Machine learning is about to revolutionize the study of ancient games

That's because the article took a paper, scanned for the most "cool" looking part of it, and reported on a distorted version of that because it's the most exciting, I guess. And note that in the context of the paper, AI here is Minimax and MCTS AI, not "neUraL NetWorK" AI.

The relevant section about ludemes starts on page 13. But, whatever they wrote about is actually relevant to these 3 paragraphs below under 4.4. Somehow that was warped into something about genetic optimization (which is mildly orthogonal to machine learning, AI, etc. regardless) of games.

Basically, the paper is 100% about studying ancient games. It's not about "AI" studying games. It's about humans creating new mathematical tools (under the umbrella of machine learning and "AI") to better study ancient games.

Prepub link: https://arxiv.org/pdf/1905.13516.pdf

While strategy games do not inherently have any genetic material (de Voogt, 1999), a tree of
ludemes may similarly be viewed as the genotype of a game from which the way that it plays
(phenotype) emerges. When games transfer to other cultures, regions, religions, parts of a
population, etc., or when they are used as inspiration for the development of a new game or
game variant, we often expect some of its ludemes – though possibly not all of them – to
also transmit.

In the field of computational phylogenetics, it is assumed that populations (of organisms, or
languages, or games, etc.) did not appear spontaneously, but developed through “evolutionary”
processes which place strong constraints on the hypotheses that can plausibly explain observed
similarities and differences. Computational models encoding assumptions based on those
constraints can be used to construct plausible networks of “ancestry”. In biology such
networks are generally assumed to be trees, but in the case of cultural artefacts such as
games we expect there to be significantly more horizontal transfer. There have been previous
attempts at using phylogenetic techniques for Mancala games (Eagle, 1999) and Chess-like
games (Kraaijeveld, 2001), but they tended to confuse the genotype and phenotype of games
in these analyses (Morris, 2013).
Cameron Browne et al. 17

Similar to the quantitative estimates of game quality discussed in Subsection 4.3, we do
not expect computational phylogenetics to allow for conclusions with respect to the origins of
games with 100% certainty. However, the resulting networks (or distributions over plausible
networks) may again provide insight into plausible hypotheses to be investigated in more
detail in further research. We envision outputs of computational phylogenetics to be used in
combination with other analyses (such as those discussed in the previous subsection) and
partial evidence from more traditional research fields (e.g. archaeology).