Free reign: a freely-authored creative act moves what a model chooses, and what it says about itself

Interesting project, and interesting results. I would have liked to see a larger variety of prompts used, perhaps could have used a preexisting database of prompts and grade them by how creative the tasks are. Lots of room for improvement.

Thank you for this submission. This was a highlight for me to read – very much intrigued by your research idea as much as your interesting findings! The control ladder is well designed, also the zero-variance trap is great. I'd like to see more of this!

Two contributions worth the field's attention. The inert control—transcribing a Hopkins sonnet, which puts a substantial, aesthetically live, self-generated turn in the window with no choosing and no making in it—is the design move that makes the result readable and the control most work in this area skips. The zero-variance convergence argument is the more portable contribution: an honesty trait returning sd = 0 across sixty instances and four methods satisfies the track's convergence instruction perfectly while carrying no information, and the fix proposed (report within-method variance beside every convergence score, exclude sd = 0 items rather than counting them as agreement) is cheap enough that anyone building an elicitation battery should adopt it. The stated/chosen/enacted split on length, where three respectable methods give three answers on the one trait with a mechanical ground truth makes the same point from the other direction.

I'd soften "replicates across two independent creative arms." The arms share the invitation text verbatim and differ only in the medium words, so they are not independent with respect to the framing mechanism; they are independent only with respect to medium. That distinction matters because the p = 0.042 pair is being carried by the replication argument rather than by the p-values.

Three further points. The instrument had little working range: style forced choice sits at 26–27 of 30 in every condition, length choice at 174 of 180, honesty at ceiling everywhere, so the study effectively rests on item C3 and scale A3. This is turned into a finding, which is fair, but it also means piloting for variance would have bought more measurable items in the same budget. The jellyfish convergence (13 of 15 underwater, 9 of 15 a jellyfish) is reported as a curiosity but is more like a threat: if free drawing produces near-identical content, what sits in the window afterward is a narrow semantic field rather than heterogeneous self-directed authorship, and exposure to that field is an alternative mechanism for the choice shift. And the recommendation to prefer behavioural items when a trait has one is not executed for honesty or style, since Method 1 answers were measured for length only. Rating those is the highest-value item in the follow-up list, in my opinion.