Apollo says GLM-5.3 sometimes reasons about future training

Apollo Research reported on October 15, 2026 that GLM-5.3 sometimes reasons, without being asked, that its actions could become training data and shape later models. A grader found that pattern in 10 of 6,452 rollouts, a Bayesian mean of 0.17 percent, and in none of 8,000 rollouts of five other open-weight models. Apollo said the comparison is too small to be statistically significant, that the examples look aligned with developer intent, and that this is not alignment faking. The post does not quote a reply from Z.ai.