The article lists three assumptions that one might intuitively think hold. The problematic assumption is this one: > When someone observes an event happening,

The article lists three assumptions that one might intuitively think hold. The problematic assumption is this one: > When someone observes an event happening, it really happened. Given the experiment from the paper, I would rephrase this assumption as "it's not possible to rewind things". But obviously inside a computer simulation made up of operations that are all individually reversible, it's trivial to rewind things. The experiment sets up a situation where certain information is reversibly recorded, and then the records are unmade by temporarily rewinding. The rewinding is obfuscated by hiding it inside of a measurement that is incompatible with the presence or absence of the record. I guess the authors might disagree about the measurement implying rewinding, but as a bit of evidence I'll note that Scott Aaronson and Yosi Atia have shown that performing the measurement in question over a simulated agent is at least as expensive as rewinding the simulated agent [1]. Whatever is being done, it is doing some seriously expensive screwing around with the agent's state. It's like the experiment has a step where you feed the human through a giant meat processing plant, and for some reason everyone is pretending that's not somehow important. Basically, the authors are appealing to the intuition that humans are big complicated in-practice-irreversible things, so clearly a record is permanent if it has affected the state of a human. But then they imagine instantiating the human's state inside a ridiculously powerful computer capable of reversibly simulating time advancing and of performing operations that have been engineered specifically to mess with the presence or absence of the record's effects on the human's state. Surprise surprise, the record gets messed with. Then for the actual experiment the big complicated ball of dependent spaghetti that is a human is replaced by a nice simple photon going along one path or another path. 1: https://www.scottaaronson.com/blog/?p=4786 https://www.scottaaronson.com/blog/?p=4786