Apparently we've reached the stage where we're happy that *only* 1% of the papers submitted to #ICLR2026 were fully AI-generated, and *all of* 61% were *mostly human-written*.

Sure, it's much less than the 21% fully AI-written reviews citing nonexistent references (yes, in 2025) and asking for meaningless additional experiments.

I guess we've made our bed and we're lying in it.

Did anyone do a version of the "this is fine" flame-surrounded dog, but it's ML papers and reviews burning? (Please don't automatically generate one, thanks.)

[All numbers estimated by one LLM detection tool.]