massive bingos to the D of DITZ, such as ACQUIRED and QUARRIED. I’m not loving my chances after MOO, but that’s the precise reason I feel the need to make the play.
This move is also described in a recent Will Anderson video.
But Macondo apparently preferred TOQUE. Now, it’s hard to say for sure that the engine is wrong. There are at least arguments in favor of MOO, and it might depend somewhat on exactly what sort of opponent you’re facing, but — I mean, maybe TOQUE is in fact just better, it’s not like you can just go ask an engine to adjudicate this argument.
I do think that to some extent the Scrabble community folklore might just be outdated, though, even if you can make various arguments about places where Macondo has gaps. Sure, it’s not a situation anything like Chess, where the engines are obviously just far and away better — some of that is due to Scrabble being a luck-based game that you can win by getting lucky even if the engine is a stronger player than you, but it’s certainly not all that! But it’s really not so bad.
I think a lot of people just have impressions that are still based on when Quackle was the state of the art or whatever. Quackle definitely makes mistakes. But in my (very brief) research tonight I haven’t been able to find any unambiguous examples of the current state of the art bots making outright errors, rather than there just being cases where you can sort of make the argument that a move the engine disprefers might be better.
Macondo has clear deficiencies. For most of the game it has two modes: static and sim. Sim is generally preferred, and I think empirically 5-ply sim has had best results. But 5-ply sim is Monte Carlo using an empirical winrate lookup table at the end of a 5-turn rollout assuming both players play like a simple static evaluation (that is, playing solely based on points + rack leave value). So it models its future self poorly, and potentially its opponent as well. This is basically unchanged from Quackle, see http://www.breakingthegame.net/computers3 .
There are also other bots, see the solvers channel of the Woogles Discord for discussion. I think some work has been done on nested sims in MAGPIE but idk how well they perform yet. Also discussion and trial of other techniques.
Oops, meant to link to http://www.breakingthegame.net/computers5
There are cases where engine-preferred move is worse, but it's mostly because of how current engines work, doing a Monte Carlo simulation and assuming your opponent will play his top equity play.
On closed boards, macondo sometimes struggles to find the best moves (tho it's often arguable which of the moves is the best, when you cannot trust the engine), especially when it's trailing, it tends to assume that opponent will open up a bingo spot, which is very unlikely to happen.
https://woogles.io/game/9uDRKeYKf8?turn=22
For example VITAE here is a clear error to any human player. It's hard to say what's best for sure, but this move basically concedes the game. I would strongly prefer something like INTI or NIT (hard to say their real win %, it should be lower than what the analysis says).
The engines give VITAE about 30% win rate, in reality i believe it is 0%, because the only bingos possible to draw are through the AG, and any sane player will simply block the spot (as Mack did).
The endgame and pre-endgame (1 in the bag) are played perfectly by the current engines if they look deep enough. On woogles website the bots sometimes make errors because they are solving the endgame at insufficient depth.
1-in-the-bag preendgame is still not played quite perfectly and plausibly never will be, as such would involve inference.