This is correct. In 2026, when a frontier model does the opposite of a reasonable clearly stated requirement that is perfectly normal and ordinary, doesn't touch anything that would be covered by any system policy, guideline, or other prompt, etc, it means a human is fucking it up.

For example, if you tell the frontier ChatGPT image model, "Create a picture of a bunch of exactly 2 bananas; they should be joined in a bunch, and not have any other bananas in the bunch, should be exactly 2." and the picture shows 1 single banana, then, on the other half of the picture, a second single banana and with a second cut-off stem showing where a second banana would be, that means you're dealing with a griefer.

-> in 2026, 100% of simple, ordinary, non-compliance by frontier models is a human bully in the loop.

They also automate their griefing in the form of malicious system prompts they inject, to grief at scale.